diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2022-12-15.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2023-01-29.sh similarity index 100% rename from .github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2022-12-15.sh rename to .github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2023-01-29.sh index 761eb72e2..7d2853c17 100755 --- a/.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2022-12-15.sh +++ b/.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2023-01-29.sh @@ -21,9 +21,9 @@ tree $repo/ ls -lh $repo/test_wavs/*.wav pushd $repo/exp -git lfs pull --include "data/lang_bpe_500/HLG.pt" git lfs pull --include "data/lang_bpe_500/L.pt" git lfs pull --include "data/lang_bpe_500/LG.pt" +git lfs pull --include "data/lang_bpe_500/HLG.pt" git lfs pull --include "data/lang_bpe_500/Linv.pt" git lfs pull --include "data/lang_bpe_500/bpe.model" git lfs pull --include "exp/cpu_jit.pt" diff --git a/.github/scripts/run-librispeech-streaming-zipformer-2023-05-18.sh b/.github/scripts/run-librispeech-streaming-zipformer-2023-05-18.sh index 45324cb27..f4e2124b1 100755 --- a/.github/scripts/run-librispeech-streaming-zipformer-2023-05-18.sh +++ b/.github/scripts/run-librispeech-streaming-zipformer-2023-05-18.sh @@ -23,6 +23,7 @@ ls -lh $repo/test_wavs/*.wav pushd $repo/exp git lfs pull --include "data/lang_bpe_500/bpe.model" +git lfs pull --include "data/lang_bpe_500/tokens.txt" git lfs pull --include "exp/jit_script_chunk_16_left_128.pt" git lfs pull --include "exp/pretrained.pt" ln -s pretrained.pt epoch-99.pt @@ -33,7 +34,7 @@ log "Export to torchscript model" ./zipformer/export.py \ --exp-dir $repo/exp \ --use-averaged-model false \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ --causal 1 \ --chunk-size 16 \ --left-context-frames 128 \ @@ -46,7 +47,7 @@ ls -lh $repo/exp/*.pt log "Decode with models exported by torch.jit.script()" ./zipformer/jit_pretrained_streaming.py \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ --nn-model-filename $repo/exp/jit_script_chunk_16_left_128.pt \ $repo/test_wavs/1089-134686-0001.wav @@ -60,7 +61,7 @@ for method in greedy_search modified_beam_search fast_beam_search; do --method $method \ --beam-size 4 \ --checkpoint $repo/exp/pretrained.pt \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ $repo/test_wavs/1089-134686-0001.wav \ $repo/test_wavs/1221-135766-0001.wav \ $repo/test_wavs/1221-135766-0002.wav diff --git a/.github/scripts/run-librispeech-zipformer-2023-05-18.sh b/.github/scripts/run-librispeech-zipformer-2023-05-18.sh index 6aac1793e..fb1a0149d 100755 --- a/.github/scripts/run-librispeech-zipformer-2023-05-18.sh +++ b/.github/scripts/run-librispeech-zipformer-2023-05-18.sh @@ -23,6 +23,7 @@ ls -lh $repo/test_wavs/*.wav pushd $repo/exp git lfs pull --include "data/lang_bpe_500/bpe.model" +git lfs pull --include "data/lang_bpe_500/tokens.txt" git lfs pull --include "exp/jit_script.pt" git lfs pull --include "exp/pretrained.pt" ln -s pretrained.pt epoch-99.pt @@ -33,7 +34,7 @@ log "Export to torchscript model" ./zipformer/export.py \ --exp-dir $repo/exp \ --use-averaged-model false \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ --epoch 99 \ --avg 1 \ --jit 1 @@ -43,7 +44,7 @@ ls -lh $repo/exp/*.pt log "Decode with models exported by torch.jit.script()" ./zipformer/jit_pretrained.py \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ --nn-model-filename $repo/exp/jit_script.pt \ $repo/test_wavs/1089-134686-0001.wav \ $repo/test_wavs/1221-135766-0001.wav \ @@ -56,7 +57,7 @@ for method in greedy_search modified_beam_search fast_beam_search; do --method $method \ --beam-size 4 \ --checkpoint $repo/exp/pretrained.pt \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ $repo/test_wavs/1089-134686-0001.wav \ $repo/test_wavs/1221-135766-0001.wav \ $repo/test_wavs/1221-135766-0002.wav diff --git a/.github/scripts/run-librispeech-zipformer-ctc-2023-06-14.sh b/.github/scripts/run-librispeech-zipformer-ctc-2023-06-14.sh index cfa9c420c..0026d2109 100755 --- a/.github/scripts/run-librispeech-zipformer-ctc-2023-06-14.sh +++ b/.github/scripts/run-librispeech-zipformer-ctc-2023-06-14.sh @@ -23,6 +23,7 @@ ls -lh $repo/test_wavs/*.wav pushd $repo/exp git lfs pull --include "data/lang_bpe_500/bpe.model" +git lfs pull --include "data/lang_bpe_500/tokens.txt" git lfs pull --include "data/lang_bpe_500/HLG.pt" git lfs pull --include "data/lang_bpe_500/L.pt" git lfs pull --include "data/lang_bpe_500/LG.pt" @@ -40,7 +41,7 @@ log "Export to torchscript model" --use-transducer 1 \ --use-ctc 1 \ --use-averaged-model false \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ --epoch 99 \ --avg 1 \ --jit 1 @@ -51,7 +52,7 @@ log "Decode with models exported by torch.jit.script()" for method in ctc-decoding 1best; do ./zipformer/jit_pretrained_ctc.py \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ --model-filename $repo/exp/jit_script.pt \ --HLG $repo/data/lang_bpe_500/HLG.pt \ --words-file $repo/data/lang_bpe_500/words.txt \ @@ -71,8 +72,7 @@ for method in ctc-decoding 1best; do --use-ctc 1 \ --method $method \ --checkpoint $repo/exp/pretrained.pt \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ - --words-file $repo/data/lang_bpe_500/words.txt \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ --HLG $repo/data/lang_bpe_500/HLG.pt \ --G $repo/data/lm/G_4_gram.pt \ --words-file $repo/data/lang_bpe_500/words.txt \ diff --git a/.github/scripts/test-ncnn-export.sh b/.github/scripts/test-ncnn-export.sh index 52491d2ea..ac16131d0 100755 --- a/.github/scripts/test-ncnn-export.sh +++ b/.github/scripts/test-ncnn-export.sh @@ -195,14 +195,14 @@ git lfs pull --include "data/lang_char_bpe/Linv.pt" git lfs pull --include "exp/pretrained.pt" cd exp -ln -s pretrained.pt epoch-99.pt +ln -s pretrained.pt epoch-9999.pt popd ./pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py \ --lang-dir $repo/data/lang_char_bpe \ --exp-dir $repo/exp \ --use-averaged-model 0 \ - --epoch 99 \ + --epoch 9999 \ --avg 1 \ --decode-chunk-len 32 \ --num-encoder-layers "2,4,3,2,4" \ diff --git a/.github/workflows/run-aishell-2022-06-20.yml b/.github/workflows/run-aishell-2022-06-20.yml index c46cea0f6..d14196f38 100644 --- a/.github/workflows/run-aishell-2022-06-20.yml +++ b/.github/workflows/run-aishell-2022-06-20.yml @@ -44,7 +44,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -119,5 +119,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: aishell-torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless3-2022-06-20 + name: aishell-torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless3-2022-06-20 path: egs/aishell/ASR/pruned_transducer_stateless3/exp/ diff --git a/.github/workflows/run-gigaspeech-2022-05-13.yml b/.github/workflows/run-gigaspeech-2022-05-13.yml index f8ee25cc4..0e47f7538 100644 --- a/.github/workflows/run-gigaspeech-2022-05-13.yml +++ b/.github/workflows/run-gigaspeech-2022-05-13.yml @@ -43,7 +43,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -122,5 +122,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-gigaspeech-pruned_transducer_stateless2-2022-05-12 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-gigaspeech-pruned_transducer_stateless2-2022-05-12 path: egs/gigaspeech/ASR/pruned_transducer_stateless2/exp/ diff --git a/.github/workflows/run-librispeech-2022-03-12.yml b/.github/workflows/run-librispeech-2022-03-12.yml index d42202b79..3edbe43ec 100644 --- a/.github/workflows/run-librispeech-2022-03-12.yml +++ b/.github/workflows/run-librispeech-2022-03-12.yml @@ -43,7 +43,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -155,5 +155,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless-2022-03-12 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless-2022-03-12 path: egs/librispeech/ASR/pruned_transducer_stateless/exp/ diff --git a/.github/workflows/run-librispeech-2022-04-29.yml b/.github/workflows/run-librispeech-2022-04-29.yml index f42c8f27a..bb44a073b 100644 --- a/.github/workflows/run-librispeech-2022-04-29.yml +++ b/.github/workflows/run-librispeech-2022-04-29.yml @@ -43,7 +43,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -174,12 +174,12 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless2-2022-04-29 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless2-2022-04-29 path: egs/librispeech/ASR/pruned_transducer_stateless2/exp/ - name: Upload decoding results for pruned_transducer_stateless3 uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless3-2022-04-29 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless3-2022-04-29 path: egs/librispeech/ASR/pruned_transducer_stateless3/exp/ diff --git a/.github/workflows/run-librispeech-2022-05-13.yml b/.github/workflows/run-librispeech-2022-05-13.yml index 1fbd96157..e7b53b21c 100644 --- a/.github/workflows/run-librispeech-2022-05-13.yml +++ b/.github/workflows/run-librispeech-2022-05-13.yml @@ -43,7 +43,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -155,5 +155,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless5-2022-05-13 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless5-2022-05-13 path: egs/librispeech/ASR/pruned_transducer_stateless5/exp/ diff --git a/.github/workflows/run-librispeech-2022-11-11-stateless7.yml b/.github/workflows/run-librispeech-2022-11-11-stateless7.yml index 596596bd9..7e378c9a1 100644 --- a/.github/workflows/run-librispeech-2022-11-11-stateless7.yml +++ b/.github/workflows/run-librispeech-2022-11-11-stateless7.yml @@ -155,5 +155,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless7-2022-11-11 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless7-2022-11-11 path: egs/librispeech/ASR/pruned_transducer_stateless7/exp/ diff --git a/.github/workflows/run-librispeech-2022-11-14-stateless8.yml b/.github/workflows/run-librispeech-2022-11-14-stateless8.yml index dca7d6d25..a2c1a0ad6 100644 --- a/.github/workflows/run-librispeech-2022-11-14-stateless8.yml +++ b/.github/workflows/run-librispeech-2022-11-14-stateless8.yml @@ -155,5 +155,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless8-2022-11-14 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless8-2022-11-14 path: egs/librispeech/ASR/pruned_transducer_stateless8/exp/ diff --git a/.github/workflows/run-librispeech-2022-12-01-stateless7-ctc.yml b/.github/workflows/run-librispeech-2022-12-01-stateless7-ctc.yml index cd41e988e..500ab1736 100644 --- a/.github/workflows/run-librispeech-2022-12-01-stateless7-ctc.yml +++ b/.github/workflows/run-librispeech-2022-12-01-stateless7-ctc.yml @@ -159,5 +159,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless7-ctc-2022-12-01 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless7-ctc-2022-12-01 path: egs/librispeech/ASR/pruned_transducer_stateless7_ctc/exp/ diff --git a/.github/workflows/run-librispeech-2022-12-08-zipformer-mmi.yml b/.github/workflows/run-librispeech-2022-12-08-zipformer-mmi.yml index 91242c401..1a7f9f594 100644 --- a/.github/workflows/run-librispeech-2022-12-08-zipformer-mmi.yml +++ b/.github/workflows/run-librispeech-2022-12-08-zipformer-mmi.yml @@ -163,5 +163,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-zipformer_mmi-2022-12-08 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-zipformer_mmi-2022-12-08 path: egs/librispeech/ASR/zipformer_mmi/exp/ diff --git a/.github/workflows/run-librispeech-2022-12-29-stateless7-streaming.yml b/.github/workflows/run-librispeech-2022-12-29-stateless7-streaming.yml index 8490a62fc..68014e20c 100644 --- a/.github/workflows/run-librispeech-2022-12-29-stateless7-streaming.yml +++ b/.github/workflows/run-librispeech-2022-12-29-stateless7-streaming.yml @@ -168,5 +168,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless7-streaming-2022-12-29 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless7-streaming-2022-12-29 path: egs/librispeech/ASR/pruned_transducer_stateless7_streaming/exp/ diff --git a/.github/workflows/run-librispeech-2022-12-15-stateless7-ctc-bs.yml b/.github/workflows/run-librispeech-2023-01-29-stateless7-ctc-bs.yml similarity index 96% rename from .github/workflows/run-librispeech-2022-12-15-stateless7-ctc-bs.yml rename to .github/workflows/run-librispeech-2023-01-29-stateless7-ctc-bs.yml index e0130a636..821abc25d 100644 --- a/.github/workflows/run-librispeech-2022-12-15-stateless7-ctc-bs.yml +++ b/.github/workflows/run-librispeech-2023-01-29-stateless7-ctc-bs.yml @@ -14,7 +14,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -name: run-librispeech-2022-12-15-stateless7-ctc-bs +name: run-librispeech-2023-01-29-stateless7-ctc-bs # zipformer on: @@ -34,7 +34,7 @@ on: - cron: "50 15 * * *" jobs: - run_librispeech_2022_12_15_zipformer_ctc_bs: + run_librispeech_2023_01_29_zipformer_ctc_bs: if: github.event.label.name == 'run-decode' || github.event.label.name == 'blank-skip' || github.event_name == 'push' || github.event_name == 'schedule' runs-on: ${{ matrix.os }} strategy: @@ -124,7 +124,7 @@ jobs: export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH - .github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2022-12-15.sh + .github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2023-01-29.sh - name: Display decoding results for librispeech pruned_transducer_stateless7_ctc_bs if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' @@ -159,5 +159,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless7-ctc-bs-2022-12-15 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless7-ctc-bs-2023-01-29 path: egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/exp/ diff --git a/.github/workflows/run-librispeech-conformer-ctc3-2022-11-28.yml b/.github/workflows/run-librispeech-conformer-ctc3-2022-11-28.yml index 40a37da57..905515dc4 100644 --- a/.github/workflows/run-librispeech-conformer-ctc3-2022-11-28.yml +++ b/.github/workflows/run-librispeech-conformer-ctc3-2022-11-28.yml @@ -151,5 +151,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-conformer_ctc3-2022-11-28 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-conformer_ctc3-2022-11-28 path: egs/librispeech/ASR/conformer_ctc3/exp/ diff --git a/.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml b/.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml index aba29d066..501fae38c 100644 --- a/.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml +++ b/.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml @@ -26,7 +26,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.8] fail-fast: false @@ -159,5 +159,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'shallow-fusion' || github.event.label.name == 'LODR' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-lstm_transducer_stateless2-2022-09-03 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-lstm_transducer_stateless2-2022-09-03 path: egs/librispeech/ASR/lstm_transducer_stateless2/exp/ diff --git a/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml b/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml index fd497601d..bf73d4f18 100644 --- a/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml +++ b/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml @@ -43,7 +43,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -153,5 +153,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless3-2022-04-29 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless3-2022-04-29 path: egs/librispeech/ASR/pruned_transducer_stateless3/exp/ diff --git a/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml b/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml index 57fe5b999..6ea308468 100644 --- a/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml +++ b/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml @@ -43,7 +43,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -155,5 +155,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless2-2022-06-26 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless2-2022-06-26 path: egs/librispeech/ASR/pruned_transducer_stateless2/exp/ diff --git a/.github/workflows/run-librispeech-streaming-zipformer-2023-05-18.yml b/.github/workflows/run-librispeech-streaming-zipformer-2023-05-18.yml index ed934d56d..5145fb43c 100644 --- a/.github/workflows/run-librispeech-streaming-zipformer-2023-05-18.yml +++ b/.github/workflows/run-librispeech-streaming-zipformer-2023-05-18.yml @@ -170,5 +170,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-zipformer-2022-11-11 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-zipformer-2022-11-11 path: egs/librispeech/ASR/zipformer/exp/ diff --git a/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml b/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml index 515122a66..9fe2f0389 100644 --- a/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml +++ b/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml @@ -43,7 +43,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -155,5 +155,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-transducer_stateless2-2022-04-19 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-transducer_stateless2-2022-04-19 path: egs/librispeech/ASR/transducer_stateless2/exp/ diff --git a/.github/workflows/run-librispeech-zipformer-2023-05-18.yml b/.github/workflows/run-librispeech-zipformer-2023-05-18.yml index 7ecf0d2a0..e9d235ad1 100644 --- a/.github/workflows/run-librispeech-zipformer-2023-05-18.yml +++ b/.github/workflows/run-librispeech-zipformer-2023-05-18.yml @@ -155,5 +155,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-zipformer-2022-11-11 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-zipformer-2022-11-11 path: egs/librispeech/ASR/zipformer/exp/ diff --git a/.github/workflows/run-librispeech-zipformer-ctc-2023-06-14.yml b/.github/workflows/run-librispeech-zipformer-ctc-2023-06-14.yml index 569ce48fc..48f0b1532 100644 --- a/.github/workflows/run-librispeech-zipformer-ctc-2023-06-14.yml +++ b/.github/workflows/run-librispeech-zipformer-ctc-2023-06-14.yml @@ -151,5 +151,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-zipformer-2022-11-11 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-zipformer-2022-11-11 path: egs/librispeech/ASR/zipformer/exp/ diff --git a/.github/workflows/run-pretrained-conformer-ctc.yml b/.github/workflows/run-pretrained-conformer-ctc.yml index 8aaea35f6..bcd326b9d 100644 --- a/.github/workflows/run-pretrained-conformer-ctc.yml +++ b/.github/workflows/run-pretrained-conformer-ctc.yml @@ -33,7 +33,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false diff --git a/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml b/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml index 03a1df48e..1e5b25f5c 100644 --- a/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml +++ b/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml @@ -42,7 +42,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -154,5 +154,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-transducer_stateless_multi_datasets-100h-2022-02-21 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-transducer_stateless_multi_datasets-100h-2022-02-21 path: egs/librispeech/ASR/transducer_stateless_multi_datasets/exp/ diff --git a/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml b/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml index 8da4ff56a..9063c0ed6 100644 --- a/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml +++ b/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml @@ -42,7 +42,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -154,5 +154,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-transducer_stateless_multi_datasets-100h-2022-03-01 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-transducer_stateless_multi_datasets-100h-2022-03-01 path: egs/librispeech/ASR/transducer_stateless_multi_datasets/exp/ diff --git a/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml b/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml index 0b3e70d77..2d24528d3 100644 --- a/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml +++ b/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml @@ -33,7 +33,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false diff --git a/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml b/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml index a6a59d339..761b26131 100644 --- a/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml +++ b/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml @@ -33,7 +33,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false diff --git a/.github/workflows/run-pretrained-transducer-stateless.yml b/.github/workflows/run-pretrained-transducer-stateless.yml index 98d84bf96..e46b9a849 100644 --- a/.github/workflows/run-pretrained-transducer-stateless.yml +++ b/.github/workflows/run-pretrained-transducer-stateless.yml @@ -42,7 +42,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false @@ -154,5 +154,5 @@ jobs: uses: actions/upload-artifact@v2 if: github.event_name == 'schedule' || github.event.label.name == 'run-decode' with: - name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-transducer_stateless-2022-02-07 + name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-transducer_stateless-2022-02-07 path: egs/librispeech/ASR/transducer_stateless/exp/ diff --git a/.github/workflows/run-pretrained-transducer.yml b/.github/workflows/run-pretrained-transducer.yml index 8c1a652e0..190e446bc 100644 --- a/.github/workflows/run-pretrained-transducer.yml +++ b/.github/workflows/run-pretrained-transducer.yml @@ -33,7 +33,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.7, 3.8, 3.9] fail-fast: false diff --git a/.github/workflows/run-wenetspeech-pruned-transducer-stateless2.yml b/.github/workflows/run-wenetspeech-pruned-transducer-stateless2.yml index 6c70c646b..319a5558a 100644 --- a/.github/workflows/run-wenetspeech-pruned-transducer-stateless2.yml +++ b/.github/workflows/run-wenetspeech-pruned-transducer-stateless2.yml @@ -33,7 +33,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - os: [ubuntu-18.04] + os: [ubuntu-latest] python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-yesno-recipe.yml b/.github/workflows/run-yesno-recipe.yml index f997e634a..8a2c94829 100644 --- a/.github/workflows/run-yesno-recipe.yml +++ b/.github/workflows/run-yesno-recipe.yml @@ -33,7 +33,7 @@ jobs: runs-on: ${{ matrix.os }} strategy: matrix: - # os: [ubuntu-18.04, macos-10.15] + # os: [ubuntu-latest, macos-10.15] # TODO: enable macOS for CPU testing os: [ubuntu-latest] python-version: [3.8] diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml index e04fb5655..363556bb7 100644 --- a/.github/workflows/test.yml +++ b/.github/workflows/test.yml @@ -35,9 +35,9 @@ jobs: matrix: os: [ubuntu-latest] python-version: ["3.8"] - torch: ["1.10.0"] - torchaudio: ["0.10.0"] - k2-version: ["1.23.2.dev20221201"] + torch: ["1.13.0"] + torchaudio: ["0.13.0"] + k2-version: ["1.24.3.dev20230719"] fail-fast: false @@ -66,14 +66,14 @@ jobs: pip install torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html pip install torchaudio==${{ matrix.torchaudio }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html - pip install k2==${{ matrix.k2-version }}+cpu.torch${{ matrix.torch }} -f https://k2-fsa.org/nightly/ + pip install k2==${{ matrix.k2-version }}+cpu.torch${{ matrix.torch }} -f https://k2-fsa.github.io/k2/cpu.html pip install git+https://github.com/lhotse-speech/lhotse # icefall requirements pip uninstall -y protobuf pip install --no-binary protobuf protobuf==3.20.* pip install kaldifst - pip install onnxruntime + pip install onnxruntime matplotlib pip install -r requirements.txt - name: Install graphviz @@ -83,13 +83,6 @@ jobs: python3 -m pip install -qq graphviz sudo apt-get -qq install graphviz - - name: Install graphviz - if: startsWith(matrix.os, 'macos') - shell: bash - run: | - python3 -m pip install -qq graphviz - brew install -q graphviz - - name: Run tests if: startsWith(matrix.os, 'ubuntu') run: | @@ -129,40 +122,10 @@ jobs: cd ../transducer_lstm pytest -v -s - - name: Run tests - if: startsWith(matrix.os, 'macos') - run: | - ls -lh - export PYTHONPATH=$PWD:$PWD/lhotse:$PYTHONPATH - lib_path=$(python -c "from distutils.sysconfig import get_python_lib; print(get_python_lib())") - echo "lib_path: $lib_path" - export DYLD_LIBRARY_PATH=$lib_path:$DYLD_LIBRARY_PATH - pytest -v -s ./test - - # run tests for conformer ctc - cd egs/librispeech/ASR/conformer_ctc + cd ../zipformer pytest -v -s - cd ../pruned_transducer_stateless - pytest -v -s - - cd ../pruned_transducer_stateless2 - pytest -v -s - - cd ../pruned_transducer_stateless3 - pytest -v -s - - cd ../pruned_transducer_stateless4 - pytest -v -s - - cd ../transducer_stateless - pytest -v -s - - # cd ../transducer - # pytest -v -s - - cd ../transducer_stateless2 - pytest -v -s - - cd ../transducer_lstm - pytest -v -s + - uses: actions/upload-artifact@v2 + with: + path: egs/librispeech/ASR/zipformer/swoosh.pdf + name: swoosh.pdf diff --git a/docs/source/conf.py b/docs/source/conf.py index 6901dec02..0ff3f801c 100644 --- a/docs/source/conf.py +++ b/docs/source/conf.py @@ -86,6 +86,7 @@ rst_epilog = """ .. _git-lfs: https://git-lfs.com/ .. _ncnn: https://github.com/tencent/ncnn .. _LibriSpeech: https://www.openslr.org/12 +.. _Gigaspeech: https://github.com/SpeechColab/GigaSpeech .. _musan: http://www.openslr.org/17/ .. _ONNX: https://github.com/onnx/onnx .. _onnxruntime: https://github.com/microsoft/onnxruntime diff --git a/docs/source/decoding-with-langugage-models/LODR.rst b/docs/source/decoding-with-langugage-models/LODR.rst new file mode 100644 index 000000000..7ffa0c128 --- /dev/null +++ b/docs/source/decoding-with-langugage-models/LODR.rst @@ -0,0 +1,184 @@ +.. _LODR: + +LODR for RNN Transducer +======================= + + +As a type of E2E model, neural transducers are usually considered as having an internal +language model, which learns the language level information on the training corpus. +In real-life scenario, there is often a mismatch between the training corpus and the target corpus space. +This mismatch can be a problem when decoding for neural transducer models with language models as its internal +language can act "against" the external LM. In this tutorial, we show how to use +`Low-order Density Ratio `_ to alleviate this effect to further improve the performance +of langugae model integration. + +.. note:: + + This tutorial is based on the recipe + `pruned_transducer_stateless7_streaming `_, + which is a streaming transducer model trained on `LibriSpeech`_. + However, you can easily apply LODR to other recipes. + If you encounter any problems, please open an issue here `icefall `__. + + +.. note:: + + For simplicity, the training and testing corpus in this tutorial are the same (`LibriSpeech`_). However, + you can change the testing set to any other domains (e.g `GigaSpeech`_) and prepare the language models + using that corpus. + +First, let's have a look at some background information. As the predecessor of LODR, Density Ratio (DR) is first proposed `here `_ +to address the language information mismatch between the training +corpus (source domain) and the testing corpus (target domain). Assuming that the source domain and the test domain +are acoustically similar, DR derives the following formular for decoding with Bayes' theorem: + +.. math:: + + \text{score}\left(y_u|\mathit{x},y\right) = + \log p\left(y_u|\mathit{x},y_{1:u-1}\right) + + \lambda_1 \log p_{\text{Target LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) - + \lambda_2 \log p_{\text{Source LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) + + +where :math:`\lambda_1` and :math:`\lambda_2` are the weights of LM scores for target domain and source domain respectively. +Here, the source domain LM is trained on the training corpus. The only difference in the above formular compared to +shallow fusion is the subtraction of the source domain LM. + +Some works treat the predictor and the joiner of the neural transducer as its internal LM. However, the LM is +considered to be weak and can only capture low-level language information. Therefore, `LODR `__ proposed to use +a low-order n-gram LM as an approximation of the ILM of the neural transducer. This leads to the following formula +during decoding for transducer model: + +.. math:: + + \text{score}\left(y_u|\mathit{x},y\right) = + \log p_{rnnt}\left(y_u|\mathit{x},y_{1:u-1}\right) + + \lambda_1 \log p_{\text{Target LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) - + \lambda_2 \log p_{\text{bi-gram}}\left(y_u|\mathit{x},y_{1:u-1}\right) + +In LODR, an additional bi-gram LM estimated on the source domain (e.g training corpus) is required. Comared to DR, +the only difference lies in the choice of source domain LM. According to the original `paper `_, +LODR achieves similar performance compared DR in both intra-domain and cross-domain settings. +As a bi-gram is much faster to evaluate, LODR is usually much faster. + +Now, we will show you how to use LODR in ``icefall``. +For illustration purpose, we will use a pre-trained ASR model from this `link `_. +If you want to train your model from scratch, please have a look at :ref:`non_streaming_librispeech_pruned_transducer_stateless`. +The testing scenario here is intra-domain (we decode the model trained on `LibriSpeech`_ on `LibriSpeech`_ testing sets). + +As the initial step, let's download the pre-trained model. + +.. code-block:: bash + + $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29 + $ pushd icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp + $ git lfs pull --include "pretrained.pt" + $ ln -s pretrained.pt epoch-99.pt # create a symbolic link so that the checkpoint can be loaded + +To test the model, let's have a look at the decoding results **without** using LM. This can be done via the following command: + +.. code-block:: bash + + $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/ + $ ./pruned_transducer_stateless7_streaming/decode.py \ + --epoch 99 \ + --avg 1 \ + --use-averaged-model False \ + --exp-dir $exp_dir \ + --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model + --max-duration 600 \ + --decode-chunk-len 32 \ + --decoding-method modified_beam_search + +The following WERs are achieved on test-clean and test-other: + +.. code-block:: text + + $ For test-clean, WER of different settings are: + $ beam_size_4 3.11 best for test-clean + $ For test-other, WER of different settings are: + $ beam_size_4 7.93 best for test-other + +Then, we download the external language model and bi-gram LM that are necessary for LODR. +Note that the bi-gram is estimated on the LibriSpeech 960 hours' text. + +.. code-block:: bash + + $ # download the external LM + $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm + $ # create a symbolic link so that the checkpoint can be loaded + $ pushd icefall-librispeech-rnn-lm/exp + $ git lfs pull --include "pretrained.pt" + $ ln -s pretrained.pt epoch-99.pt + $ popd + $ + $ # download the bi-gram + $ git lfs install + $ git clone https://huggingface.co/marcoyang/librispeech_bigram + $ pushd data/lang_bpe_500 + $ ln -s ../../librispeech_bigram/2gram.fst.txt . + $ popd + +Then, we perform LODR decoding by setting ``--decoding-method`` to ``modified_beam_search_lm_LODR``: + +.. code-block:: bash + + $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp + $ lm_dir=./icefall-librispeech-rnn-lm/exp + $ lm_scale=0.42 + $ LODR_scale=-0.24 + $ ./pruned_transducer_stateless7_streaming/decode.py \ + --epoch 99 \ + --avg 1 \ + --use-averaged-model False \ + --beam-size 4 \ + --exp-dir $exp_dir \ + --max-duration 600 \ + --decode-chunk-len 32 \ + --decoding-method modified_beam_search_lm_LODR \ + --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model + --use-shallow-fusion 1 \ + --lm-type rnn \ + --lm-exp-dir $lm_dir \ + --lm-epoch 99 \ + --lm-scale $lm_scale \ + --lm-avg 1 \ + --rnn-lm-embedding-dim 2048 \ + --rnn-lm-hidden-dim 2048 \ + --rnn-lm-num-layers 3 \ + --lm-vocab-size 500 \ + --tokens-ngram 2 \ + --ngram-lm-scale $LODR_scale + +There are two extra arguments that need to be given when doing LODR. ``--tokens-ngram`` specifies the order of n-gram. As we +are using a bi-gram, we set it to 2. ``--ngram-lm-scale`` is the scale of the bi-gram, it should be a negative number +as we are subtracting the bi-gram's score during decoding. + +The decoding results obtained with the above command are shown below: + +.. code-block:: text + + $ For test-clean, WER of different settings are: + $ beam_size_4 2.61 best for test-clean + $ For test-other, WER of different settings are: + $ beam_size_4 6.74 best for test-other + +Recall that the lowest WER we obtained in :ref:`shallow_fusion` with beam size of 4 is ``2.77/7.08``, LODR +indeed **further improves** the WER. We can do even better if we increase ``--beam-size``: + +.. list-table:: WER of LODR with different beam sizes + :widths: 25 25 50 + :header-rows: 1 + + * - Beam size + - test-clean + - test-other + * - 4 + - 2.61 + - 6.74 + * - 8 + - 2.45 + - 6.38 + * - 12 + - 2.4 + - 6.23 \ No newline at end of file diff --git a/docs/source/decoding-with-langugage-models/index.rst b/docs/source/decoding-with-langugage-models/index.rst new file mode 100644 index 000000000..6e5e3a4d9 --- /dev/null +++ b/docs/source/decoding-with-langugage-models/index.rst @@ -0,0 +1,33 @@ +Decoding with language models +============================= + +This section describes how to use external langugage models +during decoding to improve the WER of transducer models. + +The following decoding methods with external langugage models are available: + + +.. list-table:: LM-rescoring-based methods vs shallow-fusion-based methods (The numbers in each field is WER on test-clean, WER on test-other and decoding time on test-clean) + :widths: 25 50 + :header-rows: 1 + + * - Decoding method + - beam=4 + * - ``modified_beam_search`` + - Beam search (i.e. really n-best decoding, the "beam" is the value of n), similar to the original RNN-T paper. Note, this method does not use language model. + * - ``modified_beam_search_lm_shallow_fusion`` + - As ``modified_beam_search``, but interpolate RNN-T scores with language model scores, also known as shallow fusion + * - ``modified_beam_search_LODR`` + - As ``modified_beam_search_lm_shallow_fusion``, but subtract score of a (BPE-symbol-level) bigram backoff language model used as an approximation to the internal language model of RNN-T. + * - ``modified_beam_search_lm_rescore`` + - As ``modified_beam_search``, but rescore the n-best hypotheses with external language model (e.g. RNNLM) and re-rank them. + * - ``modified_beam_search_lm_rescore_LODR`` + - As ``modified_beam_search_lm_rescore``, but also subtract the score of a (BPE-symbol-level) bigram backoff language model during re-ranking. + + +.. toctree:: + :maxdepth: 2 + + shallow-fusion + LODR + rescoring diff --git a/docs/source/decoding-with-langugage-models/rescoring.rst b/docs/source/decoding-with-langugage-models/rescoring.rst new file mode 100644 index 000000000..de7e700d0 --- /dev/null +++ b/docs/source/decoding-with-langugage-models/rescoring.rst @@ -0,0 +1,256 @@ +.. _rescoring: + +LM rescoring for Transducer +================================= + +LM rescoring is a commonly used approach to incorporate external LM information. Unlike shallow-fusion-based +<<<<<<< HEAD +methods (see :ref:`shallow_fusion`, :ref:`LODR`), rescoring is usually performed to re-rank the n-best hypotheses after beam search. +======= +methods (see :ref:`shallow-fusion`, :ref:`LODR`), rescoring is usually performed to re-rank the n-best hypotheses after beam search. +>>>>>>> 80d922c1583b9b7fb7e9b47008302cdc74ef58b7 +Rescoring is usually more efficient than shallow fusion since less computation is performed on the external LM. +In this tutorial, we will show you how to use external LM to rescore the n-best hypotheses decoded from neural transducer models in +`icefall `__. + +.. note:: + + This tutorial is based on the recipe + `pruned_transducer_stateless7_streaming `_, + which is a streaming transducer model trained on `LibriSpeech`_. + However, you can easily apply shallow fusion to other recipes. + If you encounter any problems, please open an issue `here `_. + +.. note:: + + For simplicity, the training and testing corpus in this tutorial is the same (`LibriSpeech`_). However, you can change the testing set + to any other domains (e.g `GigaSpeech`_) and use an external LM trained on that domain. + +.. HINT:: + + We recommend you to use a GPU for decoding. + +For illustration purpose, we will use a pre-trained ASR model from this `link `__. +If you want to train your model from scratch, please have a look at :ref:`non_streaming_librispeech_pruned_transducer_stateless`. + +As the initial step, let's download the pre-trained model. + +.. code-block:: bash + + $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29 + $ pushd icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp + $ git lfs pull --include "pretrained.pt" + $ ln -s pretrained.pt epoch-99.pt # create a symbolic link so that the checkpoint can be loaded + +As usual, we first test the model's performance without external LM. This can be done via the following command: + +.. code-block:: bash + + $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/ + $ ./pruned_transducer_stateless7_streaming/decode.py \ + --epoch 99 \ + --avg 1 \ + --use-averaged-model False \ + --exp-dir $exp_dir \ + --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model + --max-duration 600 \ + --decode-chunk-len 32 \ + --decoding-method modified_beam_search + +The following WERs are achieved on test-clean and test-other: + +.. code-block:: text + + $ For test-clean, WER of different settings are: + $ beam_size_4 3.11 best for test-clean + $ For test-other, WER of different settings are: + $ beam_size_4 7.93 best for test-other + +Now, we will try to improve the above WER numbers via external LM rescoring. We will download +a pre-trained LM from this `link `__. + +.. note:: + + This is an RNN LM trained on the LibriSpeech text corpus. So it might not be ideal for other corpus. + You may also train a RNN LM from scratch. Please refer to this `script `__ + for training a RNN LM and this `script `__ to train a transformer LM. + +.. code-block:: bash + + $ # download the external LM + $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm + $ # create a symbolic link so that the checkpoint can be loaded + $ pushd icefall-librispeech-rnn-lm/exp + $ git lfs pull --include "pretrained.pt" + $ ln -s pretrained.pt epoch-99.pt + $ popd + + +With the RNNLM available, we can rescore the n-best hypotheses generated from `modified_beam_search`. Here, +`n` should be the number of beams, i.e ``--beam-size``. The command for LM rescoring is +as follows. Note that the ``--decoding-method`` is set to `modified_beam_search_lm_rescore` and ``--use-shallow-fusion`` +is set to `False`. + +.. code-block:: bash + + $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp + $ lm_dir=./icefall-librispeech-rnn-lm/exp + $ lm_scale=0.43 + $ ./pruned_transducer_stateless7_streaming/decode.py \ + --epoch 99 \ + --avg 1 \ + --use-averaged-model False \ + --beam-size 4 \ + --exp-dir $exp_dir \ + --max-duration 600 \ + --decode-chunk-len 32 \ + --decoding-method modified_beam_search_lm_rescore \ + --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model + --use-shallow-fusion 0 \ + --lm-type rnn \ + --lm-exp-dir $lm_dir \ + --lm-epoch 99 \ + --lm-scale $lm_scale \ + --lm-avg 1 \ + --rnn-lm-embedding-dim 2048 \ + --rnn-lm-hidden-dim 2048 \ + --rnn-lm-num-layers 3 \ + --lm-vocab-size 500 + +.. code-block:: text + + $ For test-clean, WER of different settings are: + $ beam_size_4 2.93 best for test-clean + $ For test-other, WER of different settings are: + $ beam_size_4 7.6 best for test-other + +Great! We made some improvements! Increasing the size of the n-best hypotheses will further boost the performance, +see the following table: + +.. list-table:: WERs of LM rescoring with different beam sizes + :widths: 25 25 25 + :header-rows: 1 + + * - Beam size + - test-clean + - test-other + * - 4 + - 2.93 + - 7.6 + * - 8 + - 2.67 + - 7.11 + * - 12 + - 2.59 + - 6.86 + +In fact, we can also apply LODR (see :ref:`LODR`) when doing LM rescoring. To do so, we need to +download the bi-gram required by LODR: + +.. code-block:: bash + + $ # download the bi-gram + $ git lfs install + $ git clone https://huggingface.co/marcoyang/librispeech_bigram + $ pushd data/lang_bpe_500 + $ ln -s ../../librispeech_bigram/2gram.arpa . + $ popd + +Then we can performn LM rescoring + LODR by changing the decoding method to `modified_beam_search_lm_rescore_LODR`. + +.. note:: + + This decoding method requires the dependency of `kenlm `_. You can install it + via this command: `pip install https://github.com/kpu/kenlm/archive/master.zip`. + +.. code-block:: bash + + $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp + $ lm_dir=./icefall-librispeech-rnn-lm/exp + $ lm_scale=0.43 + $ ./pruned_transducer_stateless7_streaming/decode.py \ + --epoch 99 \ + --avg 1 \ + --use-averaged-model False \ + --beam-size 4 \ + --exp-dir $exp_dir \ + --max-duration 600 \ + --decode-chunk-len 32 \ + --decoding-method modified_beam_search_lm_rescore_LODR \ + --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model + --use-shallow-fusion 0 \ + --lm-type rnn \ + --lm-exp-dir $lm_dir \ + --lm-epoch 99 \ + --lm-scale $lm_scale \ + --lm-avg 1 \ + --rnn-lm-embedding-dim 2048 \ + --rnn-lm-hidden-dim 2048 \ + --rnn-lm-num-layers 3 \ + --lm-vocab-size 500 + +You should see the following WERs after executing the commands above: + +.. code-block:: text + + $ For test-clean, WER of different settings are: + $ beam_size_4 2.9 best for test-clean + $ For test-other, WER of different settings are: + $ beam_size_4 7.57 best for test-other + +It's slightly better than LM rescoring. If we further increase the beam size, we will see +further improvements from LM rescoring + LODR: + +.. list-table:: WERs of LM rescoring + LODR with different beam sizes + :widths: 25 25 25 + :header-rows: 1 + + * - Beam size + - test-clean + - test-other + * - 4 + - 2.9 + - 7.57 + * - 8 + - 2.63 + - 7.04 + * - 12 + - 2.52 + - 6.73 + +As mentioned earlier, LM rescoring is usually faster than shallow-fusion based methods. +Here, we benchmark the WERs and decoding speed of them: + +.. list-table:: LM-rescoring-based methods vs shallow-fusion-based methods (The numbers in each field is WER on test-clean, WER on test-other and decoding time on test-clean) + :widths: 25 25 25 25 + :header-rows: 1 + + * - Decoding method + - beam=4 + - beam=8 + - beam=12 + * - ``modified_beam_search`` + - 3.11/7.93; 132s + - 3.1/7.95; 177s + - 3.1/7.96; 210s + * - ``modified_beam_search_lm_shallow_fusion`` + - 2.77/7.08; 262s + - 2.62/6.65; 352s + - 2.58/6.65; 488s + * - ``modified_beam_search_LODR`` + - 2.61/6.74; 400s + - 2.45/6.38; 610s + - 2.4/6.23; 870s + * - ``modified_beam_search_lm_rescore`` + - 2.93/7.6; 156s + - 2.67/7.11; 203s + - 2.59/6.86; 255s + * - ``modified_beam_search_lm_rescore_LODR`` + - 2.9/7.57; 160s + - 2.63/7.04; 203s + - 2.52/6.73; 263s + +.. note:: + + Decoding is performed with a single 32G V100, we set ``--max-duration`` to 600. + Decoding time here is only for reference and it may vary. \ No newline at end of file diff --git a/docs/source/decoding-with-langugage-models/shallow-fusion.rst b/docs/source/decoding-with-langugage-models/shallow-fusion.rst new file mode 100644 index 000000000..0d2837372 --- /dev/null +++ b/docs/source/decoding-with-langugage-models/shallow-fusion.rst @@ -0,0 +1,176 @@ +.. _shallow_fusion: + +Shallow fusion for Transducer +================================= + +External language models (LM) are commonly used to improve WERs for E2E ASR models. +This tutorial shows you how to perform ``shallow fusion`` with an external LM +to improve the word-error-rate of a transducer model. + +.. note:: + + This tutorial is based on the recipe + `pruned_transducer_stateless7_streaming `_, + which is a streaming transducer model trained on `LibriSpeech`_. + However, you can easily apply shallow fusion to other recipes. + If you encounter any problems, please open an issue here `icefall `_. + +.. note:: + + For simplicity, the training and testing corpus in this tutorial is the same (`LibriSpeech`_). However, you can change the testing set + to any other domains (e.g `GigaSpeech`_) and use an external LM trained on that domain. + +.. HINT:: + + We recommend you to use a GPU for decoding. + +For illustration purpose, we will use a pre-trained ASR model from this `link `__. +If you want to train your model from scratch, please have a look at :ref:`non_streaming_librispeech_pruned_transducer_stateless`. + +As the initial step, let's download the pre-trained model. + +.. code-block:: bash + + $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29 + $ pushd icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp + $ git lfs pull --include "pretrained.pt" + $ ln -s pretrained.pt epoch-99.pt # create a symbolic link so that the checkpoint can be loaded + +To test the model, let's have a look at the decoding results without using LM. This can be done via the following command: + +.. code-block:: bash + + $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/ + $ ./pruned_transducer_stateless7_streaming/decode.py \ + --epoch 99 \ + --avg 1 \ + --use-averaged-model False \ + --exp-dir $exp_dir \ + --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model + --max-duration 600 \ + --decode-chunk-len 32 \ + --decoding-method modified_beam_search + +The following WERs are achieved on test-clean and test-other: + +.. code-block:: text + + $ For test-clean, WER of different settings are: + $ beam_size_4 3.11 best for test-clean + $ For test-other, WER of different settings are: + $ beam_size_4 7.93 best for test-other + +These are already good numbers! But we can further improve it by using shallow fusion with external LM. +Training a language model usually takes a long time, we can download a pre-trained LM from this `link `__. + +.. code-block:: bash + + $ # download the external LM + $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm + $ # create a symbolic link so that the checkpoint can be loaded + $ pushd icefall-librispeech-rnn-lm/exp + $ git lfs pull --include "pretrained.pt" + $ ln -s pretrained.pt epoch-99.pt + $ popd + +.. note:: + + This is an RNN LM trained on the LibriSpeech text corpus. So it might not be ideal for other corpus. + You may also train a RNN LM from scratch. Please refer to this `script `__ + for training a RNN LM and this `script `__ to train a transformer LM. + +To use shallow fusion for decoding, we can execute the following command: + +.. code-block:: bash + + $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp + $ lm_dir=./icefall-librispeech-rnn-lm/exp + $ lm_scale=0.29 + $ ./pruned_transducer_stateless7_streaming/decode.py \ + --epoch 99 \ + --avg 1 \ + --use-averaged-model False \ + --beam-size 4 \ + --exp-dir $exp_dir \ + --max-duration 600 \ + --decode-chunk-len 32 \ + --decoding-method modified_beam_search_lm_shallow_fusion \ + --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model + --use-shallow-fusion 1 \ + --lm-type rnn \ + --lm-exp-dir $lm_dir \ + --lm-epoch 99 \ + --lm-scale $lm_scale \ + --lm-avg 1 \ + --rnn-lm-embedding-dim 2048 \ + --rnn-lm-hidden-dim 2048 \ + --rnn-lm-num-layers 3 \ + --lm-vocab-size 500 + +Note that we set ``--decoding-method modified_beam_search_lm_shallow_fusion`` and ``--use-shallow-fusion True`` +to use shallow fusion. ``--lm-type`` specifies the type of neural LM we are going to use, you can either choose +between ``rnn`` or ``transformer``. The following three arguments are associated with the rnn: + +- ``--rnn-lm-embedding-dim`` + The embedding dimension of the RNN LM + +- ``--rnn-lm-hidden-dim`` + The hidden dimension of the RNN LM + +- ``--rnn-lm-num-layers`` + The number of RNN layers in the RNN LM. + + +The decoding result obtained with the above command are shown below. + +.. code-block:: text + + $ For test-clean, WER of different settings are: + $ beam_size_4 2.77 best for test-clean + $ For test-other, WER of different settings are: + $ beam_size_4 7.08 best for test-other + +The improvement of shallow fusion is very obvious! The relative WER reduction on test-other is around 10.5%. +A few parameters can be tuned to further boost the performance of shallow fusion: + +- ``--lm-scale`` + + Controls the scale of the LM. If too small, the external language model may not be fully utilized; if too large, + the LM score may dominant during decoding, leading to bad WER. A typical value of this is around 0.3. + +- ``--beam-size`` + + The number of active paths in the search beam. It controls the trade-off between decoding efficiency and accuracy. + +Here, we also show how `--beam-size` effect the WER and decoding time: + +.. list-table:: WERs and decoding time (on test-clean) of shallow fusion with different beam sizes + :widths: 25 25 25 25 + :header-rows: 1 + + * - Beam size + - test-clean + - test-other + - Decoding time on test-clean (s) + * - 4 + - 2.77 + - 7.08 + - 262 + * - 8 + - 2.62 + - 6.65 + - 352 + * - 12 + - 2.58 + - 6.65 + - 488 + +As we see, a larger beam size during shallow fusion improves the WER, but is also slower. + + + + + + + + diff --git a/docs/source/index.rst b/docs/source/index.rst index 8d76eb68b..a7d365a15 100644 --- a/docs/source/index.rst +++ b/docs/source/index.rst @@ -34,3 +34,8 @@ speech recognition recipes using `k2 `_. contributing/index huggingface/index + +.. toctree:: + :maxdepth: 2 + + decoding-with-langugage-models/index \ No newline at end of file diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst b/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst index ea9f350cd..2e8d0893a 100644 --- a/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst +++ b/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst @@ -1,7 +1,7 @@ Distillation with HuBERT ======================== -This tutorial shows you how to perform knowledge distillation in `icefall`_ +This tutorial shows you how to perform knowledge distillation in `icefall `_ with the `LibriSpeech`_ dataset. The distillation method used here is called "Multi Vector Quantization Knowledge Distillation" (MVQ-KD). Please have a look at our paper `Predicting Multi-Codebook Vector Quantization Indexes for Knowledge Distillation `_ @@ -13,7 +13,7 @@ for more details about MVQ-KD. `pruned_transducer_stateless4 `_. Currently, we only implement MVQ-KD in this recipe. However, MVQ-KD is theoretically applicable to all recipes with only minor changes needed. Feel free to try out MVQ-KD in different recipes. If you - encounter any problems, please open an issue here `icefall `_. + encounter any problems, please open an issue here `icefall `__. .. note:: @@ -217,7 +217,7 @@ the following command. --exp-dir $exp_dir \ --enable-distillation True -You should get similar results as `here `_. +You should get similar results as `here `__. That's all! Feel free to experiment with your own setups and report your results. -If you encounter any problems during training, please open up an issue `here `_. +If you encounter any problems during training, please open up an issue `here `__. diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.rst b/docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.rst index 42fd3df77..1bc1dd984 100644 --- a/docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.rst +++ b/docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.rst @@ -8,10 +8,10 @@ with the `LibriSpeech `_ dataset. .. Note:: - The tutorial is suitable for `pruned_transducer_stateless `_, - `pruned_transducer_stateless2 `_, - `pruned_transducer_stateless4 `_, - `pruned_transducer_stateless5 `_, + The tutorial is suitable for `pruned_transducer_stateless `__, + `pruned_transducer_stateless2 `__, + `pruned_transducer_stateless4 `__, + `pruned_transducer_stateless5 `__, We will take pruned_transducer_stateless4 as an example in this tutorial. .. HINT:: @@ -237,7 +237,7 @@ them, please modify ``./pruned_transducer_stateless4/train.py`` directly. .. NOTE:: - The options for `pruned_transducer_stateless5 `_ are a little different from + The options for `pruned_transducer_stateless5 `__ are a little different from other recipes. It allows you to configure ``--num-encoder-layers``, ``--dim-feedforward``, ``--nhead``, ``--encoder-dim``, ``--decoder-dim``, ``--joiner-dim`` from commandline, so that you can train models with different size with pruned_transducer_stateless5. @@ -529,13 +529,13 @@ Download pretrained models If you don't want to train from scratch, you can download the pretrained models by visiting the following links: - - `pruned_transducer_stateless `_ + - `pruned_transducer_stateless `__ - - `pruned_transducer_stateless2 `_ + - `pruned_transducer_stateless2 `__ - - `pruned_transducer_stateless4 `_ + - `pruned_transducer_stateless4 `__ - - `pruned_transducer_stateless5 `_ + - `pruned_transducer_stateless5 `__ See ``_ for the details of the above pretrained models diff --git a/docs/source/recipes/Streaming-ASR/introduction.rst b/docs/source/recipes/Streaming-ASR/introduction.rst index e1382e77d..ac77a51d1 100644 --- a/docs/source/recipes/Streaming-ASR/introduction.rst +++ b/docs/source/recipes/Streaming-ASR/introduction.rst @@ -45,9 +45,9 @@ the input features. We have three variants of Emformer models in ``icefall``. - - ``pruned_stateless_emformer_rnnt2`` using Emformer from torchaudio, see `LibriSpeech recipe `_. + - ``pruned_stateless_emformer_rnnt2`` using Emformer from torchaudio, see `LibriSpeech recipe `__. - ``conv_emformer_transducer_stateless`` using ConvEmformer implemented by ourself. Different from the Emformer in torchaudio, ConvEmformer has a convolution in each layer and uses the mechanisms in our reworked conformer model. - See `LibriSpeech recipe `_. + See `LibriSpeech recipe `__. - ``conv_emformer_transducer_stateless2`` using ConvEmformer implemented by ourself. The only difference from the above one is that it uses a simplified memory bank. See `LibriSpeech recipe `_. diff --git a/docs/source/recipes/Streaming-ASR/librispeech/pruned_transducer_stateless.rst b/docs/source/recipes/Streaming-ASR/librispeech/pruned_transducer_stateless.rst index de7102ba8..2ca70bcf3 100644 --- a/docs/source/recipes/Streaming-ASR/librispeech/pruned_transducer_stateless.rst +++ b/docs/source/recipes/Streaming-ASR/librispeech/pruned_transducer_stateless.rst @@ -6,10 +6,10 @@ with the `LibriSpeech `_ dataset. .. Note:: - The tutorial is suitable for `pruned_transducer_stateless `_, - `pruned_transducer_stateless2 `_, - `pruned_transducer_stateless4 `_, - `pruned_transducer_stateless5 `_, + The tutorial is suitable for `pruned_transducer_stateless `__, + `pruned_transducer_stateless2 `__, + `pruned_transducer_stateless4 `__, + `pruned_transducer_stateless5 `__, We will take pruned_transducer_stateless4 as an example in this tutorial. .. HINT:: @@ -264,7 +264,7 @@ them, please modify ``./pruned_transducer_stateless4/train.py`` directly. .. NOTE:: - The options for `pruned_transducer_stateless5 `_ are a little different from + The options for `pruned_transducer_stateless5 `__ are a little different from other recipes. It allows you to configure ``--num-encoder-layers``, ``--dim-feedforward``, ``--nhead``, ``--encoder-dim``, ``--decoder-dim``, ``--joiner-dim`` from commandline, so that you can train models with different size with pruned_transducer_stateless5. diff --git a/docs/source/recipes/Streaming-ASR/librispeech/zipformer_transducer.rst b/docs/source/recipes/Streaming-ASR/librispeech/zipformer_transducer.rst index f0e8961d7..8b75473c6 100644 --- a/docs/source/recipes/Streaming-ASR/librispeech/zipformer_transducer.rst +++ b/docs/source/recipes/Streaming-ASR/librispeech/zipformer_transducer.rst @@ -6,7 +6,7 @@ with the `LibriSpeech `_ dataset. .. Note:: - The tutorial is suitable for `pruned_transducer_stateless7_streaming `_, + The tutorial is suitable for `pruned_transducer_stateless7_streaming `__, .. HINT:: @@ -642,7 +642,7 @@ Download pretrained models If you don't want to train from scratch, you can download the pretrained models by visiting the following links: - - `pruned_transducer_stateless7_streaming `_ + - `pruned_transducer_stateless7_streaming `__ See ``_ for the details of the above pretrained models diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/train.py b/egs/aishell/ASR/pruned_transducer_stateless7/train.py index ef536c035..cbb7db086 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7/train.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7/train.py @@ -240,7 +240,7 @@ def get_parser(): parser.add_argument( "--exp-dir", type=str, - default="pruned_transducer_stateless3/exp", + default="pruned_transducer_stateless7/exp", help="""The experiment dir. It specifies the directory where all training related files, e.g., checkpoints, log, etc, are saved diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/train2.py b/egs/aishell/ASR/pruned_transducer_stateless7/train2.py index fb35a6c95..c30f6f960 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7/train2.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7/train2.py @@ -243,7 +243,7 @@ def get_parser(): parser.add_argument( "--exp-dir", type=str, - default="pruned_transducer_stateless3/exp", + default="pruned_transducer_stateless7/exp", help="""The experiment dir. It specifies the directory where all training related files, e.g., checkpoints, log, etc, are saved diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/decode.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/decode.py index fcb0ebc4e..da9000164 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/decode.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/decode.py @@ -397,7 +397,6 @@ def decode_one_batch( beam=params.beam, max_contexts=params.max_contexts, max_states=params.max_states, - subtract_ilme=True, ilme_scale=params.ilme_scale, ) for hyp in hyp_tokens: diff --git a/egs/ami/SURT/README.md b/egs/ami/SURT/README.md new file mode 100644 index 000000000..74a8ba014 --- /dev/null +++ b/egs/ami/SURT/README.md @@ -0,0 +1,156 @@ +# Introduction + +This is a multi-talker ASR recipe for the AMI and ICSI datasets. We train a Streaming +Unmixing and Recognition Transducer (SURT) model for the task. + +Please refer to the `egs/libricss/SURT` recipe README for details about the task and the +model. + +## Description of the recipe + +### Pre-requisites + +The recipes in this directory need the following packages to be installed: + +- [meeteval](https://github.com/fgnt/meeteval) +- [einops](https://github.com/arogozhnikov/einops) + +Additionally, we initialize the model with the pre-trained model from the LibriCSS recipe. +Please download this checkpoint (see below) or train the LibriCSS recipe first. + +### Training + +To train the model, run the following from within `egs/ami/SURT`: + +```bash +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +python dprnn_zipformer/train.py \ + --use-fp16 True \ + --exp-dir dprnn_zipformer/exp/surt_base \ + --world-size 4 \ + --max-duration 500 \ + --max-duration-valid 250 \ + --max-cuts 200 \ + --num-buckets 50 \ + --num-epochs 30 \ + --enable-spec-aug True \ + --enable-musan False \ + --ctc-loss-scale 0.2 \ + --heat-loss-scale 0.2 \ + --base-lr 0.004 \ + --model-init-ckpt exp/libricss_base.pt \ + --chunk-width-randomization True \ + --num-mask-encoder-layers 4 \ + --num-encoder-layers 2,2,2,2,2 +``` + +The above is for SURT-base (~26M). For SURT-large (~38M), use: + +```bash + --model-init-ckpt exp/libricss_large.pt \ + --num-mask-encoder-layers 6 \ + --num-encoder-layers 2,4,3,2,4 \ + --model-init-ckpt exp/zipformer_large.pt \ +``` + +**NOTE:** You may need to decrease the `--max-duration` for SURT-large to avoid OOM. + +### Adaptation + +The training step above only trains on simulated mixtures. For best results, we also +adapt the final model on the AMI+ICSI train set. For this, run the following from within +`egs/ami/SURT`: + +```bash +export CUDA_VISIBLE_DEVICES="0" + +python dprnn_zipformer/train_adapt.py \ + --use-fp16 True \ + --exp-dir dprnn_zipformer/exp/surt_base_adapt \ + --world-size 4 \ + --max-duration 500 \ + --max-duration-valid 250 \ + --max-cuts 200 \ + --num-buckets 50 \ + --num-epochs 8 \ + --lr-epochs 2 \ + --enable-spec-aug True \ + --enable-musan False \ + --ctc-loss-scale 0.2 \ + --base-lr 0.0004 \ + --model-init-ckpt dprnn_zipformer/exp/surt_base/epoch-30.pt \ + --chunk-width-randomization True \ + --num-mask-encoder-layers 4 \ + --num-encoder-layers 2,2,2,2,2 +``` + +For SURT-large, use the following config: + +```bash + --num-mask-encoder-layers 6 \ + --num-encoder-layers 2,4,3,2,4 \ + --model-init-ckpt dprnn_zipformer/exp/surt_large/epoch-30.pt \ + --num-epochs 15 \ + --lr-epochs 4 \ +``` + + +### Decoding + +To decode the model, run the following from within `egs/ami/SURT`: + +#### Greedy search + +```bash +export CUDA_VISIBLE_DEVICES="0" + +python dprnn_zipformer/decode.py \ + --epoch 20 --avg 1 --use-averaged-model False \ + --exp-dir dprnn_zipformer/exp/surt_base_adapt \ + --max-duration 250 \ + --decoding-method greedy_search +``` + +#### Beam search + +```bash +python dprnn_zipformer/decode.py \ + --epoch 20 --avg 1 --use-averaged-model False \ + --exp-dir dprnn_zipformer/exp/surt_base_adapt \ + --max-duration 250 \ + --decoding-method modified_beam_search \ + --beam-size 4 +``` + +## Results (using beam search) + +**AMI** + +| Model | IHM-Mix | SDM | MDM | +|------------|:-------:|:----:|:----:| +| SURT-base | 39.8 | 65.4 | 46.6 | +| + adapt | 37.4 | 46.9 | 43.7 | +| SURT-large | 36.8 | 62.5 | 44.4 | +| + adapt | **35.1** | **44.6** | **41.4** | + +**ICSI** + +| Model | IHM-Mix | SDM | +|------------|:-------:|:----:| +| SURT-base | 28.3 | 60.0 | +| + adapt | 26.3 | 33.9 | +| SURT-large | 27.8 | 59.7 | +| + adapt | **24.4** | **32.3** | + +## Pre-trained models and logs + +* LibriCSS pre-trained model (for initialization): [base](https://huggingface.co/desh2608/icefall-surt-libricss-dprnn-zipformer/tree/main/exp/surt_base) [large](https://huggingface.co/desh2608/icefall-surt-libricss-dprnn-zipformer/tree/main/exp/surt_large) + +* Pre-trained models: + +* Training logs: + - surt_base: + - surt_base_adapt: + - surt_large: + - surt_large_adapt: diff --git a/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py b/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py new file mode 100644 index 000000000..ec8106bc3 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py @@ -0,0 +1,399 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Callable, Dict, List, Optional + +import torch +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures + CutMix, + DynamicBucketingSampler, + K2SurtDataset, + PrecomputedFeatures, + SimpleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import OnTheFlyFeatures +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class AmiAsrDataModule: + """ + DataModule for k2 SURT experiments. + It assumes there is always one train and valid dataloader, + but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + and test-other). + + It contains all the common data pipeline modules used in ASR + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + - augmentation, + - on-the-fly feature extraction + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/manifests"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--max-duration-valid", + type=int, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--max-cuts", + type=int, + default=100, + help="Maximum number of cuts in a single batch. You can " + "reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--on-the-fly-feats", + type=str2bool, + default=False, + help=( + "When enabled, use on-the-fly cut mixing and feature " + "extraction. Will drop existing precomputed feature manifests " + "if available." + ), + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + + group.add_argument( + "--enable-spec-aug", + type=str2bool, + default=True, + help="When enabled, use SpecAugment for training dataset.", + ) + + group.add_argument( + "--spec-aug-time-warp-factor", + type=int, + default=80, + help="Used only when --enable-spec-aug is True. " + "It specifies the factor for time warping in SpecAugment. " + "Larger values mean more warping. " + "A value less than 1 means to disable time warp.", + ) + + group.add_argument( + "--enable-musan", + type=str2bool, + default=True, + help="When enabled, select noise from MUSAN and mix it" + "with training dataset. ", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + sampler_state_dict: Optional[Dict[str, Any]] = None, + sources: bool = False, + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + logging.info("About to get Musan cuts") + cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") + transforms.append( + CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True) + ) + else: + logging.info("Disable MUSAN") + + input_transforms = [] + if self.args.enable_spec_aug: + logging.info("Enable SpecAugment") + logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}") + # Set the value of num_frame_masks according to Lhotse's version. + # In different Lhotse's versions, the default of num_frame_masks is + # different. + num_frame_masks = 10 + num_frame_masks_parameter = inspect.signature( + SpecAugment.__init__ + ).parameters["num_frame_masks"] + if num_frame_masks_parameter.default == 1: + num_frame_masks = 2 + logging.info(f"Num frame mask: {num_frame_masks}") + input_transforms.append( + SpecAugment( + time_warp_factor=self.args.spec_aug_time_warp_factor, + num_frame_masks=num_frame_masks, + features_mask_size=27, + num_feature_masks=2, + frames_mask_size=100, + ) + ) + else: + logging.info("Disable SpecAugment") + + logging.info("About to create train dataset") + train = K2SurtDataset( + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + cut_transforms=transforms, + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + return_sources=sources, + strict=False, + ) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + quadratic_duration=30.0, + max_cuts=self.args.max_cuts, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SingleCutSampler.") + train_sampler = SimpleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + max_cuts=self.args.max_cuts, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + return train_dl + + def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: + transforms = [] + + logging.info("About to create dev dataset") + validate = K2SurtDataset( + input_strategy=OnTheFlyFeatures( + OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + ) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + cut_transforms=transforms, + return_cuts=self.args.return_cuts, + return_sources=False, + strict=False, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration_valid, + quadratic_duration=30.0, + max_cuts=self.args.max_cuts, + shuffle=False, + ) + logging.info("About to create dev dataloader") + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SurtDataset( + input_strategy=OnTheFlyFeatures( + OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + ) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + return_cuts=self.args.return_cuts, + return_sources=False, + strict=False, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration_valid, + max_cuts=self.args.max_cuts, + shuffle=False, + ) + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + return test_dl + + @lru_cache() + def aimix_train_cuts( + self, + rvb_affix: str = "clean", + sources: bool = True, + ) -> CutSet: + logging.info("About to get train cuts") + source_affix = "_sources" if sources else "" + cs = load_manifest_lazy( + self.args.manifest_dir / f"cuts_train_{rvb_affix}{source_affix}.jsonl.gz" + ) + cs = cs.filter(lambda c: c.duration >= 1.0 and c.duration <= 30.0) + return cs + + @lru_cache() + def train_cuts( + self, + ) -> CutSet: + logging.info("About to get train cuts") + return load_manifest_lazy( + self.args.manifest_dir / "cuts_train_ami_icsi.jsonl.gz" + ) + + @lru_cache() + def ami_cuts(self, split: str = "dev", type: str = "sdm") -> CutSet: + logging.info(f"About to get AMI {split} {type} cuts") + return load_manifest_lazy( + self.args.manifest_dir / f"cuts_ami-{type}_{split}.jsonl.gz" + ) + + @lru_cache() + def icsi_cuts(self, split: str = "dev", type: str = "sdm") -> CutSet: + logging.info(f"About to get ICSI {split} {type} cuts") + return load_manifest_lazy( + self.args.manifest_dir / f"cuts_icsi-{type}_{split}.jsonl.gz" + ) diff --git a/egs/ami/SURT/dprnn_zipformer/beam_search.py b/egs/ami/SURT/dprnn_zipformer/beam_search.py new file mode 120000 index 000000000..581b29833 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/beam_search.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/beam_search.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/decode.py b/egs/ami/SURT/dprnn_zipformer/decode.py new file mode 100755 index 000000000..d1a1eddc9 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/decode.py @@ -0,0 +1,622 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./dprnn_zipformer/decode.py \ + --epoch 20 \ + --avg 1 \ + --use-averaged-model false \ + --exp-dir ./dprnn_zipformer/exp_adapt \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./dprnn_zipformer/decode.py \ + --epoch 20 \ + --avg 1 \ + --use-averaged-model false \ + --exp-dir ./dprnn_zipformer/exp_adapt \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./dprnn_zipformer/decode.py \ + --epoch 20 \ + --avg 1 \ + --use-averaged-model false \ + --exp-dir ./dprnn_zipformer/exp_adapt \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 +""" + + +import argparse +import logging +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import AmiAsrDataModule +from beam_search import ( + beam_search, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from lhotse.utils import EPSILON +from train import add_model_arguments, get_params, get_surt_model + +from icefall import LmScorer, NgramLm +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_surt_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=20, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=1, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="dprnn_zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + feature_lens = batch["input_lens"].to(device) + + # Apply the mask encoder + B, T, F = feature.shape + processed = model.mask_encoder(feature) # B,T,F*num_channels + masks = processed.view(B, T, F, params.num_channels).unbind(dim=-1) + x_masked = [feature * m for m in masks] + + # Recognition + # Stack the inputs along the batch axis + h = torch.cat(x_masked, dim=0) + h_lens = torch.cat([feature_lens for _ in range(params.num_channels)], dim=0) + encoder_out, encoder_out_lens = model.encoder(x=h, x_lens=h_lens) + + if model.joint_encoder_layer is not None: + encoder_out = model.joint_encoder_layer(encoder_out) + + def _group_channels(hyps: List[str]) -> List[List[str]]: + """ + Currently we have a batch of size M*B, where M is the number of + channels and B is the batch size. We need to group the hypotheses + into B groups, each of which contains M hypotheses. + + Example: + hyps = ['a1', 'b1', 'c1', 'a2', 'b2', 'c2'] + _group_channels(hyps) = [['a1', 'a2'], ['b1', 'b2'], ['c1', 'c2']] + """ + assert len(hyps) == B * params.num_channels + out_hyps = [] + for i in range(B): + out_hyps.append(hyps[i::B]) + return out_hyps + + hyps = [] + if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp)) + + if params.decoding_method == "greedy_search": + return {"greedy_search": _group_channels(hyps)} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: _group_channels(hyps)} + else: + return {f"beam_size_{params.beam_size}": _group_channels(hyps)} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + cut_ids = [cut.id for cut in batch["cuts"]] + cuts_batch = batch["cuts"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + for cut_id, hyp_words in zip(cut_ids, hyps): + # Reference is a list of supervision texts sorted by start time. + ref_words = [ + s.text.strip() + for s in sorted( + cuts_batch[cut_id].supervisions, key=lambda s: s.start + ) + ] + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(cut_ids) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_surt_error_stats( + f, + f"{test_set_name}-{key}", + results, + enable_log=True, + num_channels=params.num_channels, + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LmScorer.add_arguments(parser) + AmiAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "modified_beam_search", + ), f"Decoding method {params.decoding_method} is not supported." + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_surt_model(params) + assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, ( + model.encoder.decode_chunk_size, + params.decode_chunk_len, + ) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + ami = AmiAsrDataModule(args) + + # NOTE(@desh2608): we filter segments longer than 120s to avoid OOM errors in decoding. + # However, 99.9% of the segments are shorter than 120s, so this should not + # substantially affect the results. In future, we will implement an overlapped + # inference method to avoid OOM errors. + + test_sets = {} + for split in ["dev", "test"]: + for type in ["ihm-mix", "sdm", "mdm8-bf"]: + test_sets[f"ami-{split}_{type}"] = ( + ami.ami_cuts(split=split, type=type) + .trim_to_supervision_groups(max_pause=0.0) + .filter(lambda c: 0.1 < c.duration < 120.0) + .to_eager() + ) + + for split in ["dev", "test"]: + for type in ["ihm-mix", "sdm"]: + test_sets[f"icsi-{split}_{type}"] = ( + ami.icsi_cuts(split=split, type=type) + .trim_to_supervision_groups(max_pause=0.0) + .filter(lambda c: 0.1 < c.duration < 120.0) + .to_eager() + ) + + for test_set, test_cuts in test_sets.items(): + test_dl = ami.test_dataloaders(test_cuts) + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/ami/SURT/dprnn_zipformer/decoder.py b/egs/ami/SURT/dprnn_zipformer/decoder.py new file mode 120000 index 000000000..c34865c25 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/decoder.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/decoder.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/dprnn.py b/egs/ami/SURT/dprnn_zipformer/dprnn.py new file mode 120000 index 000000000..8918beb32 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/dprnn.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/dprnn.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/encoder_interface.py b/egs/ami/SURT/dprnn_zipformer/encoder_interface.py new file mode 120000 index 000000000..0ba945d0f --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/export.py b/egs/ami/SURT/dprnn_zipformer/export.py new file mode 120000 index 000000000..3deae4471 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/export.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/export.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/joiner.py b/egs/ami/SURT/dprnn_zipformer/joiner.py new file mode 120000 index 000000000..79fbe8769 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/joiner.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/joiner.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/model.py b/egs/ami/SURT/dprnn_zipformer/model.py new file mode 120000 index 000000000..ae8c65c99 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/model.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/model.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/optim.py b/egs/ami/SURT/dprnn_zipformer/optim.py new file mode 120000 index 000000000..366d0f7a2 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/optim.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/optim.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/scaling.py b/egs/ami/SURT/dprnn_zipformer/scaling.py new file mode 120000 index 000000000..f11d49d77 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/scaling.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/scaling.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/scaling_converter.py b/egs/ami/SURT/dprnn_zipformer/scaling_converter.py new file mode 120000 index 000000000..1533cbe0e --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/test_model.py b/egs/ami/SURT/dprnn_zipformer/test_model.py new file mode 120000 index 000000000..1259849e0 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/test_model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless7_streaming/test_model.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/train.py b/egs/ami/SURT/dprnn_zipformer/train.py new file mode 100755 index 000000000..cd5fafc34 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/train.py @@ -0,0 +1,1420 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +cd egs/ami/SURT/ +./prepare.sh + +./dprnn_zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir dprnn_zipformer/exp \ + --max-duration 650 +""" + +import argparse +import copy +import logging +import warnings +from itertools import chain +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import AmiAsrDataModule +from decoder import Decoder +from dprnn import DPRNN +from einops.layers.torch import Rearrange +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import LOG_EPSILON, fix_random_seed +from model import SURT +from optim import Eden, ScaledAdam +from scaling import ScaledLinear, ScaledLSTM +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-mask-encoder-layers", + type=int, + default=4, + help="Number of layers in the DPRNN based mask encoder.", + ) + + parser.add_argument( + "--mask-encoder-dim", + type=int, + default=256, + help="Hidden dimension of the LSTM blocks in DPRNN.", + ) + + parser.add_argument( + "--mask-encoder-segment-size", + type=int, + default=32, + help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the " + "decode-chunk-length of the zipformer encoder.", + ) + + parser.add_argument( + "--chunk-width-randomization", + type=bool, + default=False, + help="Whether to randomize the chunk width in DPRNN.", + ) + + # Zipformer config is based on: + # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740 + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,2,2,2", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="768,768,768,768,768", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="256,256,256,256,256", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="192,192,192,192,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--use-joint-encoder-layer", + type=str, + default="lstm", + choices=["linear", "lstm", "none"], + help="Whether to use a joint layer to combine all branches.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--short-chunk-size", + type=int, + default=50, + help="""Chunk length of dynamic training, the chunk size would be either + max sequence length of current batch or uniformly sampled from (1, short_chunk_size). + """, + ) + + parser.add_argument( + "--num-left-chunks", + type=int, + default=4, + help="How many left context can be seen in chunks when calculating attention.", + ) + + parser.add_argument( + "--decode-chunk-len", + type=int, + default=32, + help="The chunk size for decoding (in frames before subsampling)", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="conv_lstm_transducer_stateless_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--model-init-ckpt", + type=str, + default=None, + help="""The model checkpoint to initialize the model (either full or part). + If not specified, the model is randomly initialized. + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.004, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--heat-loss-scale", + type=float, + default=0.2, + help="Scale for HEAT loss on separated sources.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=1, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=100, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 2000, + # parameters for SURT + "num_channels": 2, + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed + # parameters for Noam + "model_warm_step": 5000, # arg given to model, not for lrate + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "env_info": get_env_info(), + } + ) + + return params + + +def get_mask_encoder_model(params: AttributeDict) -> nn.Module: + mask_encoder = DPRNN( + feature_dim=params.feature_dim, + input_size=params.mask_encoder_dim, + hidden_size=params.mask_encoder_dim, + output_size=params.feature_dim * params.num_channels, + segment_size=params.mask_encoder_segment_size, + num_blocks=params.num_mask_encoder_layers, + chunk_width_randomization=params.chunk_width_randomization, + ) + return mask_encoder + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + num_left_chunks=params.num_left_chunks, + short_chunk_size=params.short_chunk_size, + decode_chunk_size=params.decode_chunk_len // 2, + ) + return encoder + + +def get_joint_encoder_layer(params: AttributeDict) -> nn.Module: + class TakeFirst(nn.Module): + def forward(self, x): + return x[0] + + if params.use_joint_encoder_layer == "linear": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + nn.Linear( + params.num_channels * encoder_dim, params.num_channels * encoder_dim + ), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "lstm": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + ScaledLSTM( + input_size=params.num_channels * encoder_dim, + hidden_size=params.num_channels * encoder_dim, + num_layers=1, + bias=True, + batch_first=True, + dropout=0.0, + bidirectional=False, + ), + TakeFirst(), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "none": + joint_layer = None + else: + raise ValueError( + f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}" + ) + return joint_layer + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_surt_model( + params: AttributeDict, +) -> nn.Module: + mask_encoder = get_mask_encoder_model(params) + encoder = get_encoder_model(params) + joint_layer = get_joint_encoder_layer(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = SURT( + mask_encoder=mask_encoder, + encoder=encoder, + joint_encoder_layer=joint_layer, + decoder=decoder, + joiner=joiner, + num_channels=params.num_channels, + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_heat_loss(x_masked, batch, num_channels=2) -> Tensor: + """ + Compute HEAT loss for separated sources using the output of mask encoder. + Args: + x_masked: + The output of mask encoder. It is a tensor of shape (B, T, C). + batch: + A batch of data. See `lhotse.dataset.K2SurtDatasetWithSources()` + for the content in it. + num_channels: + The number of output branches in the SURT model. + """ + B, T, D = x_masked[0].shape + device = x_masked[0].device + + # Create training targets for each channel. + targets = [] + for i in range(num_channels): + target = torch.ones_like(x_masked[i]) * LOG_EPSILON + targets.append(target) + + source_feats = batch["source_feats"] + source_boundaries = batch["source_boundaries"] + input_lens = batch["input_lens"].to(device) + # Assign sources to channels based on the HEAT criteria + for b in range(B): + cut_source_feats = source_feats[b] + cut_source_boundaries = source_boundaries[b] + last_seg_end = [0 for _ in range(num_channels)] + for source_feat, (start, end) in zip(cut_source_feats, cut_source_boundaries): + assigned = False + end = min(end, T) + source_feat = source_feat[: end - start, :] + for i in range(num_channels): + if start >= last_seg_end[i]: + targets[i][b, start:end, :] += source_feat.to(device) + last_seg_end[i] = max(end, last_seg_end[i]) + assigned = True + break + if not assigned: + min_end_channel = last_seg_end.index(min(last_seg_end)) + targets[min_end_channel][b, start:end, :] += source_feat.to(device) + last_seg_end[min_end_channel] = max(end, last_seg_end[min_end_channel]) + + # Get padding mask based on input lengths + pad_mask = torch.arange(T, device=device).expand(B, T) > input_lens.unsqueeze(1) + pad_mask = pad_mask.unsqueeze(-1) + + # Compute masked loss for each channel + losses = torch.zeros((num_channels, B, T, D), device=device) + for i in range(num_channels): + loss = nn.functional.mse_loss(x_masked[i], targets[i], reduction="none") + # Apply padding mask to loss + loss.masked_fill_(pad_mask, 0) + losses[i] = loss + + # loss: C x B x T x D. pad_mask: B x T x 1 + # We want to compute loss for each item in the batch. Each item has loss given + # by the sum over C, and average over T and D. For T, we need to use the padding. + loss = losses.sum(0).mean(-1).sum(-1) / batch["input_lens"].to(device) + return loss + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"].to(device) + feature_lens = batch["input_lens"].to(device) + + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + + # The dataloader returns text as a list of cuts, each of which is a list of channel + # text. We flatten this to a list where all channels are together, i.e., it looks like + # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2]. + text = [val for tup in zip(*batch["text"]) for val in tup] + assert len(text) == len(feature) * params.num_channels + + # Convert all channel texts to token IDs and create a ragged tensor. + y = sp.encode(text, out_type=int) + y = k2.RaggedTensor(y).to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.model_warm_step + + with torch.set_grad_enabled(is_training): + (simple_loss, pruned_loss, ctc_loss, x_masked) = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + reduction="none", + subsampling_factor=params.subsampling_factor, + ) + simple_loss_is_finite = torch.isfinite(simple_loss) + pruned_loss_is_finite = torch.isfinite(pruned_loss) + ctc_loss_is_finite = torch.isfinite(ctc_loss) + + # Compute HEAT loss + if is_training and params.heat_loss_scale > 0.0: + heat_loss = compute_heat_loss( + x_masked, batch, num_channels=params.num_channels + ) + else: + heat_loss = torch.tensor(0.0, device=device) + + heat_loss_is_finite = torch.isfinite(heat_loss) + is_finite = ( + simple_loss_is_finite + & pruned_loss_is_finite + & ctc_loss_is_finite + & heat_loss_is_finite + ) + if not torch.all(is_finite): + logging.info( + "Not all losses are finite!\n" + f"simple_losses: {simple_loss}\n" + f"pruned_losses: {pruned_loss}\n" + f"ctc_losses: {ctc_loss}\n" + f"heat_losses: {heat_loss}\n" + ) + display_and_save_batch(batch, params=params, sp=sp) + simple_loss = simple_loss[simple_loss_is_finite] + pruned_loss = pruned_loss[pruned_loss_is_finite] + ctc_loss = ctc_loss[ctc_loss_is_finite] + heat_loss = heat_loss[heat_loss_is_finite] + + # If either all simple_loss or pruned_loss is inf or nan, + # we stop the training process by raising an exception + if ( + torch.all(~simple_loss_is_finite) + or torch.all(~pruned_loss_is_finite) + or torch.all(~ctc_loss_is_finite) + or torch.all(~heat_loss_is_finite) + ): + raise ValueError( + "There are too many utterances in this batch " + "leading to inf or nan losses." + ) + + simple_loss_sum = simple_loss.sum() + pruned_loss_sum = pruned_loss.sum() + ctc_loss_sum = ctc_loss.sum() + heat_loss_sum = heat_loss.sum() + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss = ( + simple_loss_scale * simple_loss_sum + + pruned_loss_scale * pruned_loss_sum + + params.ctc_loss_scale * ctc_loss_sum + + params.heat_loss_scale * heat_loss_sum + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + # info["frames"] is an approximate number for two reasons: + # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2 + # (2) If some utterances in the batch lead to inf/nan loss, they + # are filtered out. + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances` # noqa + info["utterances"] = feature.size(0) + # averaged input duration in frames over utterances + info["utt_duration"] = feature_lens.sum().item() + # averaged padding proportion over utterances + info["utt_pad_proportion"] = ( + ((feature.size(1) - feature_lens) / feature.size(1)).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss_sum.detach().cpu().item() + info["pruned_loss"] = pruned_loss_sum.detach().cpu().item() + if params.ctc_loss_scale > 0.0: + info["ctc_loss"] = ctc_loss_sum.detach().cpu().item() + if params.heat_loss_scale > 0.0: + info["heat_loss"] = heat_loss_sum.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + torch.cuda.empty_cache() + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = batch["inputs"].shape[0] + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_surt_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + + if checkpoints is None and params.model_init_ckpt is not None: + logging.info( + f"Initializing model with checkpoint from {params.model_init_ckpt}" + ) + init_ckpt = torch.load(params.model_init_ckpt, map_location=device) + model.load_state_dict(init_ckpt["model"], strict=False) + + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + diagnostic = diagnostics.attach_diagnostics(model) + + ami = AmiAsrDataModule(args) + + train_cuts = ami.aimix_train_cuts(rvb_affix="comb", sources=True) + dev_cuts = ami.ami_cuts(split="dev", type="ihm-mix") + dev_cuts = dev_cuts.trim_to_supervision_groups(max_pause=0.0).filter( + lambda c: 0.2 <= c.duration <= 60.0 + ) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = ami.train_dataloaders( + train_cuts, + sampler_state_dict=sampler_state_dict, + sources=True, + ) + valid_dl = ami.valid_dataloaders(dev_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = [sp.encode(text_ch) for text_ch in batch["text"]] + num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y] + logging.info(f"num tokens: {num_tokens}") + + +def main(): + parser = get_parser() + AmiAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + +if __name__ == "__main__": + main() diff --git a/egs/ami/SURT/dprnn_zipformer/train_adapt.py b/egs/ami/SURT/dprnn_zipformer/train_adapt.py new file mode 100755 index 000000000..9f3b4425f --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/train_adapt.py @@ -0,0 +1,1411 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +# ./dprnn_zipformer/train.py should be run before this script. + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./dprnn_zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir dprnn_zipformer/exp_adapt \ + --model-init-ckpt dprnn_zipformer/exp/epoch-30.pt \ + --max-duration 550 +""" + +import argparse +import copy +import logging +import warnings +from itertools import chain +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import AmiAsrDataModule +from decoder import Decoder +from dprnn import DPRNN +from einops.layers.torch import Rearrange +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import LOG_EPSILON, fix_random_seed +from model import SURT +from optim import Eden, ScaledAdam +from scaling import ScaledLinear, ScaledLSTM +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-mask-encoder-layers", + type=int, + default=4, + help="Number of layers in the DPRNN based mask encoder.", + ) + + parser.add_argument( + "--mask-encoder-dim", + type=int, + default=256, + help="Hidden dimension of the LSTM blocks in DPRNN.", + ) + + parser.add_argument( + "--mask-encoder-segment-size", + type=int, + default=32, + help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the " + "decode-chunk-length of the zipformer encoder.", + ) + + parser.add_argument( + "--chunk-width-randomization", + type=bool, + default=False, + help="Whether to randomize the chunk width in DPRNN.", + ) + + # Zipformer config is based on: + # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740 + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,2,2,2", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="768,768,768,768,768", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="256,256,256,256,256", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="192,192,192,192,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--use-joint-encoder-layer", + type=str, + default="linear", + choices=["linear", "lstm", "none"], + help="Whether to use a joint layer to combine all branches.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--short-chunk-size", + type=int, + default=50, + help="""Chunk length of dynamic training, the chunk size would be either + max sequence length of current batch or uniformly sampled from (1, short_chunk_size). + """, + ) + + parser.add_argument( + "--num-left-chunks", + type=int, + default=4, + help="How many left context can be seen in chunks when calculating attention.", + ) + + parser.add_argument( + "--decode-chunk-len", + type=int, + default=32, + help="The chunk size for decoding (in frames before subsampling)", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=20, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="conv_lstm_transducer_stateless_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--model-init-ckpt", + type=str, + default=None, + help="""The model checkpoint to initialize the model (either full or part). + If not specified, the model is randomly initialized. + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.0001, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=2, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=1, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=100, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 2000, + # parameters for SURT + "num_channels": 2, + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed + # parameters for Noam + "model_warm_step": 5000, # arg given to model, not for lrate + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "env_info": get_env_info(), + } + ) + + return params + + +def get_mask_encoder_model(params: AttributeDict) -> nn.Module: + mask_encoder = DPRNN( + feature_dim=params.feature_dim, + input_size=params.mask_encoder_dim, + hidden_size=params.mask_encoder_dim, + output_size=params.feature_dim * params.num_channels, + segment_size=params.mask_encoder_segment_size, + num_blocks=params.num_mask_encoder_layers, + chunk_width_randomization=params.chunk_width_randomization, + ) + return mask_encoder + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + num_left_chunks=params.num_left_chunks, + short_chunk_size=params.short_chunk_size, + decode_chunk_size=params.decode_chunk_len // 2, + ) + return encoder + + +def get_joint_encoder_layer(params: AttributeDict) -> nn.Module: + class TakeFirst(nn.Module): + def forward(self, x): + return x[0] + + if params.use_joint_encoder_layer == "linear": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + nn.Linear( + params.num_channels * encoder_dim, params.num_channels * encoder_dim + ), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "lstm": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + ScaledLSTM( + input_size=params.num_channels * encoder_dim, + hidden_size=params.num_channels * encoder_dim, + num_layers=1, + bias=True, + batch_first=True, + dropout=0.0, + bidirectional=False, + ), + TakeFirst(), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "none": + joint_layer = None + else: + raise ValueError( + f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}" + ) + return joint_layer + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_surt_model( + params: AttributeDict, +) -> nn.Module: + mask_encoder = get_mask_encoder_model(params) + encoder = get_encoder_model(params) + joint_layer = get_joint_encoder_layer(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = SURT( + mask_encoder=mask_encoder, + encoder=encoder, + joint_encoder_layer=joint_layer, + decoder=decoder, + joiner=joiner, + num_channels=params.num_channels, + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_heat_loss(x_masked, batch, num_channels=2) -> Tensor: + """ + Compute HEAT loss for separated sources using the output of mask encoder. + Args: + x_masked: + The output of mask encoder. It is a tensor of shape (B, T, C). + batch: + A batch of data. See `lhotse.dataset.K2SurtDatasetWithSources()` + for the content in it. + num_channels: + The number of output branches in the SURT model. + """ + B, T, D = x_masked[0].shape + device = x_masked[0].device + + # Create training targets for each channel. + targets = [] + for i in range(num_channels): + target = torch.ones_like(x_masked[i]) * LOG_EPSILON + targets.append(target) + + source_feats = batch["source_feats"] + source_boundaries = batch["source_boundaries"] + input_lens = batch["input_lens"].to(device) + # Assign sources to channels based on the HEAT criteria + for b in range(B): + cut_source_feats = source_feats[b] + cut_source_boundaries = source_boundaries[b] + last_seg_end = [0 for _ in range(num_channels)] + for source_feat, (start, end) in zip(cut_source_feats, cut_source_boundaries): + assigned = False + for i in range(num_channels): + if start >= last_seg_end[i]: + targets[i][b, start:end, :] += source_feat.to(device) + last_seg_end[i] = max(end, last_seg_end[i]) + assigned = True + break + if not assigned: + min_end_channel = last_seg_end.index(min(last_seg_end)) + targets[min_end_channel][b, start:end, :] += source_feat + last_seg_end[min_end_channel] = max(end, last_seg_end[min_end_channel]) + + # Get padding mask based on input lengths + pad_mask = torch.arange(T, device=device).expand(B, T) > input_lens.unsqueeze(1) + pad_mask = pad_mask.unsqueeze(-1) + + # Compute masked loss for each channel + losses = torch.zeros((num_channels, B, T, D), device=device) + for i in range(num_channels): + loss = nn.functional.mse_loss(x_masked[i], targets[i], reduction="none") + # Apply padding mask to loss + loss.masked_fill_(pad_mask, 0) + losses[i] = loss + + # loss: C x B x T x D. pad_mask: B x T x 1 + # We want to compute loss for each item in the batch. Each item has loss given + # by the sum over C, and average over T and D. For T, we need to use the padding. + loss = losses.sum(0).mean(-1).sum(-1) / batch["input_lens"].to(device) + return loss + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"].to(device) + feature_lens = batch["input_lens"].to(device) + + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + + # The dataloader returns text as a list of cuts, each of which is a list of channel + # text. We flatten this to a list where all channels are together, i.e., it looks like + # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2]. + text = [val for tup in zip(*batch["text"]) for val in tup] + assert len(text) == len(feature) * params.num_channels + + # Convert all channel texts to token IDs and create a ragged tensor. + y = sp.encode(text, out_type=int) + y = k2.RaggedTensor(y).to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.model_warm_step + + with torch.set_grad_enabled(is_training): + (simple_loss, pruned_loss, ctc_loss, x_masked) = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + reduction="none", + subsampling_factor=params.subsampling_factor, + ) + simple_loss_is_finite = torch.isfinite(simple_loss) + pruned_loss_is_finite = torch.isfinite(pruned_loss) + ctc_loss_is_finite = torch.isfinite(ctc_loss) + + # Compute HEAT loss + if is_training and params.heat_loss_scale > 0.0: + heat_loss = compute_heat_loss( + x_masked, batch, num_channels=params.num_channels + ) + else: + heat_loss = torch.tensor(0.0, device=device) + + heat_loss_is_finite = torch.isfinite(heat_loss) + is_finite = ( + simple_loss_is_finite + & pruned_loss_is_finite + & ctc_loss_is_finite + & heat_loss_is_finite + ) + if not torch.all(is_finite): + # logging.info( + # "Not all losses are finite!\n" + # f"simple_losses: {simple_loss}\n" + # f"pruned_losses: {pruned_loss}\n" + # f"ctc_losses: {ctc_loss}\n" + # f"heat_losses: {heat_loss}\n" + # ) + # display_and_save_batch(batch, params=params, sp=sp) + simple_loss = simple_loss[simple_loss_is_finite] + pruned_loss = pruned_loss[pruned_loss_is_finite] + ctc_loss = ctc_loss[ctc_loss_is_finite] + heat_loss = heat_loss[heat_loss_is_finite] + + # If either all simple_loss or pruned_loss is inf or nan, + # we stop the training process by raising an exception + if ( + torch.all(~simple_loss_is_finite) + or torch.all(~pruned_loss_is_finite) + or torch.all(~ctc_loss_is_finite) + or torch.all(~heat_loss_is_finite) + ): + raise ValueError( + "There are too many utterances in this batch " + "leading to inf or nan losses." + ) + + simple_loss_sum = simple_loss.sum() + pruned_loss_sum = pruned_loss.sum() + ctc_loss_sum = ctc_loss.sum() + heat_loss_sum = heat_loss.sum() + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss = ( + simple_loss_scale * simple_loss_sum + + pruned_loss_scale * pruned_loss_sum + + params.ctc_loss_scale * ctc_loss_sum + + params.heat_loss_scale * heat_loss_sum + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + # info["frames"] is an approximate number for two reasons: + # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2 + # (2) If some utterances in the batch lead to inf/nan loss, they + # are filtered out. + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances` # noqa + info["utterances"] = feature.size(0) + # averaged input duration in frames over utterances + info["utt_duration"] = feature_lens.sum().item() + # averaged padding proportion over utterances + info["utt_pad_proportion"] = ( + ((feature.size(1) - feature_lens) / feature.size(1)).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss_sum.detach().cpu().item() + info["pruned_loss"] = pruned_loss_sum.detach().cpu().item() + if params.ctc_loss_scale > 0.0: + info["ctc_loss"] = ctc_loss_sum.detach().cpu().item() + if params.heat_loss_scale > 0.0: + info["heat_loss"] = heat_loss_sum.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + torch.cuda.empty_cache() + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = batch["inputs"].shape[0] + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_surt_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + + if checkpoints is None and params.model_init_ckpt is not None: + logging.info( + f"Initializing model with checkpoint from {params.model_init_ckpt}" + ) + init_ckpt = torch.load(params.model_init_ckpt, map_location=device) + model.load_state_dict(init_ckpt["model"], strict=False) + + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + diagnostic = diagnostics.attach_diagnostics(model) + + ami = AmiAsrDataModule(args) + + train_cuts = ami.train_cuts() + train_cuts = train_cuts.filter(lambda c: 0.5 <= c.duration <= 35.0) + dev_cuts = ami.ami_cuts(split="dev", type="ihm-mix") + dev_cuts = dev_cuts.trim_to_supervision_groups(max_pause=0.0).filter( + lambda c: 0.2 <= c.duration <= 60.0 + ) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = ami.train_dataloaders( + train_cuts, + sampler_state_dict=sampler_state_dict, + ) + valid_dl = ami.valid_dataloaders(dev_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = [sp.encode(text_ch) for text_ch in batch["text"]] + num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y] + logging.info(f"num tokens: {num_tokens}") + + +def main(): + parser = get_parser() + AmiAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + +if __name__ == "__main__": + main() diff --git a/egs/ami/SURT/dprnn_zipformer/zipformer.py b/egs/ami/SURT/dprnn_zipformer/zipformer.py new file mode 120000 index 000000000..59b772024 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/zipformer.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/zipformer.py \ No newline at end of file diff --git a/egs/ami/SURT/local/add_source_feats.py b/egs/ami/SURT/local/add_source_feats.py new file mode 100755 index 000000000..0917b88a6 --- /dev/null +++ b/egs/ami/SURT/local/add_source_feats.py @@ -0,0 +1,78 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file adds source features as temporal arrays to the mixture manifests. +It looks for manifests in the directory data/manifests. +""" +import logging +from pathlib import Path + +import numpy as np +from lhotse import CutSet, LilcomChunkyWriter, load_manifest, load_manifest_lazy +from tqdm import tqdm + + +def add_source_feats(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + logging.info("Reading mixed cuts") + mixed_cuts_clean = load_manifest_lazy(src_dir / "cuts_train_clean.jsonl.gz") + mixed_cuts_reverb = load_manifest_lazy(src_dir / "cuts_train_reverb.jsonl.gz") + + logging.info("Reading source cuts") + source_cuts = load_manifest(src_dir / "ihm_cuts_train_trimmed.jsonl.gz") + + logging.info("Adding source features to the mixed cuts") + pbar = tqdm(total=len(mixed_cuts_clean), desc="Adding source features") + with CutSet.open_writer( + src_dir / "cuts_train_clean_sources.jsonl.gz" + ) as cut_writer_clean, CutSet.open_writer( + src_dir / "cuts_train_reverb_sources.jsonl.gz" + ) as cut_writer_reverb, LilcomChunkyWriter( + output_dir / "feats_train_clean_sources" + ) as source_feat_writer: + for cut_clean, cut_reverb in zip(mixed_cuts_clean, mixed_cuts_reverb): + assert cut_reverb.id == cut_clean.id + "_rvb" + source_feats = [] + source_feat_offsets = [] + cur_offset = 0 + for sup in sorted( + cut_clean.supervisions, key=lambda s: (s.start, s.speaker) + ): + source_cut = source_cuts[sup.id] + source_feats.append(source_cut.load_features()) + source_feat_offsets.append(cur_offset) + cur_offset += source_cut.num_frames + cut_clean.source_feats = source_feat_writer.store_array( + cut_clean.id, np.concatenate(source_feats, axis=0) + ) + cut_clean.source_feat_offsets = source_feat_offsets + cut_writer_clean.write(cut_clean) + # Also write the reverb cut + cut_reverb.source_feats = cut_clean.source_feats + cut_reverb.source_feat_offsets = cut_clean.source_feat_offsets + cut_writer_reverb.write(cut_reverb) + pbar.update(1) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + add_source_feats() diff --git a/egs/ami/SURT/local/compute_fbank_aimix.py b/egs/ami/SURT/local/compute_fbank_aimix.py new file mode 100755 index 000000000..91b3a060b --- /dev/null +++ b/egs/ami/SURT/local/compute_fbank_aimix.py @@ -0,0 +1,185 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the synthetically mixed AMI and ICSI +train set. +It looks for manifests in the directory data/manifests. + +The generated fbank features are saved in data/fbank. +""" +import logging +import random +import warnings +from pathlib import Path + +import torch +import torch.multiprocessing +import torchaudio +from lhotse import ( + AudioSource, + LilcomChunkyWriter, + Recording, + load_manifest, + load_manifest_lazy, +) +from lhotse.audio import set_ffmpeg_torchaudio_info_enabled +from lhotse.cut import MixedCut, MixTrack, MultiCut +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.utils import fix_random_seed, uuid4 +from tqdm import tqdm + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") +torchaudio.set_audio_backend("soundfile") +set_ffmpeg_torchaudio_info_enabled(False) + + +def compute_fbank_aimix(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + logging.info("Reading manifests") + train_cuts = load_manifest_lazy(src_dir / "ai-mix_cuts_clean_full.jsonl.gz") + + # only uses RIRs and noises from REVERB challenge + real_rirs = load_manifest(src_dir / "real-rir_recordings_all.jsonl.gz").filter( + lambda r: "RVB2014" in r.id + ) + noises = load_manifest(src_dir / "iso-noise_recordings_all.jsonl.gz").filter( + lambda r: "RVB2014" in r.id + ) + + # Apply perturbation to the training cuts + logging.info("Applying perturbation to the training cuts") + train_cuts_rvb = train_cuts.map( + lambda c: augment( + c, perturb_snr=True, rirs=real_rirs, noises=noises, perturb_loudness=True + ) + ) + + logging.info("Extracting fbank features for training cuts") + _ = train_cuts.compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / "ai-mix_feats_clean", + manifest_path=src_dir / "cuts_train_clean.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + _ = train_cuts_rvb.compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / "ai-mix_feats_reverb", + manifest_path=src_dir / "cuts_train_reverb.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + +def augment(cut, perturb_snr=False, rirs=None, noises=None, perturb_loudness=False): + """ + Given a mixed cut, this function optionally applies the following augmentations: + - Perturbing the SNRs of the tracks (in range [-5, 5] dB) + - Reverberation using a randomly selected RIR + - Adding noise + - Perturbing the loudness (in range [-20, -25] dB) + """ + out_cut = cut.drop_features() + + # Perturb the SNRs (optional) + if perturb_snr: + snrs = [random.uniform(-5, 5) for _ in range(len(cut.tracks))] + for i, (track, snr) in enumerate(zip(out_cut.tracks, snrs)): + if i == 0: + # Skip the first track since it is the reference + continue + track.snr = snr + + # Reverberate the cut (optional) + if rirs is not None: + # Select an RIR at random + rir = random.choice(rirs) + # Select a channel at random + rir_channel = random.choice(list(range(rir.num_channels))) + # Reverberate the cut + out_cut = out_cut.reverb_rir(rir_recording=rir, rir_channels=[rir_channel]) + + # Add noise (optional) + if noises is not None: + # Select a noise recording at random + noise = random.choice(noises).to_cut() + if isinstance(noise, MultiCut): + noise = noise.to_mono()[0] + # Select an SNR at random + snr = random.uniform(10, 30) + # Repeat the noise to match the duration of the cut + noise = repeat_cut(noise, out_cut.duration) + out_cut = MixedCut( + id=out_cut.id, + tracks=[ + MixTrack(cut=out_cut, type="MixedCut"), + MixTrack(cut=noise, type="DataCut", snr=snr), + ], + ) + + # Perturb the loudness (optional) + if perturb_loudness: + target_loudness = random.uniform(-20, -25) + out_cut = out_cut.normalize_loudness(target_loudness, mix_first=True) + return out_cut + + +def repeat_cut(cut, duration): + while cut.duration < duration: + cut = cut.mix(cut, offset_other_by=cut.duration) + return cut.truncate(duration=duration) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + fix_random_seed(42) + compute_fbank_aimix() diff --git a/egs/ami/SURT/local/compute_fbank_ami.py b/egs/ami/SURT/local/compute_fbank_ami.py new file mode 100755 index 000000000..351b41765 --- /dev/null +++ b/egs/ami/SURT/local/compute_fbank_ami.py @@ -0,0 +1,94 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the AMI dataset. +We compute features for full recordings (i.e., without trimming to supervisions). +This way we can create arbitrary segmentations later. + +The generated fbank features are saved in data/fbank. +""" +import logging +import math +from pathlib import Path + +import torch +import torch.multiprocessing +from lhotse import CutSet, LilcomChunkyWriter +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.recipes.utils import read_manifests_if_cached + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + + +def compute_fbank_ami(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + logging.info("Reading manifests") + manifests = {} + for part in ["ihm-mix", "sdm", "mdm8-bf"]: + manifests[part] = read_manifests_if_cached( + dataset_parts=["train", "dev", "test"], + output_dir=src_dir, + prefix=f"ami-{part}", + suffix="jsonl.gz", + ) + + for part in ["ihm-mix", "sdm", "mdm8-bf"]: + for split in ["train", "dev", "test"]: + logging.info(f"Processing {part} {split}") + cuts = CutSet.from_manifests( + **manifests[part][split] + ).compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / f"ami-{part}_{split}_feats", + manifest_path=src_dir / f"cuts_ami-{part}_{split}.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + ) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + compute_fbank_ami() diff --git a/egs/ami/SURT/local/compute_fbank_icsi.py b/egs/ami/SURT/local/compute_fbank_icsi.py new file mode 100755 index 000000000..4e2ff3f3b --- /dev/null +++ b/egs/ami/SURT/local/compute_fbank_icsi.py @@ -0,0 +1,95 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the ICSI dataset. +We compute features for full recordings (i.e., without trimming to supervisions). +This way we can create arbitrary segmentations later. + +The generated fbank features are saved in data/fbank. +""" +import logging +import math +from pathlib import Path + +import torch +import torch.multiprocessing +from lhotse import CutSet, LilcomChunkyWriter +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.recipes.utils import read_manifests_if_cached + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + + +def compute_fbank_icsi(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + logging.info("Reading manifests") + manifests = {} + for part in ["ihm-mix", "sdm"]: + manifests[part] = read_manifests_if_cached( + dataset_parts=["train"], + output_dir=src_dir, + prefix=f"icsi-{part}", + suffix="jsonl.gz", + ) + + for part in ["ihm-mix", "sdm"]: + for split in ["train"]: + logging.info(f"Processing {part} {split}") + cuts = CutSet.from_manifests( + **manifests[part][split] + ).compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / f"icsi-{part}_{split}_feats", + manifest_path=src_dir / f"cuts_icsi-{part}_{split}.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + compute_fbank_icsi() diff --git a/egs/ami/SURT/local/compute_fbank_ihm.py b/egs/ami/SURT/local/compute_fbank_ihm.py new file mode 100755 index 000000000..56f54aa21 --- /dev/null +++ b/egs/ami/SURT/local/compute_fbank_ihm.py @@ -0,0 +1,101 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the trimmed sub-segments which will be +used for simulating the training mixtures. + +The generated fbank features are saved in data/fbank. +""" +import logging +import math +from pathlib import Path + +import torch +import torch.multiprocessing +import torchaudio +from lhotse import CutSet, LilcomChunkyWriter, load_manifest +from lhotse.audio import set_ffmpeg_torchaudio_info_enabled +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.recipes.utils import read_manifests_if_cached +from tqdm import tqdm + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") +torchaudio.set_audio_backend("soundfile") +set_ffmpeg_torchaudio_info_enabled(False) + + +def compute_fbank_ihm(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + logging.info("Reading manifests") + manifests = {} + for data in ["ami", "icsi"]: + manifests[data] = read_manifests_if_cached( + dataset_parts=["train"], + output_dir=src_dir, + types=["recordings", "supervisions"], + prefix=f"{data}-ihm", + suffix="jsonl.gz", + ) + + logging.info("Computing features") + for data in ["ami", "icsi"]: + cs = CutSet.from_manifests(**manifests[data]["train"]) + cs = cs.trim_to_supervisions(keep_overlapping=False) + cs = cs.normalize_loudness(target=-23.0, affix_id=False) + cs = cs + cs.perturb_speed(0.9) + cs.perturb_speed(1.1) + _ = cs.compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / f"{data}-ihm_train_feats", + manifest_path=src_dir / f"{data}-ihm_cuts_train.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + compute_fbank_ihm() diff --git a/egs/ami/SURT/local/prepare_ami_train_cuts.py b/egs/ami/SURT/local/prepare_ami_train_cuts.py new file mode 100755 index 000000000..72fced70d --- /dev/null +++ b/egs/ami/SURT/local/prepare_ami_train_cuts.py @@ -0,0 +1,146 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file creates AMI train segments. +""" +import logging +import math +from pathlib import Path + +import torch +import torch.multiprocessing +from lhotse import LilcomChunkyWriter, load_manifest_lazy +from lhotse.cut import Cut, CutSet +from lhotse.utils import EPSILON, add_durations +from tqdm import tqdm + + +def cut_into_windows(cuts: CutSet, duration: float): + """ + This function takes a CutSet and cuts each cut into windows of roughly + `duration` seconds. By roughly, we mean that we try to adjust for the last supervision + that exceeds the duration, or is shorter than the duration. + """ + res = [] + with tqdm() as pbar: + for cut in cuts: + pbar.update(1) + sups = cut.index_supervisions()[cut.id] + sr = cut.sampling_rate + start = 0.0 + end = duration + num_tries = 0 + while start < cut.duration and num_tries < 2: + # Find the supervision that are cut by the window endpoint + hitlist = [iv for iv in sups.at(end) if iv.begin < end] + # If there are no supervisions, we are done + if not hitlist: + res.append( + cut.truncate( + offset=start, + duration=add_durations(end, -start, sampling_rate=sr), + keep_excessive_supervisions=False, + ) + ) + # Update the start and end for the next window + start = end + end = add_durations(end, duration, sampling_rate=sr) + else: + # find ratio of durations cut by the window endpoint + ratios = [ + add_durations(end, -iv.end, sampling_rate=sr) / iv.length() + for iv in hitlist + ] + # we retain the supervisions that have >50% of their duration + # in the window, and discard the others + retained = [] + discarded = [] + for iv, ratio in zip(hitlist, ratios): + if ratio > 0.5: + retained.append(iv) + else: + discarded.append(iv) + cur_end = max(iv.end for iv in retained) if retained else end + res.append( + cut.truncate( + offset=start, + duration=add_durations(cur_end, -start, sampling_rate=sr), + keep_excessive_supervisions=False, + ) + ) + # For the next window, we start at the earliest discarded supervision + next_start = min(iv.begin for iv in discarded) if discarded else end + next_end = add_durations(next_start, duration, sampling_rate=sr) + # It may happen that next_start is the same as start, in which case + # we will advance the window anyway + if next_start == start: + logging.warning( + f"Next start is the same as start: {next_start} == {start} for cut {cut.id}" + ) + start = end + EPSILON + end = add_durations(start, duration, sampling_rate=sr) + num_tries += 1 + else: + start = next_start + end = next_end + return CutSet.from_cuts(res) + + +def prepare_train_cuts(): + src_dir = Path("data/manifests") + + logging.info("Loading the manifests") + train_cuts_ihm = load_manifest_lazy( + src_dir / "cuts_ami-ihm-mix_train.jsonl.gz" + ).map(lambda c: c.with_id(f"{c.id}_ihm-mix")) + train_cuts_sdm = load_manifest_lazy(src_dir / "cuts_ami-sdm_train.jsonl.gz").map( + lambda c: c.with_id(f"{c.id}_sdm") + ) + train_cuts_mdm = load_manifest_lazy( + src_dir / "cuts_ami-mdm8-bf_train.jsonl.gz" + ).map(lambda c: c.with_id(f"{c.id}_mdm8-bf")) + + # Combine all cuts into one CutSet + train_cuts = train_cuts_ihm + train_cuts_sdm + train_cuts_mdm + + train_cuts_1 = train_cuts.trim_to_supervision_groups(max_pause=0.5) + train_cuts_2 = train_cuts.trim_to_supervision_groups(max_pause=0.0) + + # Combine the two segmentations + train_all = train_cuts_1 + train_cuts_2 + + # At this point, some of the cuts may be very long. We will cut them into windows of + # roughly 30 seconds. + logging.info("Cutting the segments into windows of 30 seconds") + train_all_30 = cut_into_windows(train_all, duration=30.0) + logging.info(f"Number of cuts after cutting into windows: {len(train_all_30)}") + + # Show statistics + train_all.describe(full=True) + + # Save the cuts + logging.info("Saving the cuts") + train_all.to_file(src_dir / "cuts_train_ami.jsonl.gz") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + prepare_train_cuts() diff --git a/egs/ami/SURT/local/prepare_icsi_train_cuts.py b/egs/ami/SURT/local/prepare_icsi_train_cuts.py new file mode 100755 index 000000000..818e26bfb --- /dev/null +++ b/egs/ami/SURT/local/prepare_icsi_train_cuts.py @@ -0,0 +1,67 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file creates ICSI train segments. +""" +import logging +from pathlib import Path + +from lhotse import load_manifest_lazy +from prepare_ami_train_cuts import cut_into_windows + + +def prepare_train_cuts(): + src_dir = Path("data/manifests") + + logging.info("Loading the manifests") + train_cuts_ihm = load_manifest_lazy( + src_dir / "cuts_icsi-ihm-mix_train.jsonl.gz" + ).map(lambda c: c.with_id(f"{c.id}_ihm-mix")) + train_cuts_sdm = load_manifest_lazy(src_dir / "cuts_icsi-sdm_train.jsonl.gz").map( + lambda c: c.with_id(f"{c.id}_sdm") + ) + + # Combine all cuts into one CutSet + train_cuts = train_cuts_ihm + train_cuts_sdm + + train_cuts_1 = train_cuts.trim_to_supervision_groups(max_pause=0.5) + train_cuts_2 = train_cuts.trim_to_supervision_groups(max_pause=0.0) + + # Combine the two segmentations + train_all = train_cuts_1 + train_cuts_2 + + # At this point, some of the cuts may be very long. We will cut them into windows of + # roughly 30 seconds. + logging.info("Cutting the segments into windows of 30 seconds") + train_all_30 = cut_into_windows(train_all, duration=30.0) + logging.info(f"Number of cuts after cutting into windows: {len(train_all_30)}") + + # Show statistics + train_all.describe(full=True) + + # Save the cuts + logging.info("Saving the cuts") + train_all.to_file(src_dir / "cuts_train_icsi.jsonl.gz") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + prepare_train_cuts() diff --git a/egs/ami/SURT/local/prepare_lang_bpe.py b/egs/ami/SURT/local/prepare_lang_bpe.py new file mode 120000 index 000000000..36b40e7fc --- /dev/null +++ b/egs/ami/SURT/local/prepare_lang_bpe.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/prepare_lang_bpe.py \ No newline at end of file diff --git a/egs/ami/SURT/local/train_bpe_model.py b/egs/ami/SURT/local/train_bpe_model.py new file mode 120000 index 000000000..6fad36421 --- /dev/null +++ b/egs/ami/SURT/local/train_bpe_model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/train_bpe_model.py \ No newline at end of file diff --git a/egs/ami/SURT/prepare.sh b/egs/ami/SURT/prepare.sh new file mode 100755 index 000000000..ea4e5baf2 --- /dev/null +++ b/egs/ami/SURT/prepare.sh @@ -0,0 +1,195 @@ +#!/usr/bin/env bash + +set -eou pipefail + +stage=-1 +stop_stage=100 + +# We assume dl_dir (download dir) contains the following +# directories and files. If not, they will be downloaded +# by this script automatically. +# +# - $dl_dir/ami +# You can find audio and transcripts for AMI in this path. +# +# - $dl_dir/icsi +# You can find audio and transcripts for ICSI in this path. +# +# - $dl_dir/rirs_noises +# This directory contains the RIRS_NOISES corpus downloaded from https://openslr.org/28/. +# +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data +vocab_size=500 + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +log "dl_dir: $dl_dir" + +if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then + log "Stage 0: Download data" + + # If you have pre-downloaded it to /path/to/amicorpus, + # you can create a symlink + # + # ln -sfv /path/to/amicorpus $dl_dir/amicorpus + # + if [ ! -d $dl_dir/amicorpus ]; then + for mic in ihm ihm-mix sdm mdm8-bf; do + lhotse download ami --mic $mic $dl_dir/amicorpus + done + fi + + # If you have pre-downloaded it to /path/to/icsi, + # you can create a symlink + # + # ln -sfv /path/to/icsi $dl_dir/icsi + # + if [ ! -d $dl_dir/icsi ]; then + lhotse download icsi $dl_dir/icsi + fi + + # If you have pre-downloaded it to /path/to/rirs_noises, + # you can create a symlink + # + # ln -sfv /path/to/rirs_noises $dl_dir/ + # + if [ ! -d $dl_dir/rirs_noises ]; then + lhotse download rirs_noises $dl_dir + fi +fi + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Prepare AMI manifests" + # We assume that you have downloaded the AMI corpus + # to $dl_dir/amicorpus. We perform text normalization for the transcripts. + mkdir -p data/manifests + for mic in ihm ihm-mix sdm mdm8-bf; do + log "Preparing AMI manifest for $mic" + lhotse prepare ami --mic $mic --max-words-per-segment 30 --merge-consecutive $dl_dir/amicorpus data/manifests/ + done +fi + +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Prepare ICSI manifests" + # We assume that you have downloaded the ICSI corpus + # to $dl_dir/icsi. We perform text normalization for the transcripts. + mkdir -p data/manifests + log "Preparing ICSI manifest" + for mic in ihm ihm-mix sdm; do + lhotse prepare icsi --mic $mic $dl_dir/icsi data/manifests/ + done +fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Prepare RIRs" + # We assume that you have downloaded the RIRS_NOISES corpus + # to $dl_dir/rirs_noises + lhotse prepare rir-noise -p real_rir -p iso_noise $dl_dir/rirs_noises data/manifests +fi + +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + log "Stage 3: Extract features for AMI and ICSI recordings" + python local/compute_fbank_ami.py + python local/compute_fbank_icsi.py +fi + +if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then + log "Stage 5: Create sources for simulating mixtures" + # In the following script, we speed-perturb the IHM recordings and extract features. + python local/compute_fbank_ihm.py + lhotse combine data/manifests/ami-ihm_cuts_train.jsonl.gz \ + data/manifests/icsi-ihm_cuts_train.jsonl.gz - |\ + lhotse cut trim-to-alignments --type word --max-pause 0.5 - - |\ + lhotse filter 'duration<=12.0' - - |\ + shuf | gzip -c > data/manifests/ihm_cuts_train_trimmed.jsonl.gz +fi + +if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then + log "Stage 6: Create training mixtures" + lhotse workflows simulate-meetings \ + --method conversational \ + --same-spk-pause 0.5 \ + --diff-spk-pause 0.5 \ + --diff-spk-overlap 1.0 \ + --prob-diff-spk-overlap 0.8 \ + --num-meetings 200000 \ + --num-speakers-per-meeting 2,3 \ + --max-duration-per-speaker 15.0 \ + --max-utterances-per-speaker 3 \ + --seed 1234 \ + --num-jobs 2 \ + data/manifests/ihm_cuts_train_trimmed.jsonl.gz \ + data/manifests/ai-mix_cuts_clean.jsonl.gz + + python local/compute_fbank_aimix.py + + # Add source features to the manifest (will be used for masking loss) + # This may take ~2 hours. + python local/add_source_feats.py + + # Combine clean and reverb + cat <(gunzip -c data/manifests/cuts_train_clean_sources.jsonl.gz) \ + <(gunzip -c data/manifests/cuts_train_reverb_sources.jsonl.gz) |\ + shuf | gzip -c > data/manifests/cuts_train_comb_sources.jsonl.gz +fi + +if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then + log "Stage 7: Create training mixtures from real sessions" + python local/prepare_ami_train_cuts.py + python local/prepare_icsi_train_cuts.py + + # Combine AMI and ICSI + cat <(gunzip -c data/manifests/cuts_train_ami.jsonl.gz) \ + <(gunzip -c data/manifests/cuts_train_icsi.jsonl.gz) |\ + shuf | gzip -c > data/manifests/cuts_train_ami_icsi.jsonl.gz +fi + +if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then + log "Stage 8: Dump transcripts for BPE model training (using AMI and ICSI)." + mkdir -p data/lm + cat <(gunzip -c data/manifests/ami-sdm_supervisions_train.jsonl.gz | jq '.text' | sed 's:"::g') \ + <(gunzip -c data/manifests/icsi-sdm_supervisions_train.jsonl.gz | jq '.text' | sed 's:"::g') \ + > data/lm/transcript_words.txt +fi + +if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then + log "Stage 9: Prepare BPE based lang (combining AMI and ICSI)" + + lang_dir=data/lang_bpe_${vocab_size} + mkdir -p $lang_dir + + # Add special words to words.txt + echo " 0" > $lang_dir/words.txt + echo "!SIL 1" >> $lang_dir/words.txt + echo " 2" >> $lang_dir/words.txt + + # Add regular words to words.txt + cat data/lm/transcript_words.txt | grep -o -E '\w+' | sort -u | awk '{print $0,NR+2}' >> $lang_dir/words.txt + + # Add remaining special word symbols expected by LM scripts. + num_words=$(cat $lang_dir/words.txt | wc -l) + echo " ${num_words}" >> $lang_dir/words.txt + num_words=$(cat $lang_dir/words.txt | wc -l) + echo " ${num_words}" >> $lang_dir/words.txt + num_words=$(cat $lang_dir/words.txt | wc -l) + echo "#0 ${num_words}" >> $lang_dir/words.txt + + ./local/train_bpe_model.py \ + --lang-dir $lang_dir \ + --vocab-size $vocab_size \ + --transcript data/lm/transcript_words.txt + + if [ ! -f $lang_dir/L_disambig.pt ]; then + ./local/prepare_lang_bpe.py --lang-dir $lang_dir + fi +fi diff --git a/egs/ami/SURT/shared b/egs/ami/SURT/shared new file mode 120000 index 000000000..4cbd91a7e --- /dev/null +++ b/egs/ami/SURT/shared @@ -0,0 +1 @@ +../../../icefall/shared \ No newline at end of file diff --git a/egs/commonvoice/ASR/local/preprocess_commonvoice.py b/egs/commonvoice/ASR/local/preprocess_commonvoice.py index c5ec14502..e60459765 100755 --- a/egs/commonvoice/ASR/local/preprocess_commonvoice.py +++ b/egs/commonvoice/ASR/local/preprocess_commonvoice.py @@ -45,7 +45,7 @@ def get_args(): def normalize_text(utt: str) -> str: utt = re.sub(r"[{0}]+".format("-"), " ", utt) - return re.sub(r"[^a-zA-Z\s]", "", utt).upper() + return re.sub(r"[^a-zA-Z\s']", "", utt).upper() def preprocess_commonvoice( diff --git a/egs/libricss/SURT/README.md b/egs/libricss/SURT/README.md new file mode 100644 index 000000000..10a1aaad1 --- /dev/null +++ b/egs/libricss/SURT/README.md @@ -0,0 +1,249 @@ +# Introduction + +This is a multi-talker ASR recipe for the LibriCSS dataset. We train a Streaming +Unmixing and Recognition Transducer (SURT) model for the task. In this README, +we will describe the task, the model, and the training process. We will also +provide links to pre-trained models and training logs. + +## Task + +LibriCSS is a multi-talker meeting corpus formed from mixing together LibriSpeech utterances +and replaying in a real meeting room. It consists of 10 1-hour sessions of audio, each +recorded on a 7-channel microphone. The sessions are recorded at a sampling rate of 16 kHz. +For more information, refer to the paper: +Z. Chen et al., "Continuous speech separation: dataset and analysis," +ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), +Barcelona, Spain, 2020 + +In this recipe, we perform the "continuous, streaming, multi-talker ASR" task on LibriCSS. + +* By "continuous", we mean that the model should be able to transcribe unsegmented audio +without the need of an external VAD. +* By "streaming", we mean that the model has limited right context. We use a right-context +of at most 32 frames (320 ms). +* By "multi-talker", we mean that the model should be able to transcribe overlapping speech +from multiple speakers. + +For now, we do not care about speaker attribution, i.e., the transcription is speaker +agnostic. The evaluation depends on the particular model type. In this case, we use +the optimal reference combination WER (ORC-WER) metric as implemented in the +[meeteval](https://github.com/fgnt/meeteval) toolkit. + +## Model + +We use the Streaming Unmixing and Recognition Transducer (SURT) model for this task. +The model is based on the papers: + +- Lu, Liang et al. “Streaming End-to-End Multi-Talker Speech Recognition.” IEEE Signal Processing Letters 28 (2020): 803-807. +- Raj, Desh et al. “Continuous Streaming Multi-Talker ASR with Dual-Path Transducers.” ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2021): 7317-7321. + +The model is a combination of a speech separation model and a speech recognition model, +but trained end-to-end with a single loss function. The overall architecture is shown +in the figure below. Note that this architecture is slightly different from the one +in the above papers. A detailed description of the model can be found in the following +paper: [SURT 2.0: Advanced in transducer-based multi-talker ASR](https://arxiv.org/abs/2306.10559). + +

+ + + Streaming Unmixing and Recognition Transducer + +

+ +In the [dprnn_zipformer](./dprnn_zipformer) recipe, for example, we use a DPRNN-based masking network +and a Zipfomer-based recognition network. But other combinations are possible as well. + +## Training objective + +We train the model using the pruned transducer loss, similar to other ASR recipes in +icefall. However, an important consideration is how to assign references to the output +channels (2 in this case). For this, we use the heuristic error assignment training (HEAT) +strategy, which assigns references to the first available channel based on their start +times. An illustrative example is shown in the figure below: + +

+ + + Illustration of HEAT-based reference assignment. + +

+ +## Description of the recipe + +### Pre-requisites + +The recipes in this directory need the following packages to be installed: + +- [meeteval](https://github.com/fgnt/meeteval) +- [einops](https://github.com/arogozhnikov/einops) + +Additionally, we initialize the "recognition" transducer with a pre-trained model, +trained on LibriSpeech. For this, please run the following from within `egs/librispeech/ASR`: + +```bash +./prepare.sh + +export CUDA_VISIBLE_DEVICES="0,1,2,3" +python pruned_transducer_stateless7_streaming/train.py \ + --use-fp16 True \ + --exp-dir pruned_transducer_stateless7_streaming/exp \ + --world-size 4 \ + --max-duration 800 \ + --num-epochs 10 \ + --keep-last-k 1 \ + --manifest-dir data/manifests \ + --enable-musan true \ + --master-port 54321 \ + --bpe-model data/lang_bpe_500/bpe.model \ + --num-encoder-layers 2,2,2,2,2 \ + --feedforward-dims 768,768,768,768,768 \ + --nhead 8,8,8,8,8 \ + --encoder-dims 256,256,256,256,256 \ + --attention-dims 192,192,192,192,192 \ + --encoder-unmasked-dims 192,192,192,192,192 \ + --zipformer-downsampling-factors 1,2,4,8,2 \ + --cnn-module-kernels 31,31,31,31,31 \ + --decoder-dim 512 \ + --joiner-dim 512 +``` + +The above is for SURT-base (~26M). For SURT-large (~38M), use `--num-encoder-layers 2,4,3,2,4`. + +Once the above model is trained for 10 epochs, copy it to `egs/libricss/SURT/exp`: + +```bash +cp -r pruned_transducer_stateless7_streaming/exp/epoch-10.pt exp/zipformer_base.pt +``` + +**NOTE:** We also provide this pre-trained checkpoint (see the section below), so you can skip +the above step if you want. + +### Training + +To train the model, run the following from within `egs/libricss/SURT`: + +```bash +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +python dprnn_zipformer/train.py \ + --use-fp16 True \ + --exp-dir dprnn_zipformer/exp/surt_base \ + --world-size 4 \ + --max-duration 500 \ + --max-duration-valid 250 \ + --max-cuts 200 \ + --num-buckets 50 \ + --num-epochs 30 \ + --enable-spec-aug True \ + --enable-musan False \ + --ctc-loss-scale 0.2 \ + --heat-loss-scale 0.2 \ + --base-lr 0.004 \ + --model-init-ckpt exp/zipformer_base.pt \ + --chunk-width-randomization True \ + --num-mask-encoder-layers 4 \ + --num-encoder-layers 2,2,2,2,2 +``` + +The above is for SURT-base (~26M). For SURT-large (~38M), use: + +```bash + --num-mask-encoder-layers 6 \ + --num-encoder-layers 2,4,3,2,4 \ + --model-init-ckpt exp/zipformer_large.pt \ +``` + +**NOTE:** You may need to decrease the `--max-duration` for SURT-large to avoid OOM. + +### Adaptation + +The training step above only trains on simulated mixtures. For best results, we also +adapt the final model on the LibriCSS dev set. For this, run the following from within +`egs/libricss/SURT`: + +```bash +export CUDA_VISIBLE_DEVICES="0" + +python dprnn_zipformer/train_adapt.py \ + --use-fp16 True \ + --exp-dir dprnn_zipformer/exp/surt_base_adapt \ + --world-size 1 \ + --max-duration 500 \ + --max-duration-valid 250 \ + --max-cuts 200 \ + --num-buckets 50 \ + --num-epochs 8 \ + --lr-epochs 2 \ + --enable-spec-aug True \ + --enable-musan False \ + --ctc-loss-scale 0.2 \ + --base-lr 0.0004 \ + --model-init-ckpt dprnn_zipformer/exp/surt_base/epoch-30.pt \ + --chunk-width-randomization True \ + --num-mask-encoder-layers 4 \ + --num-encoder-layers 2,2,2,2,2 +``` + +For SURT-large, use the following config: + +```bash + --num-mask-encoder-layers 6 \ + --num-encoder-layers 2,4,3,2,4 \ + --model-init-ckpt dprnn_zipformer/exp/surt_large/epoch-30.pt \ + --num-epochs 15 \ + --lr-epochs 4 \ +``` + + +### Decoding + +To decode the model, run the following from within `egs/libricss/SURT`: + +#### Greedy search + +```bash +export CUDA_VISIBLE_DEVICES="0" + +python dprnn_zipformer/decode.py \ + --epoch 8 --avg 1 --use-averaged-model False \ + --exp-dir dprnn_zipformer/exp/surt_base_adapt \ + --max-duration 250 \ + --decoding-method greedy_search +``` + +#### Beam search + +```bash +python dprnn_zipformer/decode.py \ + --epoch 8 --avg 1 --use-averaged-model False \ + --exp-dir dprnn_zipformer/exp/surt_base_adapt \ + --max-duration 250 \ + --decoding-method modified_beam_search \ + --beam-size 4 +``` + +## Results (using beam search) + +#### IHM-Mix + +| Model | # params | 0L | 0S | OV10 | OV20 | OV30 | OV40 | Avg. | +|------------|:-------:|:----:|:---:|----:|:----:|:----:|:----:|:----:| +| dprnn_zipformer (base) | 26.7 | 5.1 | 4.2 | 13.7 | 18.7 | 20.5 | 20.6 | 13.8 | +| dprnn_zipformer (large) | 37.9 | 4.6 | 3.8 | 12.7 | 14.3 | 16.7 | 21.2 | 12.2 | + +#### SDM + +| Model | # params | 0L | 0S | OV10 | OV20 | OV30 | OV40 | Avg. | +|------------|:-------:|:----:|:---:|----:|:----:|:----:|:----:|:----:| +| dprnn_zipformer (base) | 26.7 | 6.8 | 7.2 | 21.4 | 24.5 | 28.6 | 31.2 | 20.0 | +| dprnn_zipformer (large) | 37.9 | 6.4 | 6.9 | 17.9 | 19.7 | 25.2 | 25.5 | 16.9 | + +## Pre-trained models and logs + +* Pre-trained models: + +* Training logs: + - surt_base: + - surt_base_adapt: + - surt_large: + - surt_large_adapt: diff --git a/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py b/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py new file mode 100644 index 000000000..51df91598 --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py @@ -0,0 +1,372 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) +# Copyright 2023 Johns Hopkins Univrtsity (Author: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Callable, Dict, List, Optional + +import torch +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures + CutMix, + DynamicBucketingSampler, + K2SurtDataset, + PrecomputedFeatures, + SimpleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import OnTheFlyFeatures +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class LibriCssAsrDataModule: + """ + DataModule for k2 ASR experiments. + It assumes there is always one train and valid dataloader, + but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + and test-other). + + It contains all the common data pipeline modules used in ASR + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + - augmentation, + - on-the-fly feature extraction + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/manifests"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--max-duration-valid", + type=int, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--max-cuts", + type=int, + default=100, + help="Maximum number of cuts in a single batch. You can " + "reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--on-the-fly-feats", + type=str2bool, + default=False, + help=( + "When enabled, use on-the-fly cut mixing and feature " + "extraction. Will drop existing precomputed feature manifests " + "if available." + ), + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + + group.add_argument( + "--enable-spec-aug", + type=str2bool, + default=True, + help="When enabled, use SpecAugment for training dataset.", + ) + + group.add_argument( + "--spec-aug-time-warp-factor", + type=int, + default=80, + help="Used only when --enable-spec-aug is True. " + "It specifies the factor for time warping in SpecAugment. " + "Larger values mean more warping. " + "A value less than 1 means to disable time warp.", + ) + + group.add_argument( + "--enable-musan", + type=str2bool, + default=True, + help="When enabled, select noise from MUSAN and mix it" + "with training dataset. ", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + sampler_state_dict: Optional[Dict[str, Any]] = None, + return_sources: bool = True, + strict: bool = True, + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + logging.info("About to get Musan cuts") + cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") + transforms.append( + CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True) + ) + else: + logging.info("Disable MUSAN") + + input_transforms = [] + if self.args.enable_spec_aug: + logging.info("Enable SpecAugment") + logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}") + # Set the value of num_frame_masks according to Lhotse's version. + # In different Lhotse's versions, the default of num_frame_masks is + # different. + num_frame_masks = 10 + num_frame_masks_parameter = inspect.signature( + SpecAugment.__init__ + ).parameters["num_frame_masks"] + if num_frame_masks_parameter.default == 1: + num_frame_masks = 2 + logging.info(f"Num frame mask: {num_frame_masks}") + input_transforms.append( + SpecAugment( + time_warp_factor=self.args.spec_aug_time_warp_factor, + num_frame_masks=num_frame_masks, + features_mask_size=27, + num_feature_masks=2, + frames_mask_size=100, + ) + ) + else: + logging.info("Disable SpecAugment") + + logging.info("About to create train dataset") + train = K2SurtDataset( + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + cut_transforms=transforms, + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + return_sources=return_sources, + strict=strict, + ) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + quadratic_duration=30.0, + max_cuts=self.args.max_cuts, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SingleCutSampler.") + train_sampler = SimpleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + max_cuts=self.args.max_cuts, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + return train_dl + + def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: + transforms = [] + + logging.info("About to create dev dataset") + validate = K2SurtDataset( + input_strategy=OnTheFlyFeatures( + OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + ) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + cut_transforms=transforms, + return_cuts=self.args.return_cuts, + return_sources=False, + strict=False, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration_valid, + max_cuts=self.args.max_cuts, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=2, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SurtDataset( + input_strategy=OnTheFlyFeatures( + OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + ) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + return_cuts=self.args.return_cuts, + return_sources=False, + strict=False, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration_valid, + max_cuts=self.args.max_cuts, + shuffle=False, + ) + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl + + @lru_cache() + def lsmix_cuts( + self, + rvb_affix: str = "clean", + type_affix: str = "full", + sources: bool = True, + ) -> CutSet: + logging.info("About to get train cuts") + source_affix = "_sources" if sources else "" + cs = load_manifest_lazy( + self.args.manifest_dir + / f"cuts_train_{rvb_affix}_{type_affix}{source_affix}.jsonl.gz" + ) + cs = cs.filter(lambda c: c.duration >= 1.0 and c.duration <= 30.0) + return cs + + @lru_cache() + def libricss_cuts(self, split="dev", type="sdm") -> CutSet: + logging.info(f"About to get LibriCSS {split} {type} cuts") + cs = load_manifest_lazy( + self.args.manifest_dir / f"cuts_{split}_libricss-{type}.jsonl.gz" + ) + return cs diff --git a/egs/libricss/SURT/dprnn_zipformer/beam_search.py b/egs/libricss/SURT/dprnn_zipformer/beam_search.py new file mode 100644 index 000000000..c8e4643d0 --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/beam_search.py @@ -0,0 +1,730 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang +# Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import warnings +from dataclasses import dataclass, field +from typing import Dict, List, Optional, Tuple, Union + +import k2 +import torch +from model import SURT + +from icefall import NgramLmStateCost +from icefall.utils import DecodingResults + + +def greedy_search( + model: SURT, + encoder_out: torch.Tensor, + max_sym_per_frame: int, + return_timestamps: bool = False, +) -> Union[List[int], DecodingResults]: + """Greedy search for a single utterance. + Args: + model: + An instance of `SURT`. + encoder_out: + A tensor of shape (N, T, C) from the encoder. Support only N==1 for now. + max_sym_per_frame: + Maximum number of symbols per frame. If it is set to 0, the WER + would be 100%. + return_timestamps: + Whether to return timestamps. + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + assert encoder_out.ndim == 4 + + # support only batch_size == 1 for now + assert encoder_out.size(0) == 1, encoder_out.size(0) + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + unk_id = getattr(model, "unk_id", blank_id) + + device = next(model.parameters()).device + + decoder_input = torch.tensor( + [-1] * (context_size - 1) + [blank_id], device=device, dtype=torch.int64 + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + T = encoder_out.size(1) + t = 0 + hyp = [blank_id] * context_size + + # timestamp[i] is the frame index after subsampling + # on which hyp[i] is decoded + timestamp = [] + + # Maximum symbols per utterance. + max_sym_per_utt = 1000 + + # symbols per frame + sym_per_frame = 0 + + # symbols per utterance decoded so far + sym_per_utt = 0 + + while t < T and sym_per_utt < max_sym_per_utt: + if sym_per_frame >= max_sym_per_frame: + sym_per_frame = 0 + t += 1 + continue + + # fmt: off + current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2) + # fmt: on + logits = model.joiner( + current_encoder_out, decoder_out.unsqueeze(1), project_input=False + ) + # logits is (1, 1, 1, vocab_size) + + y = logits.argmax().item() + if y not in (blank_id, unk_id): + hyp.append(y) + timestamp.append(t) + decoder_input = torch.tensor([hyp[-context_size:]], device=device).reshape( + 1, context_size + ) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + sym_per_utt += 1 + sym_per_frame += 1 + else: + sym_per_frame = 0 + t += 1 + hyp = hyp[context_size:] # remove blanks + + if not return_timestamps: + return hyp + else: + return DecodingResults( + hyps=[hyp], + timestamps=[timestamp], + ) + + +def greedy_search_batch( + model: SURT, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + return_timestamps: bool = False, +) -> Union[List[List[int]], DecodingResults]: + """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. + Args: + model: + The SURT model. + encoder_out: + Output from the encoder. Its shape is (N, T, C), where N >= 1. + encoder_out_lens: + A 1-D tensor of shape (N,), containing number of valid frames in + encoder_out before padding. + return_timestamps: + Whether to return timestamps. + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + assert encoder_out.ndim == 3 + assert encoder_out.size(0) >= 1, encoder_out.size(0) + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + device = next(model.parameters()).device + + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + hyps = [[-1] * (context_size - 1) + [blank_id] for _ in range(N)] + + # timestamp[n][i] is the frame index after subsampling + # on which hyp[n][i] is decoded + timestamps = [[] for _ in range(N)] + + decoder_input = torch.tensor( + hyps, + device=device, + dtype=torch.int64, + ) # (N, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_out: (N, 1, decoder_out_dim) + + encoder_out = model.joiner.encoder_proj(packed_encoder_out.data) + + offset = 0 + for (t, batch_size) in enumerate(batch_size_list): + start = offset + end = offset + batch_size + current_encoder_out = encoder_out.data[start:end] + current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim) + offset = end + + decoder_out = decoder_out[:batch_size] + + logits = model.joiner( + current_encoder_out, decoder_out.unsqueeze(1), project_input=False + ) + # logits'shape (batch_size, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) # (batch_size, vocab_size) + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v not in (blank_id, unk_id): + hyps[i].append(v) + timestamps[i].append(t) + emitted = True + if emitted: + # update decoder output + decoder_input = [h[-context_size:] for h in hyps[:batch_size]] + decoder_input = torch.tensor( + decoder_input, + device=device, + dtype=torch.int64, + ) + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + sorted_ans = [h[context_size:] for h in hyps] + ans = [] + ans_timestamps = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + ans_timestamps.append(timestamps[unsorted_indices[i]]) + + if not return_timestamps: + return ans + else: + return DecodingResults( + hyps=ans, + timestamps=ans_timestamps, + ) + + +def modified_beam_search( + model: SURT, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: int = 4, + temperature: float = 1.0, + return_timestamps: bool = False, +) -> Union[List[List[int]], DecodingResults]: + """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded. + + Args: + model: + The SURT model. + encoder_out: + Output from the encoder. Its shape is (N, T, C). + encoder_out_lens: + A 1-D tensor of shape (N,), containing number of valid frames in + encoder_out before padding. + beam: + Number of active paths during the beam search. + temperature: + Softmax temperature. + return_timestamps: + Whether to return timestamps. + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert encoder_out.size(0) >= 1, encoder_out.size(0) + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + device = next(model.parameters()).device + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + B = [HypothesisList() for _ in range(N)] + for i in range(N): + B[i].add( + Hypothesis( + ys=[blank_id] * context_size, + log_prob=torch.zeros(1, dtype=torch.float32, device=device), + timestamp=[], + ) + ) + + encoder_out = model.joiner.encoder_proj(packed_encoder_out.data) + + offset = 0 + finalized_B = [] + for (t, batch_size) in enumerate(batch_size_list): + start = offset + end = offset + batch_size + current_encoder_out = encoder_out.data[start:end] + current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim) + offset = end + + finalized_B = B[batch_size:] + finalized_B + B = B[:batch_size] + + hyps_shape = get_hyps_shape(B).to(device) + + A = [list(b) for b in B] + B = [HypothesisList() for _ in range(batch_size)] + + ys_log_probs = torch.cat( + [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps] + ) # (num_hyps, 1) + + decoder_input = torch.tensor( + [hyp.ys[-context_size:] for hyps in A for hyp in hyps], + device=device, + dtype=torch.int64, + ) # (num_hyps, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_out is of shape (num_hyps, 1, 1, joiner_dim) + + # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor + # as index, so we use `to(torch.int64)` below. + current_encoder_out = torch.index_select( + current_encoder_out, + dim=0, + index=hyps_shape.row_ids(1).to(torch.int64), + ) # (num_hyps, 1, 1, encoder_out_dim) + + logits = model.joiner( + current_encoder_out, + decoder_out, + project_input=False, + ) # (num_hyps, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) # (num_hyps, vocab_size) + + log_probs = (logits / temperature).log_softmax(dim=-1) # (num_hyps, vocab_size) + + log_probs.add_(ys_log_probs) + + vocab_size = log_probs.size(-1) + + log_probs = log_probs.reshape(-1) + + row_splits = hyps_shape.row_splits(1) * vocab_size + log_probs_shape = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=log_probs.numel() + ) + ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs) + + for i in range(batch_size): + topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() + + for k in range(len(topk_hyp_indexes)): + hyp_idx = topk_hyp_indexes[k] + hyp = A[i][hyp_idx] + + new_ys = hyp.ys[:] + new_token = topk_token_indexes[k] + new_timestamp = hyp.timestamp[:] + if new_token not in (blank_id, unk_id): + new_ys.append(new_token) + new_timestamp.append(t) + + new_log_prob = topk_log_probs[k] + new_hyp = Hypothesis( + ys=new_ys, log_prob=new_log_prob, timestamp=new_timestamp + ) + B[i].add(new_hyp) + + B = B + finalized_B + best_hyps = [b.get_most_probable(length_norm=True) for b in B] + + sorted_ans = [h.ys[context_size:] for h in best_hyps] + sorted_timestamps = [h.timestamp for h in best_hyps] + ans = [] + ans_timestamps = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + ans_timestamps.append(sorted_timestamps[unsorted_indices[i]]) + + if not return_timestamps: + return ans + else: + return DecodingResults( + hyps=ans, + timestamps=ans_timestamps, + ) + + +def beam_search( + model: SURT, + encoder_out: torch.Tensor, + beam: int = 4, + temperature: float = 1.0, + return_timestamps: bool = False, +) -> Union[List[int], DecodingResults]: + """ + It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf + + espnet/nets/beam_search_SURT.py#L247 is used as a reference. + + Args: + model: + An instance of `SURT`. + encoder_out: + A tensor of shape (N, T, C) from the encoder. Support only N==1 for now. + beam: + Beam size. + temperature: + Softmax temperature. + return_timestamps: + Whether to return timestamps. + + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + assert encoder_out.ndim == 3 + + # support only batch_size == 1 for now + assert encoder_out.size(0) == 1, encoder_out.size(0) + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + + device = next(model.parameters()).device + + decoder_input = torch.tensor( + [blank_id] * context_size, + device=device, + dtype=torch.int64, + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + T = encoder_out.size(1) + t = 0 + + B = HypothesisList() + B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0, timestamp=[])) + + max_sym_per_utt = 20000 + + sym_per_utt = 0 + + decoder_cache: Dict[str, torch.Tensor] = {} + + while t < T and sym_per_utt < max_sym_per_utt: + # fmt: off + current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2) + # fmt: on + A = B + B = HypothesisList() + + joint_cache: Dict[str, torch.Tensor] = {} + + # TODO(fangjun): Implement prefix search to update the `log_prob` + # of hypotheses in A + + while True: + y_star = A.get_most_probable() + A.remove(y_star) + + cached_key = y_star.key + + if cached_key not in decoder_cache: + decoder_input = torch.tensor( + [y_star.ys[-context_size:]], + device=device, + dtype=torch.int64, + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + decoder_cache[cached_key] = decoder_out + else: + decoder_out = decoder_cache[cached_key] + + cached_key += f"-t-{t}" + if cached_key not in joint_cache: + logits = model.joiner( + current_encoder_out, + decoder_out.unsqueeze(1), + project_input=False, + ) + + # TODO(fangjun): Scale the blank posterior + log_prob = (logits / temperature).log_softmax(dim=-1) + # log_prob is (1, 1, 1, vocab_size) + log_prob = log_prob.squeeze() + # Now log_prob is (vocab_size,) + joint_cache[cached_key] = log_prob + else: + log_prob = joint_cache[cached_key] + + # First, process the blank symbol + skip_log_prob = log_prob[blank_id] + new_y_star_log_prob = y_star.log_prob + skip_log_prob + + # ys[:] returns a copy of ys + B.add( + Hypothesis( + ys=y_star.ys[:], + log_prob=new_y_star_log_prob, + timestamp=y_star.timestamp[:], + ) + ) + + # Second, process other non-blank labels + values, indices = log_prob.topk(beam + 1) + for i, v in zip(indices.tolist(), values.tolist()): + if i in (blank_id, unk_id): + continue + new_ys = y_star.ys + [i] + new_log_prob = y_star.log_prob + v + new_timestamp = y_star.timestamp + [t] + A.add( + Hypothesis( + ys=new_ys, + log_prob=new_log_prob, + timestamp=new_timestamp, + ) + ) + + # Check whether B contains more than "beam" elements more probable + # than the most probable in A + A_most_probable = A.get_most_probable() + + kept_B = B.filter(A_most_probable.log_prob) + + if len(kept_B) >= beam: + B = kept_B.topk(beam) + break + + t += 1 + + best_hyp = B.get_most_probable(length_norm=True) + ys = best_hyp.ys[context_size:] # [context_size:] to remove blanks + + if not return_timestamps: + return ys + else: + return DecodingResults(hyps=[ys], timestamps=[best_hyp.timestamp]) + + +@dataclass +class Hypothesis: + # The predicted tokens so far. + # Newly predicted tokens are appended to `ys`. + ys: List[int] + + # The log prob of ys. + # It contains only one entry. + log_prob: torch.Tensor + + # timestamp[i] is the frame index after subsampling + # on which ys[i] is decoded + timestamp: List[int] = field(default_factory=list) + + # the lm score for next token given the current ys + lm_score: Optional[torch.Tensor] = None + + # the RNNLM states (h and c in LSTM) + state: Optional[Tuple[torch.Tensor, torch.Tensor]] = None + + # N-gram LM state + state_cost: Optional[NgramLmStateCost] = None + + @property + def key(self) -> str: + """Return a string representation of self.ys""" + return "_".join(map(str, self.ys)) + + +class HypothesisList(object): + def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None: + """ + Args: + data: + A dict of Hypotheses. Its key is its `value.key`. + """ + if data is None: + self._data = {} + else: + self._data = data + + @property + def data(self) -> Dict[str, Hypothesis]: + return self._data + + def add(self, hyp: Hypothesis) -> None: + """Add a Hypothesis to `self`. + + If `hyp` already exists in `self`, its probability is updated using + `log-sum-exp` with the existed one. + + Args: + hyp: + The hypothesis to be added. + """ + key = hyp.key + if key in self: + old_hyp = self._data[key] # shallow copy + torch.logaddexp(old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob) + else: + self._data[key] = hyp + + def get_most_probable(self, length_norm: bool = False) -> Hypothesis: + """Get the most probable hypothesis, i.e., the one with + the largest `log_prob`. + + Args: + length_norm: + If True, the `log_prob` of a hypothesis is normalized by the + number of tokens in it. + Returns: + Return the hypothesis that has the largest `log_prob`. + """ + if length_norm: + return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys)) + else: + return max(self._data.values(), key=lambda hyp: hyp.log_prob) + + def remove(self, hyp: Hypothesis) -> None: + """Remove a given hypothesis. + + Caution: + `self` is modified **in-place**. + + Args: + hyp: + The hypothesis to be removed from `self`. + Note: It must be contained in `self`. Otherwise, + an exception is raised. + """ + key = hyp.key + assert key in self, f"{key} does not exist" + del self._data[key] + + def filter(self, threshold: torch.Tensor) -> "HypothesisList": + """Remove all Hypotheses whose log_prob is less than threshold. + + Caution: + `self` is not modified. Instead, a new HypothesisList is returned. + + Returns: + Return a new HypothesisList containing all hypotheses from `self` + with `log_prob` being greater than the given `threshold`. + """ + ans = HypothesisList() + for _, hyp in self._data.items(): + if hyp.log_prob > threshold: + ans.add(hyp) # shallow copy + return ans + + def topk(self, k: int) -> "HypothesisList": + """Return the top-k hypothesis.""" + hyps = list(self._data.items()) + + hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k] + + ans = HypothesisList(dict(hyps)) + return ans + + def __contains__(self, key: str): + return key in self._data + + def __iter__(self): + return iter(self._data.values()) + + def __len__(self) -> int: + return len(self._data) + + def __str__(self) -> str: + s = [] + for key in self: + s.append(key) + return ", ".join(s) + + +def get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape: + """Return a ragged shape with axes [utt][num_hyps]. + + Args: + hyps: + len(hyps) == batch_size. It contains the current hypothesis for + each utterance in the batch. + Returns: + Return a ragged shape with 2 axes [utt][num_hyps]. Note that + the shape is on CPU. + """ + num_hyps = [len(h) for h in hyps] + + # torch.cumsum() is inclusive sum, so we put a 0 at the beginning + # to get exclusive sum later. + num_hyps.insert(0, 0) + + num_hyps = torch.tensor(num_hyps) + row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32) + ans = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=row_splits[-1].item() + ) + return ans diff --git a/egs/libricss/SURT/dprnn_zipformer/decode.py b/egs/libricss/SURT/dprnn_zipformer/decode.py new file mode 100755 index 000000000..6abbffe00 --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/decode.py @@ -0,0 +1,654 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./dprnn_zipformer/decode.py \ + --epoch 30 \ + --avg 9 \ + --use-averaged-model true \ + --exp-dir ./dprnn_zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) modified beam search +./dprnn_zipformer/decode.py \ + --epoch 30 \ + --avg 9 \ + --use-averaged-model true \ + --exp-dir ./dprnn_zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 +""" + + +import argparse +import logging +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriCssAsrDataModule +from beam_search import ( + beam_search, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from lhotse.utils import EPSILON +from train import add_model_arguments, get_params, get_surt_model + +from icefall import LmScorer, NgramLm +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_surt_error_stats, +) + +OVERLAP_RATIOS = ["0L", "0S", "OV10", "OV20", "OV30", "OV40"] + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="dprnn_zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--save-masks", + type=str2bool, + default=False, + help="""If true, save masks generated by unmixing module.""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + feature_lens = batch["input_lens"].to(device) + + # Apply the mask encoder + B, T, F = feature.shape + processed = model.mask_encoder(feature) # B,T,F*num_channels + masks = processed.view(B, T, F, params.num_channels).unbind(dim=-1) + x_masked = [feature * m for m in masks] + + masks_dict = {} + if params.save_masks: + # To save the masks, we split them by batch and trim each mask to the length of + # the corresponding feature. We save them in a dict, where the key is the + # cut ID and the value is the mask. + for i in range(B): + mask = torch.cat( + [x_masked[j][i, : feature_lens[i]] for j in range(params.num_channels)], + dim=-1, + ) + mask = mask.cpu().numpy() + masks_dict[batch["cuts"][i].id] = mask + + # Recognition + # Concatenate the inputs along the batch axis + h = torch.cat(x_masked, dim=0) + h_lens = feature_lens.repeat(params.num_channels) + encoder_out, encoder_out_lens = model.encoder(x=h, x_lens=h_lens) + + if model.joint_encoder_layer is not None: + encoder_out = model.joint_encoder_layer(encoder_out) + + def _group_channels(hyps: List[str]) -> List[List[str]]: + """ + Currently we have a batch of size M*B, where M is the number of + channels and B is the batch size. We need to group the hypotheses + into B groups, each of which contains M hypotheses. + + Example: + hyps = ['a1', 'b1', 'c1', 'a2', 'b2', 'c2'] + _group_channels(hyps) = [['a1', 'a2'], ['b1', 'b2'], ['c1', 'c2']] + """ + assert len(hyps) == B * params.num_channels + out_hyps = [] + for i in range(B): + out_hyps.append(hyps[i::B]) + return out_hyps + + hyps = [] + if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp)) + + if params.decoding_method == "greedy_search": + return {"greedy_search": _group_channels(hyps)}, masks_dict + else: + return {f"beam_size_{params.beam_size}": _group_channels(hyps)}, masks_dict + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + masks = {} + for batch_idx, batch in enumerate(dl): + cut_ids = [cut.id for cut in batch["cuts"]] + cuts_batch = batch["cuts"] + + hyps_dict, masks_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + ) + masks.update(masks_dict) + + for name, hyps in hyps_dict.items(): + this_batch = [] + for cut_id, hyp_words in zip(cut_ids, hyps): + # Reference is a list of supervision texts sorted by start time. + ref_words = [ + s.text.strip() + for s in sorted( + cuts_batch[cut_id].supervisions, key=lambda s: s.start + ) + ] + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(cut_ids) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results, masks_dict + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_surt_error_stats( + f, + f"{test_set_name}-{key}", + results, + enable_log=True, + num_channels=params.num_channels, + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +def save_masks( + params: AttributeDict, + test_set_name: str, + masks: List[torch.Tensor], +): + masks_path = params.res_dir / f"masks-{test_set_name}.txt" + torch.save(masks, masks_path) + logging.info(f"The masks are stored in {masks_path}") + + +@torch.no_grad() +def main(): + parser = get_parser() + LmScorer.add_arguments(parser) + LibriCssAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "modified_beam_search", + ), f"Decoding method {params.decoding_method} is not supported." + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_surt_model(params) + assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, ( + model.encoder.decode_chunk_size, + params.decode_chunk_len, + ) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + libricss = LibriCssAsrDataModule(args) + + dev_cuts = libricss.libricss_cuts(split="dev", type="ihm-mix").to_eager() + dev_cuts_grouped = [dev_cuts.filter(lambda x: ol in x.id) for ol in OVERLAP_RATIOS] + test_cuts = libricss.libricss_cuts(split="test", type="ihm-mix").to_eager() + test_cuts_grouped = [ + test_cuts.filter(lambda x: ol in x.id) for ol in OVERLAP_RATIOS + ] + + for dev_set, ol in zip(dev_cuts_grouped, OVERLAP_RATIOS): + dev_dl = libricss.test_dataloaders(dev_set) + results_dict, masks = decode_dataset( + dl=dev_dl, + params=params, + model=model, + sp=sp, + ) + + save_results( + params=params, + test_set_name=f"dev_{ol}", + results_dict=results_dict, + ) + + if params.save_masks: + save_masks( + params=params, + test_set_name=f"dev_{ol}", + masks=masks, + ) + + for test_set, ol in zip(test_cuts_grouped, OVERLAP_RATIOS): + test_dl = libricss.test_dataloaders(test_set) + results_dict, masks = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + ) + + save_results( + params=params, + test_set_name=f"test_{ol}", + results_dict=results_dict, + ) + + if params.save_masks: + save_masks( + params=params, + test_set_name=f"test_{ol}", + masks=masks, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/libricss/SURT/dprnn_zipformer/decoder.py b/egs/libricss/SURT/dprnn_zipformer/decoder.py new file mode 120000 index 000000000..8283d8c5a --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless7/decoder.py \ No newline at end of file diff --git a/egs/libricss/SURT/dprnn_zipformer/dprnn.py b/egs/libricss/SURT/dprnn_zipformer/dprnn.py new file mode 100644 index 000000000..440dea885 --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/dprnn.py @@ -0,0 +1,305 @@ +import random +from typing import Optional, Tuple + +import torch +import torch.nn as nn +from einops import rearrange +from scaling import ActivationBalancer, BasicNorm, DoubleSwish, ScaledLinear, ScaledLSTM +from torch.autograd import Variable + +EPS = torch.finfo(torch.get_default_dtype()).eps + + +def _pad_segment(input, segment_size): + # Source: https://github.com/espnet/espnet/blob/master/espnet2/enh/layers/dprnn.py#L342 + # input is the features: (B, N, T) + batch_size, dim, seq_len = input.shape + segment_stride = segment_size // 2 + + rest = segment_size - (segment_stride + seq_len % segment_size) % segment_size + if rest > 0: + pad = Variable(torch.zeros(batch_size, dim, rest)).type(input.type()) + input = torch.cat([input, pad], 2) + + pad_aux = Variable(torch.zeros(batch_size, dim, segment_stride)).type(input.type()) + input = torch.cat([pad_aux, input, pad_aux], 2) + + return input, rest + + +def split_feature(input, segment_size): + # Source: https://github.com/espnet/espnet/blob/master/espnet2/enh/layers/dprnn.py#L358 + # split the feature into chunks of segment size + # input is the features: (B, N, T) + + input, rest = _pad_segment(input, segment_size) + batch_size, dim, seq_len = input.shape + segment_stride = segment_size // 2 + + segments1 = ( + input[:, :, :-segment_stride] + .contiguous() + .view(batch_size, dim, -1, segment_size) + ) + segments2 = ( + input[:, :, segment_stride:] + .contiguous() + .view(batch_size, dim, -1, segment_size) + ) + segments = ( + torch.cat([segments1, segments2], 3) + .view(batch_size, dim, -1, segment_size) + .transpose(2, 3) + ) + + return segments.contiguous(), rest + + +def merge_feature(input, rest): + # Source: https://github.com/espnet/espnet/blob/master/espnet2/enh/layers/dprnn.py#L385 + # merge the splitted features into full utterance + # input is the features: (B, N, L, K) + + batch_size, dim, segment_size, _ = input.shape + segment_stride = segment_size // 2 + input = ( + input.transpose(2, 3).contiguous().view(batch_size, dim, -1, segment_size * 2) + ) # B, N, K, L + + input1 = ( + input[:, :, :, :segment_size] + .contiguous() + .view(batch_size, dim, -1)[:, :, segment_stride:] + ) + input2 = ( + input[:, :, :, segment_size:] + .contiguous() + .view(batch_size, dim, -1)[:, :, :-segment_stride] + ) + + output = input1 + input2 + if rest > 0: + output = output[:, :, :-rest] + + return output.contiguous() # B, N, T + + +class RNNEncoderLayer(nn.Module): + """ + RNNEncoderLayer is made up of lstm and feedforward networks. + Args: + input_size: + The number of expected features in the input (required). + hidden_size: + The hidden dimension of rnn layer. + dropout: + The dropout value (default=0.1). + layer_dropout: + The dropout value for model-level warmup (default=0.075). + """ + + def __init__( + self, + input_size: int, + hidden_size: int, + dropout: float = 0.1, + bidirectional: bool = False, + ) -> None: + super(RNNEncoderLayer, self).__init__() + self.input_size = input_size + self.hidden_size = hidden_size + + assert hidden_size >= input_size, (hidden_size, input_size) + self.lstm = ScaledLSTM( + input_size=input_size, + hidden_size=hidden_size // 2 if bidirectional else hidden_size, + proj_size=0, + num_layers=1, + dropout=0.0, + batch_first=True, + bidirectional=bidirectional, + ) + self.norm_final = BasicNorm(input_size) + + # try to ensure the output is close to zero-mean (or at least, zero-median). # noqa + self.balancer = ActivationBalancer( + num_channels=input_size, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + max_abs=6.0, + ) + self.dropout = nn.Dropout(dropout) + + def forward( + self, + src: torch.Tensor, + states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + warmup: float = 1.0, + ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]: + """ + Pass the input through the encoder layer. + Args: + src: + The sequence to the encoder layer (required). + Its shape is (S, N, E), where S is the sequence length, + N is the batch size, and E is the feature number. + states: + A tuple of 2 tensors (optional). It is for streaming inference. + states[0] is the hidden states of all layers, + with shape of (1, N, input_size); + states[1] is the cell states of all layers, + with shape of (1, N, hidden_size). + """ + src_orig = src + + # alpha = 1.0 means fully use this encoder layer, 0.0 would mean + # completely bypass it. + alpha = warmup if self.training else 1.0 + + # lstm module + src_lstm, new_states = self.lstm(src, states) + src = self.dropout(src_lstm) + src + src = self.norm_final(self.balancer(src)) + + if alpha != 1.0: + src = alpha * src + (1 - alpha) * src_orig + + return src + + +# dual-path RNN +class DPRNN(nn.Module): + """Deep dual-path RNN. + Source: https://github.com/espnet/espnet/blob/master/espnet2/enh/layers/dprnn.py + + args: + input_size: int, dimension of the input feature. The input should have shape + (batch, seq_len, input_size). + hidden_size: int, dimension of the hidden state. + output_size: int, dimension of the output size. + dropout: float, dropout ratio. Default is 0. + num_blocks: int, number of stacked RNN layers. Default is 1. + """ + + def __init__( + self, + feature_dim, + input_size, + hidden_size, + output_size, + dropout=0.1, + num_blocks=1, + segment_size=50, + chunk_width_randomization=False, + ): + super().__init__() + + self.input_size = input_size + self.output_size = output_size + self.hidden_size = hidden_size + + self.segment_size = segment_size + self.chunk_width_randomization = chunk_width_randomization + + self.input_embed = nn.Sequential( + ScaledLinear(feature_dim, input_size), + BasicNorm(input_size), + ActivationBalancer( + num_channels=input_size, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + ), + ) + + # dual-path RNN + self.row_rnn = nn.ModuleList([]) + self.col_rnn = nn.ModuleList([]) + for _ in range(num_blocks): + # intra-RNN is non-causal + self.row_rnn.append( + RNNEncoderLayer( + input_size, hidden_size, dropout=dropout, bidirectional=True + ) + ) + self.col_rnn.append( + RNNEncoderLayer( + input_size, hidden_size, dropout=dropout, bidirectional=False + ) + ) + + # output layer + self.out_embed = nn.Sequential( + ScaledLinear(input_size, output_size), + BasicNorm(output_size), + ActivationBalancer( + num_channels=output_size, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + ), + ) + + def forward(self, input): + # input shape: B, T, F + input = self.input_embed(input) + B, T, D = input.shape + + if self.chunk_width_randomization and self.training: + segment_size = random.randint(self.segment_size // 2, self.segment_size) + else: + segment_size = self.segment_size + input, rest = split_feature(input.transpose(1, 2), segment_size) + # input shape: batch, N, dim1, dim2 + # apply RNN on dim1 first and then dim2 + # output shape: B, output_size, dim1, dim2 + # input = input.to(device) + batch_size, _, dim1, dim2 = input.shape + output = input + for i in range(len(self.row_rnn)): + row_input = ( + output.permute(0, 3, 2, 1) + .contiguous() + .view(batch_size * dim2, dim1, -1) + ) # B*dim2, dim1, N + output = self.row_rnn[i](row_input) # B*dim2, dim1, H + output = ( + output.view(batch_size, dim2, dim1, -1).permute(0, 3, 2, 1).contiguous() + ) # B, N, dim1, dim2 + + col_input = ( + output.permute(0, 2, 3, 1) + .contiguous() + .view(batch_size * dim1, dim2, -1) + ) # B*dim1, dim2, N + output = self.col_rnn[i](col_input) # B*dim1, dim2, H + output = ( + output.view(batch_size, dim1, dim2, -1).permute(0, 3, 1, 2).contiguous() + ) # B, N, dim1, dim2 + + output = merge_feature(output, rest) + output = output.transpose(1, 2) + output = self.out_embed(output) + + # Apply ReLU to the output + output = torch.relu(output) + + return output + + +if __name__ == "__main__": + + model = DPRNN( + 80, + 256, + 256, + 160, + dropout=0.1, + num_blocks=4, + segment_size=32, + chunk_width_randomization=True, + ) + input = torch.randn(2, 1002, 80) + print(sum(p.numel() for p in model.parameters())) + print(model(input).shape) diff --git a/egs/libricss/SURT/dprnn_zipformer/encoder_interface.py b/egs/libricss/SURT/dprnn_zipformer/encoder_interface.py new file mode 120000 index 000000000..0c2673d46 --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py \ No newline at end of file diff --git a/egs/libricss/SURT/dprnn_zipformer/export.py b/egs/libricss/SURT/dprnn_zipformer/export.py new file mode 100755 index 000000000..f51f2a7ab --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/export.py @@ -0,0 +1,306 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script converts several saved checkpoints +# to a single one using model averaging. +""" + +Usage: + +(1) Export to torchscript model using torch.jit.script() + +./dprnn_zipformer/export.py \ + --exp-dir ./dprnn_zipformer/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 30 \ + --avg 9 \ + --jit 1 + +It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later +load it by `torch.jit.load("cpu_jit.pt")`. + +Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python +are on CPU. You can use `to("cuda")` to move them to a CUDA device. + +Check +https://github.com/k2-fsa/sherpa +for how to use the exported models outside of icefall. + +(2) Export `model.state_dict()` + +./dprnn_zipformer/export.py \ + --exp-dir ./dprnn_zipformer/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 30 \ + --avg 9 + +It will generate a file `pretrained.pt` in the given `exp_dir`. You can later +load it by `icefall.checkpoint.load_checkpoint()`. + +To use the generated file with `dprnn_zipformer/decode.py`, +you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + ./dprnn_zipformer/decode.py \ + --exp-dir ./dprnn_zipformer/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 600 \ + --decoding-method greedy_search \ + --bpe-model data/lang_bpe_500/bpe.model +""" + +import argparse +import logging +from pathlib import Path + +import sentencepiece as spm +import torch +import torch.nn as nn +from scaling_converter import convert_scaled_to_non_scaled +from train import add_model_arguments, get_params, get_surt_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="dprnn_zipformer/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + It will generate a file named cpu_jit.pt + + Check ./jit_pretrained.py for how to use it. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + add_model_arguments(parser) + + return parser + + +@torch.no_grad() +def main(): + args = get_parser().parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_surt_model(params) + + model.to(device) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to("cpu") + model.eval() + + if params.jit is True: + convert_scaled_to_non_scaled(model, inplace=True) + # We won't use the forward() method of the model in C++, so just ignore + # it here. + # Otherwise, one of its arguments is a ragged tensor and is not + # torch scriptabe. + model.__class__.forward = torch.jit.ignore(model.__class__.forward) + logging.info("Using torch.jit.script") + model = torch.jit.script(model) + filename = params.exp_dir / "cpu_jit.pt" + model.save(str(filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torchscript. Export model.state_dict()") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/libricss/SURT/dprnn_zipformer/joiner.py b/egs/libricss/SURT/dprnn_zipformer/joiner.py new file mode 120000 index 000000000..0f0c3c90a --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless7/joiner.py \ No newline at end of file diff --git a/egs/libricss/SURT/dprnn_zipformer/model.py b/egs/libricss/SURT/dprnn_zipformer/model.py new file mode 100644 index 000000000..688e1e78d --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/model.py @@ -0,0 +1,316 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, Wei Kang) +# Copyright 2023 Johns Hopkins University (author: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import List, Optional, Tuple + +import k2 +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface + +from icefall.utils import add_sos + + +class SURT(nn.Module): + """It implements Streaming Unmixing and Recognition Transducer (SURT). + https://arxiv.org/abs/2011.13148 + """ + + def __init__( + self, + mask_encoder: nn.Module, + encoder: EncoderInterface, + joint_encoder_layer: Optional[nn.Module], + decoder: nn.Module, + joiner: nn.Module, + num_channels: int, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + ): + """ + Args: + mask_encoder: + It is the masking network. It generates a mask for each channel of the + encoder. These masks are applied to the input features, and then passed + to the transcription network. + encoder: + It is the transcription network in the paper. Its accepts + two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, encoder_dm) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, decoder_dim). + It should contain one attribute: `blank_id`. + joiner: + It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim). + Its output shape is (N, T, U, vocab_size). Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + num_channels: + It is the number of channels that the input features will be split into. + In general, it should be equal to the maximum number of simultaneously + active speakers. For most real scenarios, using 2 channels is sufficient. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + assert hasattr(decoder, "blank_id") + + self.mask_encoder = mask_encoder + self.encoder = encoder + self.joint_encoder_layer = joint_encoder_layer + self.decoder = decoder + self.joiner = joiner + self.num_channels = num_channels + + self.simple_am_proj = nn.Linear( + encoder_dim, + vocab_size, + ) + self.simple_lm_proj = nn.Linear(decoder_dim, vocab_size) + + self.ctc_output = nn.Sequential( + nn.Dropout(p=0.1), + nn.Linear(encoder_dim, vocab_size), + nn.LogSoftmax(dim=-1), + ) + + def forward_helper( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + reduction: str = "sum", + beam_size: int = 10, + use_double_scores: bool = False, + subsampling_factor: int = 1, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Compute transducer loss for one branch of the SURT model. + """ + encoder_out, x_lens = self.encoder(x, x_lens) + assert torch.all(x_lens > 0) + + if self.joint_encoder_layer is not None: + encoder_out = self.joint_encoder_layer(encoder_out) + + # compute ctc log-probs + ctc_output = self.ctc_output(encoder_out) + + # For the decoder, i.e., the prediction network + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + # sos_y_padded: [B, S + 1], start with SOS. + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + # decoder_out: [B, S + 1, decoder_dim] + decoder_out = self.decoder(sos_y_padded) + + # Note: y does not start with SOS + # y_padded : [B, S] + y_padded = y.pad(mode="constant", padding_value=0) + + y_padded = y_padded.to(torch.int64) + boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device) + boundary[:, 2] = y_lens + boundary[:, 3] = x_lens + + lm = self.simple_lm_proj(decoder_out) + am = self.simple_am_proj(encoder_out) + + with torch.cuda.amp.autocast(enabled=False): + simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed( + lm=lm.float(), + am=am.float(), + symbols=y_padded, + termination_symbol=blank_id, + lm_only_scale=lm_scale, + am_only_scale=am_scale, + boundary=boundary, + reduction=reduction, + return_grad=True, + ) + + # ranges : [B, T, prune_range] + ranges = k2.get_rnnt_prune_ranges( + px_grad=px_grad, + py_grad=py_grad, + boundary=boundary, + s_range=prune_range, + ) + + # am_pruned : [B, T, prune_range, encoder_dim] + # lm_pruned : [B, T, prune_range, decoder_dim] + am_pruned, lm_pruned = k2.do_rnnt_pruning( + am=self.joiner.encoder_proj(encoder_out), + lm=self.joiner.decoder_proj(decoder_out), + ranges=ranges, + ) + + # logits : [B, T, prune_range, vocab_size] + + # project_input=False since we applied the decoder's input projections + # prior to do_rnnt_pruning (this is an optimization for speed). + logits = self.joiner(am_pruned, lm_pruned, project_input=False) + + with torch.cuda.amp.autocast(enabled=False): + pruned_loss = k2.rnnt_loss_pruned( + logits=logits.float(), + symbols=y_padded, + ranges=ranges, + termination_symbol=blank_id, + boundary=boundary, + reduction=reduction, + ) + + # Compute ctc loss + supervision_segments = torch.stack( + ( + torch.arange(len(x_lens), device="cpu"), + torch.zeros_like(x_lens, device="cpu"), + torch.clone(x_lens).detach().cpu(), + ), + dim=1, + ).to(torch.int32) + # We need to sort supervision_segments in decreasing order of num_frames + indices = torch.argsort(supervision_segments[:, 2], descending=True) + supervision_segments = supervision_segments[indices] + + # Works with a BPE model + decoding_graph = k2.ctc_graph(y, modified=False, device=x.device) + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments, + allow_truncate=subsampling_factor - 1, + ) + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=beam_size, + reduction="none", + use_double_scores=use_double_scores, + ) + + return (simple_loss, pruned_loss, ctc_loss) + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + reduction: str = "sum", + beam_size: int = 10, + use_double_scores: bool = False, + subsampling_factor: int = 1, + return_masks: bool = False, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + y: + A ragged tensor of shape (N*num_channels, S). It contains the labels + of the N utterances. The labels are in the range [0, vocab_size). All + the channels are concatenated together one after another. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + reduction: + "sum" to sum the losses over all utterances in the batch. + "none" to return the loss in a 1-D tensor for each utterance + in the batch. + beam_size: + The beam size used in CTC decoding. + use_double_scores: + If True, use double precision for CTC decoding. + subsampling_factor: + The subsampling factor of the model. It is used to compute the + supervision segments for CTC loss. + return_masks: + If True, return the masks as well as masked features. + Returns: + Return the transducer loss. + + Note: + Regarding am_scale & lm_scale, it will make the loss-function one of + the form: + lm_scale * lm_probs + am_scale * am_probs + + (1-lm_scale-am_scale) * combined_probs + """ + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == x_lens.size(0), (x.size(), x_lens.size()) + + # Apply the mask encoder + B, T, F = x.shape + processed = self.mask_encoder(x) # B,T,F*num_channels + masks = processed.view(B, T, F, self.num_channels).unbind(dim=-1) + x_masked = [x * m for m in masks] + + # Recognition + # Stack the inputs along the batch axis + h = torch.cat(x_masked, dim=0) + h_lens = torch.cat([x_lens for _ in range(self.num_channels)], dim=0) + + simple_loss, pruned_loss, ctc_loss = self.forward_helper( + h, + h_lens, + y, + prune_range, + am_scale, + lm_scale, + reduction=reduction, + beam_size=beam_size, + use_double_scores=use_double_scores, + subsampling_factor=subsampling_factor, + ) + + # Chunks the outputs into 2 parts along batch axis and then stack them along a new axis. + simple_loss = torch.stack( + torch.chunk(simple_loss, self.num_channels, dim=0), dim=0 + ) + pruned_loss = torch.stack( + torch.chunk(pruned_loss, self.num_channels, dim=0), dim=0 + ) + ctc_loss = torch.stack(torch.chunk(ctc_loss, self.num_channels, dim=0), dim=0) + + if return_masks: + return (simple_loss, pruned_loss, ctc_loss, x_masked, masks) + else: + return (simple_loss, pruned_loss, ctc_loss, x_masked) diff --git a/egs/libricss/SURT/dprnn_zipformer/optim.py b/egs/libricss/SURT/dprnn_zipformer/optim.py new file mode 120000 index 000000000..8a05abb5f --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless7/optim.py \ No newline at end of file diff --git a/egs/libricss/SURT/dprnn_zipformer/scaling.py b/egs/libricss/SURT/dprnn_zipformer/scaling.py new file mode 120000 index 000000000..5f9be9fe0 --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless7/scaling.py \ No newline at end of file diff --git a/egs/libricss/SURT/dprnn_zipformer/scaling_converter.py b/egs/libricss/SURT/dprnn_zipformer/scaling_converter.py new file mode 120000 index 000000000..f9960e5c6 --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py \ No newline at end of file diff --git a/egs/libricss/SURT/dprnn_zipformer/train.py b/egs/libricss/SURT/dprnn_zipformer/train.py new file mode 100755 index 000000000..6598f8b5d --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/train.py @@ -0,0 +1,1452 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# 2023 Johns Hopkins University (author: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +cd egs/libricss/SURT +./prepare.sh + +./dprnn_zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir dprnn_zipformer/exp \ + --max-duration 300 + +# For mix precision training: + +./dprnn_zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir dprnn_zipformer/exp \ + --max-duration 550 +""" + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriCssAsrDataModule +from decoder import Decoder +from dprnn import DPRNN +from einops.layers.torch import Rearrange +from graph_pit.loss.optimized import optimized_graph_pit_mse_loss as gpit_mse +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import LOG_EPSILON, fix_random_seed +from model import SURT +from optim import Eden, ScaledAdam +from scaling import ScaledLSTM +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-mask-encoder-layers", + type=int, + default=4, + help="Number of layers in the DPRNN based mask encoder.", + ) + + parser.add_argument( + "--mask-encoder-dim", + type=int, + default=256, + help="Hidden dimension of the LSTM blocks in DPRNN.", + ) + + parser.add_argument( + "--mask-encoder-segment-size", + type=int, + default=32, + help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the " + "decode-chunk-length of the zipformer encoder.", + ) + + parser.add_argument( + "--chunk-width-randomization", + type=bool, + default=False, + help="Whether to randomize the chunk width in DPRNN.", + ) + + # Zipformer config is based on: + # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740 + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,2,2,2", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="768,768,768,768,768", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="256,256,256,256,256", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="192,192,192,192,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--use-joint-encoder-layer", + type=str, + default="lstm", + choices=["linear", "lstm", "none"], + help="Whether to use a joint layer to combine all branches.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--short-chunk-size", + type=int, + default=50, + help="""Chunk length of dynamic training, the chunk size would be either + max sequence length of current batch or uniformly sampled from (1, short_chunk_size). + """, + ) + + parser.add_argument( + "--num-left-chunks", + type=int, + default=4, + help="How many left context can be seen in chunks when calculating attention.", + ) + + parser.add_argument( + "--decode-chunk-len", + type=int, + default=32, + help="The chunk size for decoding (in frames before subsampling)", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="conv_lstm_transducer_stateless_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--model-init-ckpt", + type=str, + default=None, + help="""The model checkpoint to initialize the model (either full or part). + If not specified, the model is randomly initialized. + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.004, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=6, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--heat-loss-scale", + type=float, + default=0.0, + help="Scale for HEAT loss on separated sources.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=1, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=100, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 2000, + # parameters for SURT + "num_channels": 2, + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed + # parameters for Noam + "model_warm_step": 5000, # arg given to model, not for lrate + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "env_info": get_env_info(), + } + ) + + return params + + +def get_mask_encoder_model(params: AttributeDict) -> nn.Module: + mask_encoder = DPRNN( + feature_dim=params.feature_dim, + input_size=params.mask_encoder_dim, + hidden_size=params.mask_encoder_dim, + output_size=params.feature_dim * params.num_channels, + segment_size=params.mask_encoder_segment_size, + num_blocks=params.num_mask_encoder_layers, + chunk_width_randomization=params.chunk_width_randomization, + ) + return mask_encoder + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + num_left_chunks=params.num_left_chunks, + short_chunk_size=params.short_chunk_size, + decode_chunk_size=params.decode_chunk_len // 2, + ) + return encoder + + +def get_joint_encoder_layer(params: AttributeDict) -> nn.Module: + class TakeFirst(nn.Module): + def forward(self, x): + return x[0] + + if params.use_joint_encoder_layer == "linear": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + nn.Linear( + params.num_channels * encoder_dim, params.num_channels * encoder_dim + ), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "lstm": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + ScaledLSTM( + input_size=params.num_channels * encoder_dim, + hidden_size=params.num_channels * encoder_dim, + num_layers=1, + bias=True, + batch_first=True, + dropout=0.0, + bidirectional=False, + ), + TakeFirst(), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "none": + joint_layer = None + else: + raise ValueError( + f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}" + ) + return joint_layer + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_surt_model( + params: AttributeDict, +) -> nn.Module: + mask_encoder = get_mask_encoder_model(params) + encoder = get_encoder_model(params) + joint_layer = get_joint_encoder_layer(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = SURT( + mask_encoder=mask_encoder, + encoder=encoder, + joint_encoder_layer=joint_layer, + decoder=decoder, + joiner=joiner, + num_channels=params.num_channels, + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_heat_loss(x_masked, batch, num_channels=2) -> Tensor: + """ + Compute HEAT loss for separated sources using the output of mask encoder. + Args: + x_masked: + The output of mask encoder. It is a tensor of shape (B, T, C). + batch: + A batch of data. See `lhotse.dataset.K2SurtDatasetWithSources()` + for the content in it. + num_channels: + The number of output branches in the SURT model. + """ + B, T, D = x_masked[0].shape + device = x_masked[0].device + + # Create training targets for each channel. + targets = [] + for i in range(num_channels): + target = torch.ones_like(x_masked[i]) * LOG_EPSILON + targets.append(target) + + source_feats = batch["source_feats"] + source_boundaries = batch["source_boundaries"] + input_lens = batch["input_lens"].to(device) + # Assign sources to channels based on the HEAT criteria + for b in range(B): + cut_source_feats = source_feats[b] + cut_source_boundaries = source_boundaries[b] + last_seg_end = [0 for _ in range(num_channels)] + for source_feat, (start, end) in zip(cut_source_feats, cut_source_boundaries): + assigned = False + for i in range(num_channels): + if start >= last_seg_end[i]: + targets[i][b, start:end, :] += source_feat.to(device) + last_seg_end[i] = max(end, last_seg_end[i]) + assigned = True + break + if not assigned: + min_end_channel = last_seg_end.index(min(last_seg_end)) + targets[min_end_channel][b, start:end, :] += source_feat + last_seg_end[min_end_channel] = max(end, last_seg_end[min_end_channel]) + + # Get padding mask based on input lengths + pad_mask = torch.arange(T, device=device).expand(B, T) > input_lens.unsqueeze(1) + pad_mask = pad_mask.unsqueeze(-1) + + # Compute masked loss for each channel + losses = torch.zeros((num_channels, B, T, D), device=device) + for i in range(num_channels): + loss = nn.functional.mse_loss(x_masked[i], targets[i], reduction="none") + # Apply padding mask to loss + loss.masked_fill_(pad_mask, 0) + losses[i] = loss + + # loss: C x B x T x D. pad_mask: B x T x 1 + # We want to compute loss for each item in the batch. Each item has loss given + # by the sum over C, and average over T and D. For T, we need to use the padding. + loss = losses.sum(0).mean(-1).sum(-1) / batch["input_lens"].to(device) + return loss + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"].to(device) + feature_lens = batch["input_lens"].to(device) + + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + + # The dataloader returns text as a list of cuts, each of which is a list of channel + # text. We flatten this to a list where all channels are together, i.e., it looks like + # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2]. + text = [val for tup in zip(*batch["text"]) for val in tup] + assert len(text) == len(feature) * params.num_channels + + # Convert all channel texts to token IDs and create a ragged tensor. + y = sp.encode(text, out_type=int) + y = k2.RaggedTensor(y).to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.model_warm_step + + with torch.set_grad_enabled(is_training): + (simple_loss, pruned_loss, ctc_loss, x_masked) = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + reduction="none", + subsampling_factor=params.subsampling_factor, + ) + simple_loss_is_finite = torch.isfinite(simple_loss) + pruned_loss_is_finite = torch.isfinite(pruned_loss) + ctc_loss_is_finite = torch.isfinite(ctc_loss) + + # Compute HEAT loss + if is_training and params.heat_loss_scale > 0.0: + heat_loss = compute_heat_loss( + x_masked, batch, num_channels=params.num_channels + ) + else: + heat_loss = torch.tensor(0.0, device=device) + + heat_loss_is_finite = torch.isfinite(heat_loss) + is_finite = ( + simple_loss_is_finite + & pruned_loss_is_finite + & ctc_loss_is_finite + & heat_loss_is_finite + ) + if not torch.all(is_finite): + logging.info( + "Not all losses are finite!\n" + f"simple_losses: {simple_loss}\n" + f"pruned_losses: {pruned_loss}\n" + f"ctc_losses: {ctc_loss}\n" + f"heat_losses: {heat_loss}\n" + ) + display_and_save_batch(batch, params=params, sp=sp) + simple_loss = simple_loss[simple_loss_is_finite] + pruned_loss = pruned_loss[pruned_loss_is_finite] + ctc_loss = ctc_loss[ctc_loss_is_finite] + heat_loss = heat_loss[heat_loss_is_finite] + + # If either all simple_loss or pruned_loss is inf or nan, + # we stop the training process by raising an exception + if ( + torch.all(~simple_loss_is_finite) + or torch.all(~pruned_loss_is_finite) + or torch.all(~ctc_loss_is_finite) + or torch.all(~heat_loss_is_finite) + ): + raise ValueError( + "There are too many utterances in this batch " + "leading to inf or nan losses." + ) + + simple_loss_sum = simple_loss.sum() + pruned_loss_sum = pruned_loss.sum() + ctc_loss_sum = ctc_loss.sum() + heat_loss_sum = heat_loss.sum() + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss = ( + simple_loss_scale * simple_loss_sum + + pruned_loss_scale * pruned_loss_sum + + params.ctc_loss_scale * ctc_loss_sum + + params.heat_loss_scale * heat_loss_sum + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + # info["frames"] is an approximate number for two reasons: + # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2 + # (2) If some utterances in the batch lead to inf/nan loss, they + # are filtered out. + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances` # noqa + info["utterances"] = feature.size(0) + # averaged input duration in frames over utterances + info["utt_duration"] = feature_lens.sum().item() + # averaged padding proportion over utterances + info["utt_pad_proportion"] = ( + ((feature.size(1) - feature_lens) / feature.size(1)).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss_sum.detach().cpu().item() + info["pruned_loss"] = pruned_loss_sum.detach().cpu().item() + if params.ctc_loss_scale > 0.0: + info["ctc_loss"] = ctc_loss_sum.detach().cpu().item() + if params.heat_loss_scale > 0.0: + info["heat_loss"] = heat_loss_sum.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + train_dl_warmup: Optional[torch.utils.data.DataLoader], + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + train_dl_warmup: + Dataloader for the training dataset with 2 speakers. This is used during the + warmup stage. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + torch.cuda.empty_cache() + model.train() + + tot_loss = MetricsTracker() + + iter_train = iter(train_dl) + iter_train_warmup = iter(train_dl_warmup) if train_dl_warmup is not None else None + + batch_idx = 0 + + while True: + # We first sample a batch from the main dataset. This is because we want to + # make sure all epochs have the same number of batches. + try: + batch = next(iter_train) + except StopIteration: + break + + # If we are in warmup stage, get the batch from the warmup dataset. + if ( + params.batch_idx_train <= params.model_warm_step + and iter_train_warmup is not None + ): + try: + batch = next(iter_train_warmup) + except StopIteration: + iter_train_warmup = iter(train_dl_warmup) + batch = next(iter_train_warmup) + + batch_idx += 1 + + params.batch_idx_train += 1 + batch_size = batch["inputs"].shape[0] + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_surt_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + + if checkpoints is None and params.model_init_ckpt is not None: + logging.info( + f"Initializing model with checkpoint from {params.model_init_ckpt}" + ) + init_ckpt = torch.load(params.model_init_ckpt, map_location=device) + model.load_state_dict(init_ckpt["model"], strict=False) + + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + diagnostic = diagnostics.attach_diagnostics(model) + + libricss = LibriCssAsrDataModule(args) + + train_cuts = libricss.lsmix_cuts(rvb_affix="comb", type_affix="full", sources=True) + train_cuts_ov40 = libricss.lsmix_cuts( + rvb_affix="comb", type_affix="ov40", sources=True + ) + dev_cuts = libricss.libricss_cuts(split="dev", type="sdm") + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = libricss.train_dataloaders( + train_cuts, + sampler_state_dict=sampler_state_dict, + ) + train_dl_ov40 = libricss.train_dataloaders(train_cuts_ov40) + valid_dl = libricss.valid_dataloaders(dev_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + train_dl_warmup=train_dl_ov40, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = [sp.encode(text_ch) for text_ch in batch["text"]] + num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y] + logging.info(f"num tokens: {num_tokens}") + + +def main(): + parser = get_parser() + LibriCssAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + +if __name__ == "__main__": + main() diff --git a/egs/libricss/SURT/dprnn_zipformer/train_adapt.py b/egs/libricss/SURT/dprnn_zipformer/train_adapt.py new file mode 100755 index 000000000..1c1b0c28c --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/train_adapt.py @@ -0,0 +1,1343 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES=0 + +./dprnn_zipformer/train.py \ + --world-size 1 \ + --num-epochs 15 \ + --start-epoch 1 \ + --exp-dir dprnn_zipformer/exp \ + --max-duration 300 + +# For mix precision training: + +./dprnn_zipformer/train.py \ + --world-size 1 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir dprnn_zipformer/exp \ + --max-duration 550 +""" + +import argparse +import copy +import logging +import warnings +from itertools import chain +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriCssAsrDataModule +from decoder import Decoder +from dprnn import DPRNN +from einops.layers.torch import Rearrange +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import LOG_EPSILON, fix_random_seed +from model import SURT +from optim import Eden, ScaledAdam +from scaling import ScaledLinear, ScaledLSTM +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-mask-encoder-layers", + type=int, + default=4, + help="Number of layers in the DPRNN based mask encoder.", + ) + + parser.add_argument( + "--mask-encoder-dim", + type=int, + default=256, + help="Hidden dimension of the LSTM blocks in DPRNN.", + ) + + parser.add_argument( + "--mask-encoder-segment-size", + type=int, + default=32, + help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the " + "decode-chunk-length of the zipformer encoder.", + ) + + parser.add_argument( + "--chunk-width-randomization", + type=bool, + default=False, + help="Whether to randomize the chunk width in DPRNN.", + ) + + # Zipformer config is based on: + # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740 + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,2,2,2", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="768,768,768,768,768", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="256,256,256,256,256", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="192,192,192,192,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--use-joint-encoder-layer", + type=str, + default="lstm", + choices=["linear", "lstm", "none"], + help="Whether to use a joint layer to combine all branches.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--short-chunk-size", + type=int, + default=50, + help="""Chunk length of dynamic training, the chunk size would be either + max sequence length of current batch or uniformly sampled from (1, short_chunk_size). + """, + ) + + parser.add_argument( + "--num-left-chunks", + type=int, + default=4, + help="How many left context can be seen in chunks when calculating attention.", + ) + + parser.add_argument( + "--decode-chunk-len", + type=int, + default=32, + help="The chunk size for decoding (in frames before subsampling)", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=15, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="conv_lstm_transducer_stateless_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--model-init-ckpt", + type=str, + default=None, + help="""The model checkpoint to initialize the model (either full or part). + If not specified, the model is randomly initialized. + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.0004, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=1000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=2, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=1000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=5, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=100, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 10, + "reset_interval": 200, + "valid_interval": 100, + # parameters for SURT + "num_channels": 2, + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed + # parameters for Noam + "model_warm_step": 5000, # arg given to model, not for lrate + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "env_info": get_env_info(), + } + ) + + return params + + +def get_mask_encoder_model(params: AttributeDict) -> nn.Module: + mask_encoder = DPRNN( + feature_dim=params.feature_dim, + input_size=params.mask_encoder_dim, + hidden_size=params.mask_encoder_dim, + output_size=params.feature_dim * params.num_channels, + segment_size=params.mask_encoder_segment_size, + num_blocks=params.num_mask_encoder_layers, + chunk_width_randomization=params.chunk_width_randomization, + ) + return mask_encoder + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + num_left_chunks=params.num_left_chunks, + short_chunk_size=params.short_chunk_size, + decode_chunk_size=params.decode_chunk_len // 2, + ) + return encoder + + +def get_joint_encoder_layer(params: AttributeDict) -> nn.Module: + class TakeFirst(nn.Module): + def forward(self, x): + return x[0] + + if params.use_joint_encoder_layer == "linear": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + nn.Linear( + params.num_channels * encoder_dim, params.num_channels * encoder_dim + ), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "lstm": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + ScaledLSTM( + input_size=params.num_channels * encoder_dim, + hidden_size=params.num_channels * encoder_dim, + num_layers=1, + bias=True, + batch_first=True, + dropout=0.0, + bidirectional=False, + ), + TakeFirst(), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "none": + joint_layer = None + else: + raise ValueError( + f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}" + ) + return joint_layer + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_surt_model( + params: AttributeDict, +) -> nn.Module: + mask_encoder = get_mask_encoder_model(params) + encoder = get_encoder_model(params) + joint_layer = get_joint_encoder_layer(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = SURT( + mask_encoder=mask_encoder, + encoder=encoder, + joint_encoder_layer=joint_layer, + decoder=decoder, + joiner=joiner, + num_channels=params.num_channels, + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"].to(device) + feature_lens = batch["input_lens"].to(device) + + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + + # The dataloader returns text as a list of cuts, each of which is a list of channel + # text. We flatten this to a list where all channels are together, i.e., it looks like + # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2]. + text = [val for tup in zip(*batch["text"]) for val in tup] + assert len(text) == len(feature) * params.num_channels + + # Convert all channel texts to token IDs and create a ragged tensor. + y = sp.encode(text, out_type=int) + y = k2.RaggedTensor(y).to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.model_warm_step + + with torch.set_grad_enabled(is_training): + (simple_loss, pruned_loss, ctc_loss, x_masked) = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + reduction="none", + subsampling_factor=params.subsampling_factor, + ) + simple_loss_is_finite = torch.isfinite(simple_loss) + pruned_loss_is_finite = torch.isfinite(pruned_loss) + ctc_loss_is_finite = torch.isfinite(ctc_loss) + + is_finite = simple_loss_is_finite & pruned_loss_is_finite & ctc_loss_is_finite + if not torch.all(is_finite): + logging.info( + "Not all losses are finite!\n" + f"simple_losses: {simple_loss}\n" + f"pruned_losses: {pruned_loss}\n" + f"ctc_losses: {ctc_loss}\n" + ) + display_and_save_batch(batch, params=params, sp=sp) + simple_loss = simple_loss[simple_loss_is_finite] + pruned_loss = pruned_loss[pruned_loss_is_finite] + ctc_loss = ctc_loss[ctc_loss_is_finite] + + # If either all simple_loss or pruned_loss is inf or nan, + # we stop the training process by raising an exception + if ( + torch.all(~simple_loss_is_finite) + or torch.all(~pruned_loss_is_finite) + or torch.all(~ctc_loss_is_finite) + ): + raise ValueError( + "There are too many utterances in this batch " + "leading to inf or nan losses." + ) + + simple_loss_sum = simple_loss.sum() + pruned_loss_sum = pruned_loss.sum() + ctc_loss_sum = ctc_loss.sum() + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss = ( + simple_loss_scale * simple_loss_sum + + pruned_loss_scale * pruned_loss_sum + + params.ctc_loss_scale * ctc_loss_sum + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + # info["frames"] is an approximate number for two reasons: + # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2 + # (2) If some utterances in the batch lead to inf/nan loss, they + # are filtered out. + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances` # noqa + info["utterances"] = feature.size(0) + # averaged input duration in frames over utterances + info["utt_duration"] = feature_lens.sum().item() + # averaged padding proportion over utterances + info["utt_pad_proportion"] = ( + ((feature.size(1) - feature_lens) / feature.size(1)).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss_sum.detach().cpu().item() + info["pruned_loss"] = pruned_loss_sum.detach().cpu().item() + if params.ctc_loss_scale > 0.0: + info["ctc_loss"] = ctc_loss_sum.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + train_dl_warmup: + Dataloader for the training dataset with 2 speakers. This is used during the + warmup stage. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + torch.cuda.empty_cache() + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = batch["inputs"].shape[0] + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_surt_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + + if checkpoints is None and params.model_init_ckpt is not None: + logging.info( + f"Initializing model with checkpoint from {params.model_init_ckpt}" + ) + init_ckpt = torch.load(params.model_init_ckpt, map_location=device) + model.load_state_dict(init_ckpt["model"], strict=True) + + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + diagnostic = diagnostics.attach_diagnostics(model) + + libricss = LibriCssAsrDataModule(args) + + train_cuts_ihm = libricss.libricss_cuts(split="dev", type="ihm-mix") + train_cuts_sdm = libricss.libricss_cuts(split="dev", type="sdm") + train_cuts = train_cuts_ihm + train_cuts_sdm + + # This will create 2 copies of the sessions with different segmentation + train_cuts = train_cuts.trim_to_supervision_groups( + max_pause=0.1 + ) + train_cuts.trim_to_supervision_groups(max_pause=0.5) + dev_cuts = libricss.libricss_cuts(split="dev", type="sdm") + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = libricss.train_dataloaders( + train_cuts, + sampler_state_dict=sampler_state_dict, + return_sources=False, + strict=False, + ) + valid_dl = libricss.valid_dataloaders(dev_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = [sp.encode(text_ch) for text_ch in batch["text"]] + num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y] + logging.info(f"num tokens: {num_tokens}") + + +def main(): + parser = get_parser() + LibriCssAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + +if __name__ == "__main__": + main() diff --git a/egs/libricss/SURT/dprnn_zipformer/zipformer.py b/egs/libricss/SURT/dprnn_zipformer/zipformer.py new file mode 120000 index 000000000..ec183baa7 --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py \ No newline at end of file diff --git a/egs/libricss/SURT/heat.png b/egs/libricss/SURT/heat.png new file mode 100644 index 000000000..ac7ecfff4 Binary files /dev/null and b/egs/libricss/SURT/heat.png differ diff --git a/egs/libricss/SURT/local/add_source_feats.py b/egs/libricss/SURT/local/add_source_feats.py new file mode 100755 index 000000000..c9775561f --- /dev/null +++ b/egs/libricss/SURT/local/add_source_feats.py @@ -0,0 +1,85 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file adds source features as temporal arrays to the mixture manifests. +It looks for manifests in the directory data/manifests. +""" +import logging +from pathlib import Path + +import numpy as np +from lhotse import CutSet, LilcomChunkyWriter, load_manifest, load_manifest_lazy +from tqdm import tqdm + + +def add_source_feats(num_jobs=1): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + for type_affix in ["full", "ov40"]: + logging.info(f"Adding source features for {type_affix}") + mixed_name_clean = f"train_clean_{type_affix}" + mixed_name_rvb = f"train_rvb_{type_affix}" + + logging.info("Reading mixed cuts") + mixed_cuts_clean = load_manifest_lazy( + src_dir / f"cuts_{mixed_name_clean}.jsonl.gz" + ) + mixed_cuts_rvb = load_manifest_lazy(src_dir / f"cuts_{mixed_name_rvb}.jsonl.gz") + + logging.info("Reading source cuts") + source_cuts = load_manifest(src_dir / "librispeech_cuts_train_trimmed.jsonl.gz") + + logging.info("Adding source features to the mixed cuts") + with tqdm() as pbar, CutSet.open_writer( + src_dir / f"cuts_{mixed_name_clean}_sources.jsonl.gz" + ) as cut_writer_clean, CutSet.open_writer( + src_dir / f"cuts_{mixed_name_rvb}_sources.jsonl.gz" + ) as cut_writer_rvb, LilcomChunkyWriter( + output_dir / f"feats_train_{type_affix}_sources" + ) as source_feat_writer: + for cut_clean, cut_rvb in zip(mixed_cuts_clean, mixed_cuts_rvb): + assert cut_rvb.id == cut_clean.id + "_rvb" + # Create source_feats and source_feat_offsets + # (See `lhotse.datasets.K2SurtDataset` for details) + source_feats = [] + source_feat_offsets = [] + cur_offset = 0 + for sup in sorted( + cut_clean.supervisions, key=lambda s: (s.start, s.speaker) + ): + source_cut = source_cuts[sup.id] + source_feats.append(source_cut.load_features()) + source_feat_offsets.append(cur_offset) + cur_offset += source_cut.num_frames + cut_clean.source_feats = source_feat_writer.store_array( + cut_clean.id, np.concatenate(source_feats, axis=0) + ) + cut_clean.source_feat_offsets = source_feat_offsets + cut_writer_clean.write(cut_clean) + cut_rvb.source_feats = cut_clean.source_feats + cut_rvb.source_feat_offsets = cut_clean.source_feat_offsets + cut_writer_rvb.write(cut_rvb) + pbar.update(1) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + add_source_feats() diff --git a/egs/libricss/SURT/local/compute_fbank_libricss.py b/egs/libricss/SURT/local/compute_fbank_libricss.py new file mode 100755 index 000000000..afd66899c --- /dev/null +++ b/egs/libricss/SURT/local/compute_fbank_libricss.py @@ -0,0 +1,105 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the LibriCSS dataset. +It looks for manifests in the directory data/manifests. + +The generated fbank features are saved in data/fbank. +""" +import logging +from pathlib import Path + +import pyloudnorm as pyln +import torch +import torch.multiprocessing +from lhotse import LilcomChunkyWriter, load_manifest_lazy +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + + +def compute_fbank_libricss(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + logging.info("Reading manifests") + cuts_ihm_mix = load_manifest_lazy( + src_dir / "libricss-ihm-mix_segments_all.jsonl.gz" + ) + cuts_sdm = load_manifest_lazy(src_dir / "libricss-sdm_segments_all.jsonl.gz") + + for name, cuts in [("ihm-mix", cuts_ihm_mix), ("sdm", cuts_sdm)]: + dev_cuts = cuts.filter(lambda c: "session0" in c.id) + test_cuts = cuts.filter(lambda c: "session0" not in c.id) + + # If SDM cuts, apply loudness normalization + if name == "sdm": + dev_cuts = dev_cuts.normalize_loudness(target=-23.0) + test_cuts = test_cuts.normalize_loudness(target=-23.0) + + logging.info(f"Extracting fbank features for {name} dev cuts") + _ = dev_cuts.compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / f"libricss-{name}_feats_dev", + manifest_path=src_dir / f"cuts_dev_libricss-{name}.jsonl.gz", + batch_duration=500, + num_workers=2, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + logging.info(f"Extracting fbank features for {name} test cuts") + _ = test_cuts.compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / f"libricss-{name}_feats_test", + manifest_path=src_dir / f"cuts_test_libricss-{name}.jsonl.gz", + batch_duration=2000, + num_workers=4, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + compute_fbank_libricss() diff --git a/egs/libricss/SURT/local/compute_fbank_librispeech.py b/egs/libricss/SURT/local/compute_fbank_librispeech.py new file mode 100755 index 000000000..5c8aece9c --- /dev/null +++ b/egs/libricss/SURT/local/compute_fbank_librispeech.py @@ -0,0 +1,111 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the LibriSpeech dataset. +It looks for manifests in the directory data/manifests. + +The generated fbank features are saved in data/fbank. +""" + +import logging +from pathlib import Path + +import torch +from lhotse import CutSet, LilcomChunkyWriter +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.recipes.utils import read_manifests_if_cached + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + + +def compute_fbank_librispeech(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + num_mel_bins = 80 + + dataset_parts = ( + "train-clean-100", + "train-clean-360", + "train-other-500", + ) + prefix = "librispeech" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=16000), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + for partition, m in manifests.items(): + cuts_filename = f"{prefix}_cuts_{partition}.{suffix}" + if (output_dir / cuts_filename).is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + + cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1) + + cut_set = cut_set.compute_and_store_features_batch( + extractor=extractor, + storage_path=f"{output_dir}/{prefix}_feats_{partition}", + manifest_path=f"{src_dir}/{cuts_filename}", + batch_duration=4000, + num_workers=2, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + compute_fbank_librispeech() diff --git a/egs/libricss/SURT/local/compute_fbank_lsmix.py b/egs/libricss/SURT/local/compute_fbank_lsmix.py new file mode 100755 index 000000000..da42f8ba1 --- /dev/null +++ b/egs/libricss/SURT/local/compute_fbank_lsmix.py @@ -0,0 +1,188 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the synthetically mixed LibriSpeech +train and dev sets. +It looks for manifests in the directory data/manifests. + +The generated fbank features are saved in data/fbank. +""" +import logging +import random +import warnings +from pathlib import Path + +import torch +import torch.multiprocessing +from lhotse import LilcomChunkyWriter, load_manifest +from lhotse.cut import MixedCut, MixTrack, MultiCut +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.recipes.utils import read_manifests_if_cached +from lhotse.utils import fix_random_seed, uuid4 + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + + +def compute_fbank_lsmix(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + logging.info("Reading manifests") + manifests = read_manifests_if_cached( + dataset_parts=["train_clean_full", "train_clean_ov40"], + types=["cuts"], + output_dir=src_dir, + prefix="lsmix", + suffix="jsonl.gz", + lazy=True, + ) + + cs = {} + cs["clean_full"] = manifests["train_clean_full"]["cuts"] + cs["clean_ov40"] = manifests["train_clean_ov40"]["cuts"] + + # only uses RIRs and noises from REVERB challenge + real_rirs = load_manifest(src_dir / "real-rir_recordings_all.jsonl.gz").filter( + lambda r: "RVB2014" in r.id + ) + noises = load_manifest(src_dir / "iso-noise_recordings_all.jsonl.gz").filter( + lambda r: "RVB2014" in r.id + ) + + # Apply perturbation to the training cuts + logging.info("Applying perturbation to the training cuts") + cs["rvb_full"] = cs["clean_full"].map( + lambda c: augment( + c, perturb_snr=True, rirs=real_rirs, noises=noises, perturb_loudness=True + ) + ) + cs["rvb_ov40"] = cs["clean_ov40"].map( + lambda c: augment( + c, perturb_snr=True, rirs=real_rirs, noises=noises, perturb_loudness=True + ) + ) + + for type_affix in ["full", "ov40"]: + for rvb_affix in ["clean", "rvb"]: + logging.info( + f"Extracting fbank features for {type_affix} {rvb_affix} training cuts" + ) + cuts = cs[f"{rvb_affix}_{type_affix}"] + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + _ = cuts.compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir + / f"lsmix_feats_train_{rvb_affix}_{type_affix}", + manifest_path=src_dir + / f"cuts_train_{rvb_affix}_{type_affix}.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + +def augment(cut, perturb_snr=False, rirs=None, noises=None, perturb_loudness=False): + """ + Given a mixed cut, this function optionally applies the following augmentations: + - Perturbing the SNRs of the tracks (in range [-5, 5] dB) + - Reverberation using a randomly selected RIR + - Adding noise + - Perturbing the loudness (in range [-20, -25] dB) + """ + out_cut = cut.drop_features() + + # Perturb the SNRs (optional) + if perturb_snr: + snrs = [random.uniform(-5, 5) for _ in range(len(cut.tracks))] + for i, (track, snr) in enumerate(zip(out_cut.tracks, snrs)): + if i == 0: + # Skip the first track since it is the reference + continue + track.snr = snr + + # Reverberate the cut (optional) + if rirs is not None: + # Select an RIR at random + rir = random.choice(rirs) + # Select a channel at random + rir_channel = random.choice(list(range(rir.num_channels))) + # Reverberate the cut + out_cut = out_cut.reverb_rir(rir_recording=rir, rir_channels=[rir_channel]) + + # Add noise (optional) + if noises is not None: + # Select a noise recording at random + noise = random.choice(noises).to_cut() + if isinstance(noise, MultiCut): + noise = noise.to_mono()[0] + # Select an SNR at random + snr = random.uniform(10, 30) + # Repeat the noise to match the duration of the cut + noise = repeat_cut(noise, out_cut.duration) + out_cut = MixedCut( + id=out_cut.id, + tracks=[ + MixTrack(cut=out_cut, type="MixedCut"), + MixTrack(cut=noise, type="DataCut", snr=snr), + ], + ) + + # Perturb the loudness (optional) + if perturb_loudness: + target_loudness = random.uniform(-20, -25) + out_cut = out_cut.normalize_loudness(target_loudness, mix_first=True) + return out_cut + + +def repeat_cut(cut, duration): + while cut.duration < duration: + cut = cut.mix(cut, offset_other_by=cut.duration) + return cut.truncate(duration=duration) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + fix_random_seed(42) + compute_fbank_lsmix() diff --git a/egs/libricss/SURT/local/compute_fbank_musan.py b/egs/libricss/SURT/local/compute_fbank_musan.py new file mode 100755 index 000000000..1fcf951f9 --- /dev/null +++ b/egs/libricss/SURT/local/compute_fbank_musan.py @@ -0,0 +1,114 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the musan dataset. +It looks for manifests in the directory data/manifests. + +The generated fbank features are saved in data/fbank. +""" + +import logging +from pathlib import Path + +import torch +from lhotse import CutSet, LilcomChunkyWriter, combine +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.recipes.utils import read_manifests_if_cached + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + + +def compute_fbank_musan(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + dataset_parts = ( + "music", + "speech", + "noise", + ) + prefix = "musan" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + musan_cuts_path = src_dir / "musan_cuts.jsonl.gz" + + if musan_cuts_path.is_file(): + logging.info(f"{musan_cuts_path} already exists - skipping") + return + + logging.info("Extracting features for Musan") + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + # create chunks of Musan with duration 5 - 10 seconds + _ = ( + CutSet.from_manifests( + recordings=combine(part["recordings"] for part in manifests.values()) + ) + .cut_into_windows(10.0) + .filter(lambda c: c.duration > 5) + .compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / "musan_feats", + manifest_path=musan_cuts_path, + batch_duration=500, + num_workers=4, + storage_type=LilcomChunkyWriter, + ) + ) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + compute_fbank_musan() diff --git a/egs/libricss/SURT/prepare.sh b/egs/libricss/SURT/prepare.sh new file mode 100755 index 000000000..028240e44 --- /dev/null +++ b/egs/libricss/SURT/prepare.sh @@ -0,0 +1,204 @@ +#!/usr/bin/env bash + +set -eou pipefail + +stage=-1 +stop_stage=100 + +# We assume dl_dir (download dir) contains the following +# directories and files. If not, they will be downloaded +# by this script automatically. +# +# - $dl_dir/librispeech +# You can find audio and transcripts for LibriSpeech in this path. +# +# - $dl_dir/libricss +# You can find audio and transcripts for LibriCSS in this path. +# +# - $dl_dir/musan +# This directory contains the following directories downloaded from +# http://www.openslr.org/17/ +# +# - music +# - noise +# - speech +# +# - $dl_dir/rirs_noises +# This directory contains the RIRS_NOISES corpus downloaded from https://openslr.org/28/. +# +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data +vocab_size=500 + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +log "dl_dir: $dl_dir" + +if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then + log "Stage 0: Download data" + + # If you have pre-downloaded it to /path/to/librispeech, + # you can create a symlink + # + # ln -sfv /path/to/librispeech $dl_dir/librispeech + # + if [ ! -d $dl_dir/librispeech ]; then + lhotse download librispeech $dl_dir/librispeech + fi + + # If you have pre-downloaded it to /path/to/libricss, + # you can create a symlink + # + # ln -sfv /path/to/libricss $dl_dir/libricss + # + if [ ! -d $dl_dir/libricss ]; then + lhotse download libricss $dl_dir/libricss + fi + + # If you have pre-downloaded it to /path/to/musan, + # you can create a symlink + # + # ln -sfv /path/to/musan $dl_dir/ + # + if [ ! -d $dl_dir/musan ]; then + lhotse download musan $dl_dir + fi + + # If you have pre-downloaded it to /path/to/rirs_noises, + # you can create a symlink + # + # ln -sfv /path/to/rirs_noises $dl_dir/ + # + if [ ! -d $dl_dir/rirs_noises ]; then + lhotse download rirs_noises $dl_dir + fi +fi + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Prepare LibriSpeech manifests" + # We assume that you have downloaded the LibriSpeech corpus + # to $dl_dir/librispeech. We perform text normalization for the transcripts. + # NOTE: Alignments are required for this recipe. + mkdir -p data/manifests + lhotse prepare librispeech -p train-clean-100 -p train-clean-360 -p train-other-500 -p dev-clean \ + -j 4 --alignments-dir $dl_dir/libri_alignments/LibriSpeech $dl_dir/librispeech data/manifests/ +fi + +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Prepare LibriCSS manifests" + # We assume that you have downloaded the LibriCSS corpus + # to $dl_dir/libricss. We perform text normalization for the transcripts. + mkdir -p data/manifests + for mic in sdm ihm-mix; do + lhotse prepare libricss --type $mic --segmented $dl_dir/libricss data/manifests/ + done +fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Prepare musan manifest and RIRs" + # We assume that you have downloaded the musan corpus + # to $dl_dir/musan + mkdir -p data/manifests + lhotse prepare musan $dl_dir/musan data/manifests + + # We assume that you have downloaded the RIRS_NOISES corpus + # to $dl_dir/rirs_noises + lhotse prepare rir-noise -p real_rir -p iso_noise $dl_dir/rirs_noises data/manifests +fi + +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + log "Stage 4: Extract features for LibriSpeech, trim to alignments, and shuffle the cuts" + python local/compute_fbank_librispeech.py + lhotse combine data/manifests/librispeech_cuts_train* - |\ + lhotse cut trim-to-alignments --type word --max-pause 0.2 - - |\ + shuf | gzip -c > data/manifests/librispeech_cuts_train_trimmed.jsonl.gz +fi + +if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then + log "Stage 5: Create simulated mixtures from LibriSpeech (train and dev). This may take a while." + # We create a high overlap set which will be used during the model warmup phase, and a + # full training set that will be used for the subsequent training. + + gunzip -c data/manifests/libricss-sdm_supervisions_all.jsonl.gz |\ + grep -v "0L" | grep -v "OV10" |\ + gzip -c > data/manifests/libricss-sdm_supervisions_all_v1.jsonl.gz + + gunzip -c data/manifests/libricss-sdm_supervisions_all.jsonl.gz |\ + grep "OV40" |\ + gzip -c > data/manifests/libricss-sdm_supervisions_ov40.jsonl.gz + + # Warmup mixtures (100k) based on high overlap (OV40) + log "Generating 100k anechoic train mixtures for warmup" + lhotse workflows simulate-meetings \ + --method conversational \ + --fit-to-supervisions data/manifests/libricss-sdm_supervisions_ov40.jsonl.gz \ + --num-meetings 100000 \ + --num-speakers-per-meeting 2,3 \ + --max-duration-per-speaker 15.0 \ + --max-utterances-per-speaker 3 \ + --seed 1234 \ + --num-jobs 4 \ + data/manifests/librispeech_cuts_train_trimmed.jsonl.gz \ + data/manifests/lsmix_cuts_train_clean_ov40.jsonl.gz + + # Full training set (2,3 speakers) anechoic + log "Generating anechoic ${part} set (full)" + lhotse workflows simulate-meetings \ + --method conversational \ + --fit-to-supervisions data/manifests/libricss-sdm_supervisions_all_v1.jsonl.gz \ + --num-repeats 1 \ + --num-speakers-per-meeting 2,3 \ + --max-duration-per-speaker 15.0 \ + --max-utterances-per-speaker 3 \ + --seed 1234 \ + --num-jobs 4 \ + data/manifests/librispeech_cuts_train_trimmed.jsonl.gz \ + data/manifests/lsmix_cuts_train_clean_full.jsonl.gz +fi + +if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then + log "Stage 6: Compute fbank features for musan" + mkdir -p data/fbank + python local/compute_fbank_musan.py +fi + +if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then + log "Stage 7: Compute fbank features for simulated Libri-mix" + mkdir -p data/fbank + python local/compute_fbank_lsmix.py +fi + +if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then + log "Stage 8: Add source feats to mixtures (useful for auxiliary tasks)" + python local/add_source_feats.py + + log "Combining lsmix-clean and lsmix-rvb" + for type in full ov40; do + cat <(gunzip -c data/manifests/cuts_train_clean_${type}_sources.jsonl.gz) \ + <(gunzip -c data/manifests/cuts_train_rvb_${type}_sources.jsonl.gz) |\ + shuf | gzip -c > data/manifests/cuts_train_comb_${type}_sources.jsonl.gz + done +fi + +if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then + log "Stage 9: Compute fbank features for LibriCSS" + mkdir -p data/fbank + python local/compute_fbank_libricss.py +fi + +if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then + log "Stage 10: Download LibriSpeech BPE model from HuggingFace." + mkdir -p data/lang_bpe_500 + pushd data/lang_bpe_500 + wget https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/resolve/main/data/lang_bpe_500/bpe.model + popd +fi diff --git a/egs/libricss/SURT/shared b/egs/libricss/SURT/shared new file mode 120000 index 000000000..4cbd91a7e --- /dev/null +++ b/egs/libricss/SURT/shared @@ -0,0 +1 @@ +../../../icefall/shared \ No newline at end of file diff --git a/egs/libricss/SURT/surt.png b/egs/libricss/SURT/surt.png new file mode 100644 index 000000000..fcc8119d4 Binary files /dev/null and b/egs/libricss/SURT/surt.png differ diff --git a/egs/librispeech/ASR/conformer_ctc2/train.py b/egs/librispeech/ASR/conformer_ctc2/train.py index 3366af13e..c4a13b101 100755 --- a/egs/librispeech/ASR/conformer_ctc2/train.py +++ b/egs/librispeech/ASR/conformer_ctc2/train.py @@ -675,7 +675,6 @@ def train_one_epoch( for batch_idx, batch in enumerate(train_dl): params.batch_idx_train += 1 batch_size = len(batch["supervisions"]["text"]) - batch_name = batch["supervisions"]["uttid"] with torch.cuda.amp.autocast(enabled=params.use_fp16): loss, loss_info = compute_loss( @@ -698,10 +697,7 @@ def train_one_epoch( scaler.scale(loss).backward() except RuntimeError as e: if "CUDA out of memory" in str(e): - logging.error( - f"failing batch size:{batch_size} " - f"failing batch names {batch_name}" - ) + logging.error(f"failing batch size:{batch_size} ") raise scheduler.step_batch(params.batch_idx_train) @@ -756,10 +752,7 @@ def train_one_epoch( if loss_info["ctc_loss"] == float("inf") or loss_info["att_loss"] == float( "inf" ): - logging.error( - "Your loss contains inf, something goes wrong" - f"failing batch names {batch_name}" - ) + logging.error("Your loss contains inf, something goes wrong") if tb_writer is not None: tb_writer.add_scalar( "train/learning_rate", cur_lr, params.batch_idx_train diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py index 1bbad6946..fd59d4b7f 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py @@ -22,6 +22,7 @@ from typing import Dict, List, Optional, Tuple, Union import k2 import sentencepiece as spm import torch +from torch import nn from icefall import ContextGraph, ContextState, NgramLm, NgramLmStateCost from icefall.decode import Nbest, one_best_decoding @@ -35,7 +36,6 @@ from icefall.utils import ( get_texts, get_texts_with_timestamp, ) -from torch import nn def fast_beam_search_one_best( @@ -47,9 +47,10 @@ def fast_beam_search_one_best( max_states: int, max_contexts: int, temperature: float = 1.0, - subtract_ilme: bool = False, - ilme_scale: float = 0.1, + ilme_scale: float = 0.0, + blank_penalty: float = 0.0, return_timestamps: bool = False, + allow_partial: bool = False, ) -> Union[List[List[int]], DecodingResults]: """It limits the maximum number of symbols per frame to 1. @@ -90,8 +91,9 @@ def fast_beam_search_one_best( max_states=max_states, max_contexts=max_contexts, temperature=temperature, - subtract_ilme=subtract_ilme, ilme_scale=ilme_scale, + allow_partial=allow_partial, + blank_penalty=blank_penalty, ) best_path = one_best_decoding(lattice) @@ -114,7 +116,10 @@ def fast_beam_search_nbest_LG( nbest_scale: float = 0.5, use_double_scores: bool = True, temperature: float = 1.0, + blank_penalty: float = 0.0, + ilme_scale: float = 0.0, return_timestamps: bool = False, + allow_partial: bool = False, ) -> Union[List[List[int]], DecodingResults]: """It limits the maximum number of symbols per frame to 1. @@ -168,6 +173,9 @@ def fast_beam_search_nbest_LG( max_states=max_states, max_contexts=max_contexts, temperature=temperature, + allow_partial=allow_partial, + blank_penalty=blank_penalty, + ilme_scale=ilme_scale, ) nbest = Nbest.from_lattice( @@ -240,7 +248,9 @@ def fast_beam_search_nbest( nbest_scale: float = 0.5, use_double_scores: bool = True, temperature: float = 1.0, + blank_penalty: float = 0.0, return_timestamps: bool = False, + allow_partial: bool = False, ) -> Union[List[List[int]], DecodingResults]: """It limits the maximum number of symbols per frame to 1. @@ -293,7 +303,9 @@ def fast_beam_search_nbest( beam=beam, max_states=max_states, max_contexts=max_contexts, + blank_penalty=blank_penalty, temperature=temperature, + allow_partial=allow_partial, ) nbest = Nbest.from_lattice( @@ -331,7 +343,9 @@ def fast_beam_search_nbest_oracle( use_double_scores: bool = True, nbest_scale: float = 0.5, temperature: float = 1.0, + blank_penalty: float = 0.0, return_timestamps: bool = False, + allow_partial: bool = False, ) -> Union[List[List[int]], DecodingResults]: """It limits the maximum number of symbols per frame to 1. @@ -389,6 +403,8 @@ def fast_beam_search_nbest_oracle( max_states=max_states, max_contexts=max_contexts, temperature=temperature, + allow_partial=allow_partial, + blank_penalty=blank_penalty, ) nbest = Nbest.from_lattice( @@ -434,6 +450,8 @@ def fast_beam_search( temperature: float = 1.0, subtract_ilme: bool = False, ilme_scale: float = 0.1, + allow_partial: bool = False, + blank_penalty: float = 0.0, ) -> k2.Fsa: """It limits the maximum number of symbols per frame to 1. @@ -503,8 +521,13 @@ def fast_beam_search( project_input=False, ) logits = logits.squeeze(1).squeeze(1) + + if blank_penalty != 0: + logits[:, 0] -= blank_penalty + log_probs = (logits / temperature).log_softmax(dim=-1) - if subtract_ilme: + + if ilme_scale != 0: ilme_logits = model.joiner( torch.zeros_like( current_encoder_out, device=current_encoder_out.device @@ -513,11 +536,16 @@ def fast_beam_search( project_input=False, ) ilme_logits = ilme_logits.squeeze(1).squeeze(1) + if blank_penalty != 0: + ilme_logits[:, 0] -= blank_penalty ilme_log_probs = (ilme_logits / temperature).log_softmax(dim=-1) log_probs -= ilme_scale * ilme_log_probs + decoding_streams.advance(log_probs) decoding_streams.terminate_and_flush_to_streams() - lattice = decoding_streams.format_output(encoder_out_lens.tolist()) + lattice = decoding_streams.format_output( + encoder_out_lens.tolist(), allow_partial=allow_partial + ) return lattice @@ -526,6 +554,7 @@ def greedy_search( model: nn.Module, encoder_out: torch.Tensor, max_sym_per_frame: int, + blank_penalty: float = 0.0, return_timestamps: bool = False, ) -> Union[List[int], DecodingResults]: """Greedy search for a single utterance. @@ -595,6 +624,9 @@ def greedy_search( ) # logits is (1, 1, 1, vocab_size) + if blank_penalty != 0: + logits[:, :, :, 0] -= blank_penalty + y = logits.argmax().item() if y not in (blank_id, unk_id): hyp.append(y) @@ -626,6 +658,7 @@ def greedy_search_batch( model: nn.Module, encoder_out: torch.Tensor, encoder_out_lens: torch.Tensor, + blank_penalty: float = 0, return_timestamps: bool = False, ) -> Union[List[List[int]], DecodingResults]: """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. @@ -703,6 +736,10 @@ def greedy_search_batch( logits = logits.squeeze(1).squeeze(1) # (batch_size, vocab_size) assert logits.ndim == 2, logits.shape + + if blank_penalty != 0: + logits[:, 0] -= blank_penalty + y = logits.argmax(dim=1).tolist() emitted = False for i, v in enumerate(y): @@ -923,6 +960,7 @@ def modified_beam_search( context_graph: Optional[ContextGraph] = None, beam: int = 4, temperature: float = 1.0, + blank_penalty: float = 0.0, return_timestamps: bool = False, ) -> Union[List[List[int]], DecodingResults]: """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded. @@ -1028,6 +1066,9 @@ def modified_beam_search( logits = logits.squeeze(1).squeeze(1) # (num_hyps, vocab_size) + if blank_penalty != 0: + logits[:, 0] -= blank_penalty + log_probs = (logits / temperature).log_softmax(dim=-1) # (num_hyps, vocab_size) log_probs.add_(ys_log_probs) @@ -1662,6 +1703,7 @@ def beam_search( encoder_out: torch.Tensor, beam: int = 4, temperature: float = 1.0, + blank_penalty: float = 0.0, return_timestamps: bool = False, ) -> Union[List[int], DecodingResults]: """ @@ -1758,6 +1800,9 @@ def beam_search( project_input=False, ) + if blank_penalty != 0: + logits[:, :, :, 0] -= blank_penalty + # TODO(fangjun): Scale the blank posterior log_prob = (logits / temperature).log_softmax(dim=-1) # log_prob is (1, 1, 1, vocab_size) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py index 9bac46004..bcd419fb7 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py @@ -849,6 +849,8 @@ class RelPositionalEncoding(torch.nn.Module): torch.Tensor: Encoded tensor (batch, 2*time-1, `*`). """ + if isinstance(left_context, torch.Tensor): + left_context = left_context.item() self.extend_pe(x, left_context) x_size_1 = x.size(1) + left_context pos_emb = self.pe[ diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/test_onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless3/test_onnx.py index 598fcf344..810da8da6 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless3/test_onnx.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/test_onnx.py @@ -113,7 +113,7 @@ def test_rel_pos(): torch.onnx.export( encoder_pos, - x, + (x, torch.zeros(1, dtype=torch.int64)), filename, verbose=False, opset_version=opset_version, @@ -139,7 +139,9 @@ def test_rel_pos(): assert input_nodes[0].name == "x" assert input_nodes[0].shape == ["N", "T", num_features] - inputs = {input_nodes[0].name: x.numpy()} + inputs = { + input_nodes[0].name: x.numpy(), + } onnx_y, onnx_pos_emb = session.run(["y", "pos_emb"], inputs) onnx_y = torch.from_numpy(onnx_y) onnx_pos_emb = torch.from_numpy(onnx_pos_emb) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/test_onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless7/test_onnx.py index 2440d267c..1e9b67226 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/test_onnx.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/test_onnx.py @@ -265,7 +265,7 @@ def test_zipformer_encoder(): torch.onnx.export( encoder, - (x), + (x, torch.ones(1, dtype=torch.float32)), filename, verbose=False, opset_version=opset_version, @@ -289,6 +289,7 @@ def test_zipformer_encoder(): input_nodes = session.get_inputs() inputs = { input_nodes[0].name: x.numpy(), + input_nodes[1].name: torch.ones(1, dtype=torch.float32).numpy(), } onnx_y = session.run(["y"], inputs)[0] onnx_y = torch.from_numpy(onnx_y) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py index 0841f7cf1..c44cb1eaf 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py @@ -81,20 +81,20 @@ class FrameReducer(nn.Module): fake_limit_indexes = torch.topk( ctc_output[:, :, blank_id], max_limit_len ).indices - T = ( + T_arange = ( torch.arange(max_limit_len) .expand_as( fake_limit_indexes, ) .to(device=x.device) ) - T = torch.remainder(T, limit_lens.unsqueeze(1)) - limit_indexes = torch.gather(fake_limit_indexes, 1, T) + T_arange = torch.remainder(T_arange, limit_lens.unsqueeze(1)) + limit_indexes = torch.gather(fake_limit_indexes, 1, T_arange) limit_mask = torch.full_like( non_blank_mask, - False, + 0, device=x.device, - ).scatter_(1, limit_indexes, True) + ).scatter_(1, limit_indexes, 1) non_blank_mask = non_blank_mask | ~limit_mask @@ -108,9 +108,9 @@ class FrameReducer(nn.Module): ) - out_lens ) - max_pad_len = pad_lens_list.max() + max_pad_len = int(pad_lens_list.max()) - out = F.pad(x, (0, 0, 0, max_pad_len)) + out = F.pad(x, [0, 0, 0, max_pad_len]) valid_pad_mask = ~make_pad_mask(pad_lens_list) total_valid_mask = torch.concat([non_blank_mask, valid_pad_mask], dim=1) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export.py index 5735ee692..c191b5bcc 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export.py @@ -856,6 +856,10 @@ def main(): # Otherwise, one of its arguments is a ragged tensor and is not # torch scriptabe. model.__class__.forward = torch.jit.ignore(model.__class__.forward) + model.encoder.__class__.non_streaming_forward = model.encoder.__class__.forward + model.encoder.__class__.non_streaming_forward = torch.jit.export( + model.encoder.__class__.non_streaming_forward + ) model.encoder.__class__.forward = model.encoder.__class__.streaming_forward logging.info("Using torch.jit.script") model = torch.jit.script(model) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py index 4fd5e1820..c8301b2da 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py @@ -252,7 +252,7 @@ def main(): feature_lengths = torch.tensor(feature_lengths, device=device) - encoder_out, encoder_out_lens = model.encoder( + encoder_out, encoder_out_lens = model.encoder.non_streaming_forward( x=features, x_lens=feature_lengths, ) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py index b76272e66..a0f54b6e1 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py @@ -22,7 +22,7 @@ Usage: --avg 15 \ --decode-chunk-len 32 \ --exp-dir ./pruned_transducer_stateless7_streaming/exp \ - --decoding_method greedy_search \ + --decoding-method greedy_search \ --num-decode-streams 2000 """ diff --git a/egs/librispeech/ASR/zipformer/.gitignore b/egs/librispeech/ASR/zipformer/.gitignore new file mode 100644 index 000000000..e47ac1582 --- /dev/null +++ b/egs/librispeech/ASR/zipformer/.gitignore @@ -0,0 +1 @@ +swoosh.pdf diff --git a/egs/librispeech/ASR/zipformer/export-onnx-streaming.py b/egs/librispeech/ASR/zipformer/export-onnx-streaming.py index 8cec09869..3eb06f68c 100755 --- a/egs/librispeech/ASR/zipformer/export-onnx-streaming.py +++ b/egs/librispeech/ASR/zipformer/export-onnx-streaming.py @@ -1,6 +1,6 @@ #!/usr/bin/env python3 # -# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang) +# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Wei Kang) # Copyright 2023 Danqing Fu (danqing.fu@gmail.com) """ @@ -19,7 +19,7 @@ GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url repo=$(basename $repo_url) pushd $repo -git lfs pull --include "data/lang_bpe_500/bpe.model" +git lfs pull --include "data/lang_bpe_500/tokens.txt" git lfs pull --include "exp/pretrained.pt" cd exp @@ -29,7 +29,7 @@ popd 2. Export the model to ONNX ./zipformer/export-onnx-streaming.py \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ --use-averaged-model 0 \ --epoch 99 \ --avg 1 \ @@ -57,9 +57,9 @@ whose value is "64,128,256,-1". It will generate the following 3 files inside $repo/exp: - - encoder-epoch-99-avg-1.onnx - - decoder-epoch-99-avg-1.onnx - - joiner-epoch-99-avg-1.onnx + - encoder-epoch-99-avg-1-chunk-16-left-64.onnx + - decoder-epoch-99-avg-1-chunk-16-left-64.onnx + - joiner-epoch-99-avg-1-chunk-16-left-64.onnx See ./onnx_pretrained-streaming.py for how to use the exported ONNX models. """ @@ -69,14 +69,15 @@ import logging from pathlib import Path from typing import Dict, List, Tuple +import k2 import onnx -import sentencepiece as spm import torch import torch.nn as nn from decoder import Decoder +from export import num_tokens from onnxruntime.quantization import QuantType, quantize_dynamic from scaling_converter import convert_scaled_to_non_scaled -from train import add_model_arguments, get_params, get_model +from train import add_model_arguments, get_model, get_params from zipformer import Zipformer2 from icefall.checkpoint import ( @@ -85,7 +86,7 @@ from icefall.checkpoint import ( find_checkpoints, load_checkpoint, ) -from icefall.utils import str2bool, make_pad_mask +from icefall.utils import str2bool def get_parser(): @@ -142,10 +143,10 @@ def get_parser(): ) parser.add_argument( - "--bpe-model", + "--tokens", type=str, - default="data/lang_bpe_500/bpe.model", - help="Path to the BPE model", + default="data/lang_bpe_500/tokens.txt", + help="Path to the tokens.txt", ) parser.add_argument( @@ -217,7 +218,7 @@ class OnnxEncoder(nn.Module): ) assert x.size(1) == self.chunk_size, (x.size(1), self.chunk_size) - src_key_padding_mask = make_pad_mask(x_lens) + src_key_padding_mask = torch.zeros(N, self.chunk_size, dtype=torch.bool) # processed_mask is used to mask out initial states processed_mask = torch.arange(left_context_len, device=x.device).expand( @@ -271,6 +272,7 @@ class OnnxEncoder(nn.Module): states = self.encoder.get_init_states(batch_size, device) embed_states = self.encoder_embed.get_init_states(batch_size, device) + states.append(embed_states) processed_lens = torch.zeros(batch_size, dtype=torch.int64, device=device) @@ -585,12 +587,9 @@ def main(): logging.info(f"device: {device}") - sp = spm.SentencePieceProcessor() - sp.load(params.bpe_model) - - # is defined in local/train_bpe_model.py - params.blank_id = sp.piece_to_id("") - params.vocab_size = sp.get_piece_size() + token_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 logging.info(params) @@ -709,6 +708,8 @@ def main(): suffix = f"epoch-{params.epoch}" suffix += f"-avg-{params.avg}" + suffix += f"-chunk-{params.chunk_size}" + suffix += f"-left-{params.left_context_frames}" opset_version = 13 @@ -756,7 +757,7 @@ def main(): quantize_dynamic( model_input=decoder_filename, model_output=decoder_filename_int8, - op_types_to_quantize=["MatMul"], + op_types_to_quantize=["MatMul", "Gather"], weight_type=QuantType.QInt8, ) diff --git a/egs/librispeech/ASR/zipformer/export-onnx.py b/egs/librispeech/ASR/zipformer/export-onnx.py index f5b01ce71..724fdd2a6 100755 --- a/egs/librispeech/ASR/zipformer/export-onnx.py +++ b/egs/librispeech/ASR/zipformer/export-onnx.py @@ -1,6 +1,6 @@ #!/usr/bin/env python3 # -# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang) +# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Wei Kang) # Copyright 2023 Danqing Fu (danqing.fu@gmail.com) """ @@ -19,7 +19,7 @@ GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url repo=$(basename $repo_url) pushd $repo -git lfs pull --include "data/lang_bpe_500/bpe.model" +git lfs pull --include "data/lang_bpe_500/tokens.txt" git lfs pull --include "exp/pretrained.pt" cd exp @@ -29,12 +29,11 @@ popd 2. Export the model to ONNX ./zipformer/export-onnx.py \ - --bpe-model $repo/data/lang_bpe_500/bpe.model \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ --use-averaged-model 0 \ --epoch 99 \ --avg 1 \ --exp-dir $repo/exp \ - \ --num-encoder-layers "2,2,3,4,3,2" \ --downsampling-factor "1,2,4,8,4,2" \ --feedforward-dim "512,768,1024,1536,1024,768" \ @@ -67,14 +66,15 @@ import logging from pathlib import Path from typing import Dict, Tuple +import k2 import onnx -import sentencepiece as spm import torch import torch.nn as nn from decoder import Decoder +from export import num_tokens from onnxruntime.quantization import QuantType, quantize_dynamic from scaling_converter import convert_scaled_to_non_scaled -from train import add_model_arguments, get_params, get_model +from train import add_model_arguments, get_model, get_params from zipformer import Zipformer2 from icefall.checkpoint import ( @@ -83,7 +83,7 @@ from icefall.checkpoint import ( find_checkpoints, load_checkpoint, ) -from icefall.utils import str2bool, make_pad_mask +from icefall.utils import make_pad_mask, str2bool def get_parser(): @@ -140,10 +140,10 @@ def get_parser(): ) parser.add_argument( - "--bpe-model", + "--tokens", type=str, - default="data/lang_bpe_500/bpe.model", - help="Path to the BPE model", + default="data/lang_bpe_500/tokens.txt", + help="Path to the tokens.txt", ) parser.add_argument( @@ -434,12 +434,9 @@ def main(): logging.info(f"device: {device}") - sp = spm.SentencePieceProcessor() - sp.load(params.bpe_model) - - # is defined in local/train_bpe_model.py - params.blank_id = sp.piece_to_id("") - params.vocab_size = sp.get_piece_size() + token_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 logging.info(params) @@ -605,7 +602,7 @@ def main(): quantize_dynamic( model_input=decoder_filename, model_output=decoder_filename_int8, - op_types_to_quantize=["MatMul"], + op_types_to_quantize=["MatMul", "Gather"], weight_type=QuantType.QInt8, ) diff --git a/egs/librispeech/ASR/zipformer/export.py b/egs/librispeech/ASR/zipformer/export.py index a100cbb8d..4a48d5bad 100755 --- a/egs/librispeech/ASR/zipformer/export.py +++ b/egs/librispeech/ASR/zipformer/export.py @@ -1,6 +1,8 @@ #!/usr/bin/env python3 # -# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, Zengwei Yao) +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Wei Kang) # # See ../../../../LICENSE for clarification regarding multiple authors # @@ -22,13 +24,16 @@ Usage: +Note: This is a example for librispeech dataset, if you are using different +dataset, you should change the argument values according to your dataset. + (1) Export to torchscript model using torch.jit.script() - For non-streaming model: ./zipformer/export.py \ --exp-dir ./zipformer/exp \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 \ --jit 1 @@ -48,7 +53,7 @@ for how to use the exported models outside of icefall. --causal 1 \ --chunk-size 16 \ --left-context-frames 128 \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 \ --jit 1 @@ -67,7 +72,7 @@ for how to use the exported models outside of icefall. ./zipformer/export.py \ --exp-dir ./zipformer/exp \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 @@ -76,7 +81,7 @@ for how to use the exported models outside of icefall. ./zipformer/export.py \ --exp-dir ./zipformer/exp \ --causal 1 \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 @@ -155,13 +160,15 @@ with the following commands: import argparse import logging +import re from pathlib import Path from typing import List, Tuple -import sentencepiece as spm +import k2 import torch +from scaling_converter import convert_scaled_to_non_scaled from torch import Tensor, nn -from train import add_model_arguments, get_params, get_model +from train import add_model_arguments, get_model, get_params from icefall.checkpoint import ( average_checkpoints, @@ -170,7 +177,26 @@ from icefall.checkpoint import ( load_checkpoint, ) from icefall.utils import make_pad_mask, str2bool -from scaling_converter import convert_scaled_to_non_scaled + + +def num_tokens( + token_table: k2.SymbolTable, disambig_pattern: str = re.compile(r"^#\d+$") +) -> int: + """Return the number of tokens excluding those from + disambiguation symbols. + + Caution: + 0 is not a token ID so it is excluded from the return value. + """ + symbols = token_table.symbols + ans = [] + for s in symbols: + if not disambig_pattern.match(s): + ans.append(token_table[s]) + num_tokens = len(ans) + if 0 in ans: + num_tokens -= 1 + return num_tokens def get_parser(): @@ -227,10 +253,10 @@ def get_parser(): ) parser.add_argument( - "--bpe-model", + "--tokens", type=str, - default="data/lang_bpe_500/bpe.model", - help="Path to the BPE model", + default="data/lang_bpe_500/tokens.txt", + help="Path to the tokens.txt", ) parser.add_argument( @@ -238,7 +264,7 @@ def get_parser(): type=str2bool, default=False, help="""True to save a model after applying torch.jit.script. - It will generate a file named cpu_jit.pt. + It will generate a file named jit_script.pt. Check ./jit_pretrained.py for how to use it. """, ) @@ -257,6 +283,7 @@ def get_parser(): class EncoderModel(nn.Module): """A wrapper for encoder and encoder_embed""" + def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None: super().__init__() self.encoder = encoder @@ -275,9 +302,7 @@ class EncoderModel(nn.Module): src_key_padding_mask = make_pad_mask(x_lens) x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) - encoder_out, encoder_out_lens = self.encoder( - x, x_lens, src_key_padding_mask - ) + encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask) encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) return encoder_out, encoder_out_lens @@ -398,12 +423,9 @@ def main(): logging.info(f"device: {device}") - sp = spm.SentencePieceProcessor() - sp.load(params.bpe_model) - - # is defined in local/train_bpe_model.py - params.blank_id = sp.piece_to_id("") - params.vocab_size = sp.get_piece_size() + token_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 logging.info(params) diff --git a/egs/librispeech/ASR/zipformer/generate_averaged_model.py b/egs/librispeech/ASR/zipformer/generate_averaged_model.py index e0c7b52cb..68111fad7 100755 --- a/egs/librispeech/ASR/zipformer/generate_averaged_model.py +++ b/egs/librispeech/ASR/zipformer/generate_averaged_model.py @@ -40,16 +40,11 @@ You can later load it by `torch.load("iter-22000-avg-5.pt")`. import argparse from pathlib import Path -import sentencepiece as spm +import k2 import torch -from asr_datamodule import LibriSpeechAsrDataModule +from train import add_model_arguments, get_model, get_params -from train import add_model_arguments, get_params, get_model - -from icefall.checkpoint import ( - average_checkpoints_with_averaged_model, - find_checkpoints, -) +from icefall.checkpoint import average_checkpoints_with_averaged_model, find_checkpoints def get_parser(): @@ -93,10 +88,10 @@ def get_parser(): ) parser.add_argument( - "--bpe-model", + "--tokens", type=str, - default="data/lang_bpe_500/bpe.model", - help="Path to the BPE model", + default="data/lang_bpe_500/tokens.txt", + help="Path to the tokens.txt", ) parser.add_argument( @@ -114,7 +109,6 @@ def get_parser(): @torch.no_grad() def main(): parser = get_parser() - LibriSpeechAsrDataModule.add_arguments(parser) args = parser.parse_args() args.exp_dir = Path(args.exp_dir) @@ -131,13 +125,10 @@ def main(): device = torch.device("cpu") print(f"Device: {device}") - sp = spm.SentencePieceProcessor() - sp.load(params.bpe_model) - - # is defined in local/train_bpe_model.py - params.blank_id = sp.piece_to_id("") - params.unk_id = sp.piece_to_id("") - params.vocab_size = sp.get_piece_size() + symbol_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = symbol_table[""] + params.unk_id = symbol_table[""] + params.vocab_size = len(symbol_table) print("About to create model") model = get_model(params) diff --git a/egs/librispeech/ASR/zipformer/jit_pretrained.py b/egs/librispeech/ASR/zipformer/jit_pretrained.py index 87cd5102c..a41fbc1c9 100755 --- a/egs/librispeech/ASR/zipformer/jit_pretrained.py +++ b/egs/librispeech/ASR/zipformer/jit_pretrained.py @@ -21,7 +21,7 @@ You can use the following command to get the exported models: ./zipformer/export.py \ --exp-dir ./zipformer/exp \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 \ --jit 1 @@ -30,7 +30,7 @@ Usage of this script: ./zipformer/jit_pretrained.py \ --nn-model-filename ./zipformer/exp/cpu_jit.pt \ - --bpe-model ./data/lang_bpe_500/bpe.model \ + --tokens ./data/lang_bpe_500/tokens.txt \ /path/to/foo.wav \ /path/to/bar.wav """ @@ -40,8 +40,8 @@ import logging import math from typing import List +import k2 import kaldifeat -import sentencepiece as spm import torch import torchaudio from torch.nn.utils.rnn import pad_sequence @@ -60,9 +60,9 @@ def get_parser(): ) parser.add_argument( - "--bpe-model", + "--tokens", type=str, - help="""Path to bpe.model.""", + help="""Path to tokens.txt.""", ) parser.add_argument( @@ -128,7 +128,7 @@ def greedy_search( ) device = encoder_out.device - blank_id = 0 # hard-code to 0 + blank_id = model.decoder.blank_id batch_size_list = packed_encoder_out.batch_sizes.tolist() N = encoder_out.size(0) @@ -215,9 +215,6 @@ def main(): model.to(device) - sp = spm.SentencePieceProcessor() - sp.load(args.bpe_model) - logging.info("Constructing Fbank computer") opts = kaldifeat.FbankOptions() opts.device = device @@ -256,10 +253,21 @@ def main(): encoder_out=encoder_out, encoder_out_lens=encoder_out_lens, ) + s = "\n" + + token_table = k2.SymbolTable.from_file(args.tokens) + + def token_ids_to_words(token_ids: List[int]) -> str: + text = "" + for i in token_ids: + text += token_table[i] + return text.replace("▁", " ").strip() + for filename, hyp in zip(args.sound_files, hyps): - words = sp.decode(hyp) - s += f"{filename}:\n{words}\n\n" + words = token_ids_to_words(hyp) + s += f"{filename}:\n{words}\n" + logging.info(s) logging.info("Decoding Done") diff --git a/egs/librispeech/ASR/zipformer/jit_pretrained_ctc.py b/egs/librispeech/ASR/zipformer/jit_pretrained_ctc.py index 1ec390d5b..904d8cd76 100755 --- a/egs/librispeech/ASR/zipformer/jit_pretrained_ctc.py +++ b/egs/librispeech/ASR/zipformer/jit_pretrained_ctc.py @@ -24,7 +24,7 @@ You can generate the checkpoint with the following command: ./zipformer/export.py \ --exp-dir ./zipformer/exp \ --use-ctc 1 \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 \ --jit 1 @@ -35,7 +35,7 @@ You can generate the checkpoint with the following command: --exp-dir ./zipformer/exp \ --use-ctc 1 \ --causal 1 \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 \ --jit 1 @@ -45,7 +45,7 @@ Usage of this script: (1) ctc-decoding ./zipformer/jit_pretrained_ctc.py \ --model-filename ./zipformer/exp/jit_script.pt \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --method ctc-decoding \ --sample-rate 16000 \ /path/to/foo.wav \ @@ -91,10 +91,10 @@ from typing import List import k2 import kaldifeat -import sentencepiece as spm import torch import torchaudio from ctc_decode import get_decoding_params +from export import num_tokens from torch.nn.utils.rnn import pad_sequence from train import get_params @@ -136,9 +136,9 @@ def get_parser(): ) parser.add_argument( - "--bpe-model", + "--tokens", type=str, - help="""Path to bpe.model. + help="""Path to tokens.txt. Used only when method is ctc-decoding. """, ) @@ -149,8 +149,8 @@ def get_parser(): default="1best", help="""Decoding method. Possible values are: - (0) ctc-decoding - Use CTC decoding. It uses a sentence - piece model, i.e., lang_dir/bpe.model, to convert + (0) ctc-decoding - Use CTC decoding. It uses a token table, + i.e., lang_dir/token.txt, to convert word pieces to words. It needs neither a lexicon nor an n-gram LM. (1) 1best - Use the best path as decoding output. Only @@ -263,10 +263,8 @@ def main(): params.update(get_decoding_params()) params.update(vars(args)) - sp = spm.SentencePieceProcessor() - sp.load(params.bpe_model) - - params.vocab_size = sp.get_piece_size() + token_table = k2.SymbolTable.from_file(params.tokens) + params.vocab_size = num_tokens(token_table) + 1 logging.info(f"{params}") @@ -340,8 +338,7 @@ def main(): lattice=lattice, use_double_scores=params.use_double_scores ) token_ids = get_texts(best_path) - hyps = sp.decode(token_ids) - hyps = [s.split() for s in hyps] + hyps = [[token_table[i] for i in ids] for ids in token_ids] elif params.method in [ "1best", "nbest-rescoring", @@ -415,6 +412,7 @@ def main(): s = "\n" for filename, hyp in zip(params.sound_files, hyps): words = " ".join(hyp) + words = words.replace("▁", " ").strip() s += f"{filename}:\n{words}\n\n" logging.info(s) diff --git a/egs/librispeech/ASR/zipformer/jit_pretrained_streaming.py b/egs/librispeech/ASR/zipformer/jit_pretrained_streaming.py index 58d736685..d4ceacefd 100755 --- a/egs/librispeech/ASR/zipformer/jit_pretrained_streaming.py +++ b/egs/librispeech/ASR/zipformer/jit_pretrained_streaming.py @@ -25,7 +25,7 @@ You can use the following command to get the exported models: --causal 1 \ --chunk-size 16 \ --left-context-frames 128 \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 \ --jit 1 @@ -34,7 +34,7 @@ Usage of this script: ./zipformer/jit_pretrained_streaming.py \ --nn-model-filename ./zipformer/exp-causal/jit_script_chunk_16_left_128.pt \ - --bpe-model ./data/lang_bpe_500/bpe.model \ + --tokens ./data/lang_bpe_500/tokens.txt \ /path/to/foo.wav \ """ @@ -43,8 +43,8 @@ import logging import math from typing import List, Optional +import k2 import kaldifeat -import sentencepiece as spm import torch import torchaudio from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature @@ -60,13 +60,13 @@ def get_parser(): "--nn-model-filename", type=str, required=True, - help="Path to the torchscript model cpu_jit.pt", + help="Path to the torchscript model jit_script.pt", ) parser.add_argument( - "--bpe-model", + "--tokens", type=str, - help="""Path to bpe.model.""", + help="""Path to tokens.txt.""", ) parser.add_argument( @@ -120,8 +120,8 @@ def greedy_search( device: torch.device = torch.device("cpu"), ): assert encoder_out.ndim == 2 - context_size = 2 - blank_id = 0 + context_size = decoder.context_size + blank_id = decoder.blank_id if decoder_out is None: assert hyp is None, hyp @@ -190,8 +190,8 @@ def main(): decoder = model.decoder joiner = model.joiner - sp = spm.SentencePieceProcessor() - sp.load(args.bpe_model) + token_table = k2.SymbolTable.from_file(args.tokens) + context_size = decoder.context_size logging.info("Constructing Fbank computer") online_fbank = create_streaming_feature_extractor(args.sample_rate) @@ -250,9 +250,13 @@ def main(): decoder, joiner, encoder_out.squeeze(0), decoder_out, hyp, device=device ) - context_size = 2 + text = "" + for i in hyp[context_size:]: + text += token_table[i] + text = text.replace("▁", " ").strip() + logging.info(args.sound_file) - logging.info(sp.decode(hyp[context_size:])) + logging.info(text) logging.info("Decoding Done") diff --git a/egs/librispeech/ASR/zipformer/model.py b/egs/librispeech/ASR/zipformer/model.py index b541ee697..f2f86af47 100644 --- a/egs/librispeech/ASR/zipformer/model.py +++ b/egs/librispeech/ASR/zipformer/model.py @@ -320,7 +320,7 @@ class AsrModel(nn.Module): assert x_lens.ndim == 1, x_lens.shape assert y.num_axes == 2, y.num_axes - assert x.size(0) == x_lens.size(0) == y.dim0 + assert x.size(0) == x_lens.size(0) == y.dim0, (x.shape, x_lens.shape, y.dim0) # Compute encoder outputs encoder_out, encoder_out_lens = self.forward_encoder(x, x_lens) diff --git a/egs/librispeech/ASR/zipformer/onnx_check.py b/egs/librispeech/ASR/zipformer/onnx_check.py new file mode 100755 index 000000000..b38b875d0 --- /dev/null +++ b/egs/librispeech/ASR/zipformer/onnx_check.py @@ -0,0 +1,241 @@ +#!/usr/bin/env python3 +# +# Copyright 2022 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +This script checks that exported onnx models produce the same output +with the given torchscript model for the same input. + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "data/lang_bpe_500/tokens.txt" +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model via torchscript (torch.jit.script()) + +./zipformer/export.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp/ \ + --jit 1 + +It will generate the following file in $repo/exp: + - jit_script.pt + +3. Export the model to ONNX + +./zipformer/export-onnx.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp/ + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1.onnx + - decoder-epoch-99-avg-1.onnx + - joiner-epoch-99-avg-1.onnx + +4. Run this file + +./zipformer/onnx_check.py \ + --jit-filename $repo/exp/jit_script.pt \ + --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \ + --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \ + --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx +""" + +import argparse +import logging + +import torch +from onnx_pretrained import OnnxModel + +from icefall import is_module_available + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--jit-filename", + required=True, + type=str, + help="Path to the torchscript model", + ) + + parser.add_argument( + "--onnx-encoder-filename", + required=True, + type=str, + help="Path to the onnx encoder model", + ) + + parser.add_argument( + "--onnx-decoder-filename", + required=True, + type=str, + help="Path to the onnx decoder model", + ) + + parser.add_argument( + "--onnx-joiner-filename", + required=True, + type=str, + help="Path to the onnx joiner model", + ) + + return parser + + +def test_encoder( + torch_model: torch.jit.ScriptModule, + onnx_model: OnnxModel, +): + C = 80 + for i in range(3): + N = torch.randint(low=1, high=20, size=(1,)).item() + T = torch.randint(low=30, high=50, size=(1,)).item() + logging.info(f"test_encoder: iter {i}, N={N}, T={T}") + + x = torch.rand(N, T, C) + x_lens = torch.randint(low=30, high=T + 1, size=(N,)) + x_lens[0] = T + + torch_encoder_out, torch_encoder_out_lens = torch_model.encoder(x, x_lens) + torch_encoder_out = torch_model.joiner.encoder_proj(torch_encoder_out) + + onnx_encoder_out, onnx_encoder_out_lens = onnx_model.run_encoder(x, x_lens) + + assert torch.allclose(torch_encoder_out, onnx_encoder_out, atol=1e-05), ( + (torch_encoder_out - onnx_encoder_out).abs().max() + ) + + +def test_decoder( + torch_model: torch.jit.ScriptModule, + onnx_model: OnnxModel, +): + context_size = onnx_model.context_size + vocab_size = onnx_model.vocab_size + for i in range(10): + N = torch.randint(1, 100, size=(1,)).item() + logging.info(f"test_decoder: iter {i}, N={N}") + x = torch.randint( + low=1, + high=vocab_size, + size=(N, context_size), + dtype=torch.int64, + ) + torch_decoder_out = torch_model.decoder(x, need_pad=torch.tensor([False])) + torch_decoder_out = torch_model.joiner.decoder_proj(torch_decoder_out) + torch_decoder_out = torch_decoder_out.squeeze(1) + + onnx_decoder_out = onnx_model.run_decoder(x) + assert torch.allclose(torch_decoder_out, onnx_decoder_out, atol=1e-4), ( + (torch_decoder_out - onnx_decoder_out).abs().max() + ) + + +def test_joiner( + torch_model: torch.jit.ScriptModule, + onnx_model: OnnxModel, +): + encoder_dim = torch_model.joiner.encoder_proj.weight.shape[1] + decoder_dim = torch_model.joiner.decoder_proj.weight.shape[1] + for i in range(10): + N = torch.randint(1, 100, size=(1,)).item() + logging.info(f"test_joiner: iter {i}, N={N}") + encoder_out = torch.rand(N, encoder_dim) + decoder_out = torch.rand(N, decoder_dim) + + projected_encoder_out = torch_model.joiner.encoder_proj(encoder_out) + projected_decoder_out = torch_model.joiner.decoder_proj(decoder_out) + + torch_joiner_out = torch_model.joiner(encoder_out, decoder_out) + onnx_joiner_out = onnx_model.run_joiner( + projected_encoder_out, projected_decoder_out + ) + + assert torch.allclose(torch_joiner_out, onnx_joiner_out, atol=1e-4), ( + (torch_joiner_out - onnx_joiner_out).abs().max() + ) + + +@torch.no_grad() +def main(): + args = get_parser().parse_args() + logging.info(vars(args)) + + torch_model = torch.jit.load(args.jit_filename) + + onnx_model = OnnxModel( + encoder_model_filename=args.onnx_encoder_filename, + decoder_model_filename=args.onnx_decoder_filename, + joiner_model_filename=args.onnx_joiner_filename, + ) + + logging.info("Test encoder") + test_encoder(torch_model, onnx_model) + + logging.info("Test decoder") + test_decoder(torch_model, onnx_model) + + logging.info("Test joiner") + test_joiner(torch_model, onnx_model) + logging.info("Finished checking ONNX models") + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +# See https://github.com/pytorch/pytorch/issues/38342 +# and https://github.com/pytorch/pytorch/issues/33354 +# +# If we don't do this, the delay increases whenever there is +# a new request that changes the actual batch size. +# If you use `py-spy dump --pid --native`, you will +# see a lot of time is spent in re-compiling the torch script model. +torch._C._jit_set_profiling_executor(False) +torch._C._jit_set_profiling_mode(False) +torch._C._set_graph_executor_optimize(False) +if __name__ == "__main__": + torch.manual_seed(20220727) + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/librispeech/ASR/zipformer/onnx_decode.py b/egs/librispeech/ASR/zipformer/onnx_decode.py new file mode 100755 index 000000000..2aca36ca9 --- /dev/null +++ b/egs/librispeech/ASR/zipformer/onnx_decode.py @@ -0,0 +1,323 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This script loads ONNX exported models and uses them to decode the test sets. + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "data/lang_bpe_500/bpe.model" +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model to ONNX + +./zipformer/export-onnx.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp \ + --causal False + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1.onnx + - decoder-epoch-99-avg-1.onnx + - joiner-epoch-99-avg-1.onnx + +2. Run this file + +./zipformer/onnx_decode.py \ + --exp-dir $repo/exp \ + --max-duration 600 \ + --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \ + --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \ + --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ +""" + + +import argparse +import logging +import time +from pathlib import Path +from typing import List, Tuple + +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule + +from onnx_pretrained import greedy_search, OnnxModel + +from icefall.utils import setup_logger, store_transcripts, write_error_stats +from k2 import SymbolTable + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--encoder-model-filename", + type=str, + required=True, + help="Path to the encoder onnx model. ", + ) + + parser.add_argument( + "--decoder-model-filename", + type=str, + required=True, + help="Path to the decoder onnx model. ", + ) + + parser.add_argument( + "--joiner-model-filename", + type=str, + required=True, + help="Path to the joiner onnx model. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--tokens", + type=str, + help="""Path to tokens.txt.""", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="Valid values are greedy_search and modified_beam_search", + ) + + return parser + + +def decode_one_batch( + model: OnnxModel, token_table: SymbolTable, batch: dict +) -> List[List[str]]: + """Decode one batch and return the result. + Currently it only greedy_search is supported. + + Args: + model: + The neural model. + token_table: + The token table. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + + Returns: + Return the decoded results for each utterance. + """ + feature = batch["inputs"] + assert feature.ndim == 3 + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(dtype=torch.int64) + + encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens) + + hyps = greedy_search( + model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens + ) + + def token_ids_to_words(token_ids: List[int]) -> str: + text = "" + for i in token_ids: + text += token_table[i] + return text.replace("▁", " ").strip() + + hyps = [token_ids_to_words(h).split() for h in hyps] + return hyps + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + model: nn.Module, + token_table: SymbolTable, +) -> Tuple[List[Tuple[str, List[str], List[str]]], float]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + model: + The neural model. + token_table: + The token table. + + Returns: + - A list of tuples. Each tuple contains three elements: + - cut_id, + - reference transcript, + - predicted result. + - The total duration (in seconds) of the dataset. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + log_interval = 10 + total_duration = 0 + + results = [] + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]]) + + hyps = decode_one_batch(model=model, token_table=token_table, batch=batch) + + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results.extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + + return results, total_duration + + +def save_results( + res_dir: Path, + test_set_name: str, + results: List[Tuple[str, List[str], List[str]]], +): + recog_path = res_dir / f"recogs-{test_set_name}.txt" + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = res_dir / f"errs-{test_set_name}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True) + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + errs_info = res_dir / f"wer-summary-{test_set_name}.txt" + with open(errs_info, "w") as f: + print("WER", file=f) + print(wer, file=f) + + s = "\nFor {}, WER is {}:\n".format(test_set_name, wer) + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + + assert ( + args.decoding_method == "greedy_search" + ), "Only supports greedy_search currently." + res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}" + + setup_logger(f"{res_dir}/log-decode") + logging.info("Decoding started") + + device = torch.device("cpu") + logging.info(f"Device: {device}") + + token_table = SymbolTable.from_file(args.tokens) + + logging.info(vars(args)) + + logging.info("About to create model") + model = OnnxModel( + encoder_model_filename=args.encoder_model_filename, + decoder_model_filename=args.decoder_model_filename, + joiner_model_filename=args.joiner_model_filename, + ) + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + start_time = time.time() + results, total_duration = decode_dataset(dl=test_dl, model=model, token_table=token_table) + end_time = time.time() + elapsed_seconds = end_time - start_time + rtf = elapsed_seconds / total_duration + + logging.info(f"Elapsed time: {elapsed_seconds:.3f} s") + logging.info(f"Wave duration: {total_duration:.3f} s") + logging.info( + f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}" + ) + + save_results(res_dir=res_dir, test_set_name=test_set, results=results) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained-streaming.py b/egs/librispeech/ASR/zipformer/onnx_pretrained-streaming.py index 273f883df..2ce4506a8 100755 --- a/egs/librispeech/ASR/zipformer/onnx_pretrained-streaming.py +++ b/egs/librispeech/ASR/zipformer/onnx_pretrained-streaming.py @@ -524,11 +524,11 @@ def main(): hyp, ) - symbol_table = k2.SymbolTable.from_file(args.tokens) + token_table = k2.SymbolTable.from_file(args.tokens) text = "" for i in hyp[context_size:]: - text += symbol_table[i] + text += token_table[i] text = text.replace("▁", " ").strip() logging.info(args.sound_file) diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained.py b/egs/librispeech/ASR/zipformer/onnx_pretrained.py deleted file mode 120000 index 0069288fe..000000000 --- a/egs/librispeech/ASR/zipformer/onnx_pretrained.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless7/onnx_pretrained.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained.py b/egs/librispeech/ASR/zipformer/onnx_pretrained.py new file mode 100755 index 000000000..e8a521460 --- /dev/null +++ b/egs/librispeech/ASR/zipformer/onnx_pretrained.py @@ -0,0 +1,419 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This script loads ONNX models and uses them to decode waves. +You can use the following command to get the exported models: + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "data/lang_bpe_500/tokens.txt" +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model to ONNX + +./zipformer/export-onnx.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp \ + --causal False + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1.onnx + - decoder-epoch-99-avg-1.onnx + - joiner-epoch-99-avg-1.onnx + +3. Run this file + +./zipformer/onnx_pretrained.py \ + --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \ + --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \ + --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + $repo/test_wavs/1089-134686-0001.wav \ + $repo/test_wavs/1221-135766-0001.wav \ + $repo/test_wavs/1221-135766-0002.wav +""" + +import argparse +import logging +import math +from typing import List, Tuple + +import k2 +import kaldifeat +import onnxruntime as ort +import torch +import torchaudio +from torch.nn.utils.rnn import pad_sequence + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--encoder-model-filename", + type=str, + required=True, + help="Path to the encoder onnx model. ", + ) + + parser.add_argument( + "--decoder-model-filename", + type=str, + required=True, + help="Path to the decoder onnx model. ", + ) + + parser.add_argument( + "--joiner-model-filename", + type=str, + required=True, + help="Path to the joiner onnx model. ", + ) + + parser.add_argument( + "--tokens", + type=str, + help="""Path to tokens.txt.""", + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + return parser + + +class OnnxModel: + def __init__( + self, + encoder_model_filename: str, + decoder_model_filename: str, + joiner_model_filename: str, + ): + session_opts = ort.SessionOptions() + session_opts.inter_op_num_threads = 1 + session_opts.intra_op_num_threads = 4 + + self.session_opts = session_opts + + self.init_encoder(encoder_model_filename) + self.init_decoder(decoder_model_filename) + self.init_joiner(joiner_model_filename) + + def init_encoder(self, encoder_model_filename: str): + self.encoder = ort.InferenceSession( + encoder_model_filename, + sess_options=self.session_opts, + ) + + def init_decoder(self, decoder_model_filename: str): + self.decoder = ort.InferenceSession( + decoder_model_filename, + sess_options=self.session_opts, + ) + + decoder_meta = self.decoder.get_modelmeta().custom_metadata_map + self.context_size = int(decoder_meta["context_size"]) + self.vocab_size = int(decoder_meta["vocab_size"]) + + logging.info(f"context_size: {self.context_size}") + logging.info(f"vocab_size: {self.vocab_size}") + + def init_joiner(self, joiner_model_filename: str): + self.joiner = ort.InferenceSession( + joiner_model_filename, + sess_options=self.session_opts, + ) + + joiner_meta = self.joiner.get_modelmeta().custom_metadata_map + self.joiner_dim = int(joiner_meta["joiner_dim"]) + + logging.info(f"joiner_dim: {self.joiner_dim}") + + def run_encoder( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + A 3-D tensor of shape (N, T, C) + x_lens: + A 2-D tensor of shape (N,). Its dtype is torch.int64 + Returns: + Return a tuple containing: + - encoder_out, its shape is (N, T', joiner_dim) + - encoder_out_lens, its shape is (N,) + """ + out = self.encoder.run( + [ + self.encoder.get_outputs()[0].name, + self.encoder.get_outputs()[1].name, + ], + { + self.encoder.get_inputs()[0].name: x.numpy(), + self.encoder.get_inputs()[1].name: x_lens.numpy(), + }, + ) + return torch.from_numpy(out[0]), torch.from_numpy(out[1]) + + def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor: + """ + Args: + decoder_input: + A 2-D tensor of shape (N, context_size) + Returns: + Return a 2-D tensor of shape (N, joiner_dim) + """ + out = self.decoder.run( + [self.decoder.get_outputs()[0].name], + {self.decoder.get_inputs()[0].name: decoder_input.numpy()}, + )[0] + + return torch.from_numpy(out) + + def run_joiner( + self, encoder_out: torch.Tensor, decoder_out: torch.Tensor + ) -> torch.Tensor: + """ + Args: + encoder_out: + A 2-D tensor of shape (N, joiner_dim) + decoder_out: + A 2-D tensor of shape (N, joiner_dim) + Returns: + Return a 2-D tensor of shape (N, vocab_size) + """ + out = self.joiner.run( + [self.joiner.get_outputs()[0].name], + { + self.joiner.get_inputs()[0].name: encoder_out.numpy(), + self.joiner.get_inputs()[1].name: decoder_out.numpy(), + }, + )[0] + + return torch.from_numpy(out) + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert ( + sample_rate == expected_sample_rate + ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0]) + return ans + + +def greedy_search( + model: OnnxModel, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, +) -> List[List[int]]: + """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. + Args: + model: + The transducer model. + encoder_out: + A 3-D tensor of shape (N, T, joiner_dim) + encoder_out_lens: + A 1-D tensor of shape (N,). + Returns: + Return the decoded results for each utterance. + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert encoder_out.size(0) >= 1, encoder_out.size(0) + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + blank_id = 0 # hard-code to 0 + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + context_size = model.context_size + hyps = [[blank_id] * context_size for _ in range(N)] + + decoder_input = torch.tensor( + hyps, + dtype=torch.int64, + ) # (N, context_size) + + decoder_out = model.run_decoder(decoder_input) + + offset = 0 + for batch_size in batch_size_list: + start = offset + end = offset + batch_size + current_encoder_out = packed_encoder_out.data[start:end] + # current_encoder_out's shape: (batch_size, joiner_dim) + offset = end + + decoder_out = decoder_out[:batch_size] + logits = model.run_joiner(current_encoder_out, decoder_out) + + # logits'shape (batch_size, vocab_size) + + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v != blank_id: + hyps[i].append(v) + emitted = True + if emitted: + # update decoder output + decoder_input = [h[-context_size:] for h in hyps[:batch_size]] + decoder_input = torch.tensor( + decoder_input, + dtype=torch.int64, + ) + decoder_out = model.run_decoder(decoder_input) + + sorted_ans = [h[context_size:] for h in hyps] + ans = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + logging.info(vars(args)) + model = OnnxModel( + encoder_model_filename=args.encoder_model_filename, + decoder_model_filename=args.decoder_model_filename, + joiner_model_filename=args.joiner_model_filename, + ) + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = "cpu" + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = args.sample_rate + opts.mel_opts.num_bins = 80 + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {args.sound_files}") + waves = read_sound_files( + filenames=args.sound_files, + expected_sample_rate=args.sample_rate, + ) + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence( + features, + batch_first=True, + padding_value=math.log(1e-10), + ) + + feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64) + encoder_out, encoder_out_lens = model.run_encoder(features, feature_lengths) + + hyps = greedy_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + s = "\n" + + token_table = k2.SymbolTable.from_file(args.tokens) + + def token_ids_to_words(token_ids: List[int]) -> str: + text = "" + for i in token_ids: + text += token_table[i] + return text.replace("▁", " ").strip() + + for filename, hyp in zip(args.sound_files, hyps): + words = token_ids_to_words(hyp) + s += f"{filename}:\n{words}\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/librispeech/ASR/zipformer/pretrained.py b/egs/librispeech/ASR/zipformer/pretrained.py index 2944f79e3..3104b6084 100755 --- a/egs/librispeech/ASR/zipformer/pretrained.py +++ b/egs/librispeech/ASR/zipformer/pretrained.py @@ -18,11 +18,14 @@ This script loads a checkpoint and uses it to decode waves. You can generate the checkpoint with the following command: +Note: This is a example for librispeech dataset, if you are using different +dataset, you should change the argument values according to your dataset. + - For non-streaming model: ./zipformer/export.py \ --exp-dir ./zipformer/exp \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 @@ -31,7 +34,7 @@ You can generate the checkpoint with the following command: ./zipformer/export.py \ --exp-dir ./zipformer/exp \ --causal 1 \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 @@ -42,7 +45,7 @@ Usage of this script: (1) greedy search ./zipformer/pretrained.py \ --checkpoint ./zipformer/exp/pretrained.pt \ - --bpe-model ./data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --method greedy_search \ /path/to/foo.wav \ /path/to/bar.wav @@ -50,7 +53,7 @@ Usage of this script: (2) modified beam search ./zipformer/pretrained.py \ --checkpoint ./zipformer/exp/pretrained.pt \ - --bpe-model ./data/lang_bpe_500/bpe.model \ + --tokens ./data/lang_bpe_500/tokens.txt \ --method modified_beam_search \ /path/to/foo.wav \ /path/to/bar.wav @@ -58,7 +61,7 @@ Usage of this script: (3) fast beam search ./zipformer/pretrained.py \ --checkpoint ./zipformer/exp/pretrained.pt \ - --bpe-model ./data/lang_bpe_500/bpe.model \ + --tokens ./data/lang_bpe_500/tokens.txt \ --method fast_beam_search \ /path/to/foo.wav \ /path/to/bar.wav @@ -71,7 +74,7 @@ Usage of this script: --causal 1 \ --chunk-size 16 \ --left-context-frames 128 \ - --bpe-model ./data/lang_bpe_500/bpe.model \ + --tokens ./data/lang_bpe_500/tokens.txt \ --method greedy_search \ /path/to/foo.wav \ /path/to/bar.wav @@ -82,7 +85,7 @@ Usage of this script: --causal 1 \ --chunk-size 16 \ --left-context-frames 128 \ - --bpe-model ./data/lang_bpe_500/bpe.model \ + --tokens ./data/lang_bpe_500/tokens.txt \ --method modified_beam_search \ /path/to/foo.wav \ /path/to/bar.wav @@ -93,7 +96,7 @@ Usage of this script: --causal 1 \ --chunk-size 16 \ --left-context-frames 128 \ - --bpe-model ./data/lang_bpe_500/bpe.model \ + --tokens ./data/lang_bpe_500/tokens.txt \ --method fast_beam_search \ /path/to/foo.wav \ /path/to/bar.wav @@ -112,7 +115,6 @@ from typing import List import k2 import kaldifeat -import sentencepiece as spm import torch import torchaudio from beam_search import ( @@ -120,8 +122,11 @@ from beam_search import ( greedy_search_batch, modified_beam_search, ) +from export import num_tokens from torch.nn.utils.rnn import pad_sequence -from train import add_model_arguments, get_params, get_model +from train import add_model_arguments, get_model, get_params + +from icefall.utils import make_pad_mask def get_parser(): @@ -139,9 +144,9 @@ def get_parser(): ) parser.add_argument( - "--bpe-model", + "--tokens", type=str, - help="""Path to bpe.model.""", + help="""Path to tokens.txt.""", ) parser.add_argument( @@ -258,13 +263,11 @@ def main(): params.update(vars(args)) - sp = spm.SentencePieceProcessor() - sp.load(params.bpe_model) + token_table = k2.SymbolTable.from_file(params.tokens) - # is defined in local/train_bpe_model.py - params.blank_id = sp.piece_to_id("") - params.unk_id = sp.piece_to_id("") - params.vocab_size = sp.get_piece_size() + params.blank_id = token_table[""] + params.unk_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 logging.info(f"{params}") @@ -323,6 +326,12 @@ def main(): msg = f"Using {params.method}" logging.info(msg) + def token_ids_to_words(token_ids: List[int]) -> str: + text = "" + for i in token_ids: + text += token_table[i] + return text.replace("▁", " ").strip() + if params.method == "fast_beam_search": decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) hyp_tokens = fast_beam_search_one_best( @@ -334,8 +343,8 @@ def main(): max_contexts=params.max_contexts, max_states=params.max_states, ) - for hyp in sp.decode(hyp_tokens): - hyps.append(hyp.split()) + for hyp in hyp_tokens: + hyps.append(token_ids_to_words(hyp)) elif params.method == "modified_beam_search": hyp_tokens = modified_beam_search( model=model, @@ -344,23 +353,22 @@ def main(): beam=params.beam_size, ) - for hyp in sp.decode(hyp_tokens): - hyps.append(hyp.split()) + for hyp in hyp_tokens: + hyps.append(token_ids_to_words(hyp)) elif params.method == "greedy_search" and params.max_sym_per_frame == 1: hyp_tokens = greedy_search_batch( model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens, ) - for hyp in sp.decode(hyp_tokens): - hyps.append(hyp.split()) + for hyp in hyp_tokens: + hyps.append(token_ids_to_words(hyp)) else: raise ValueError(f"Unsupported method: {params.method}") s = "\n" for filename, hyp in zip(params.sound_files, hyps): - words = " ".join(hyp) - s += f"{filename}:\n{words}\n\n" + s += f"{filename}:\n{hyp}\n\n" logging.info(s) logging.info("Decoding Done") diff --git a/egs/librispeech/ASR/zipformer/pretrained_ctc.py b/egs/librispeech/ASR/zipformer/pretrained_ctc.py index f10d95449..be239e9c3 100755 --- a/egs/librispeech/ASR/zipformer/pretrained_ctc.py +++ b/egs/librispeech/ASR/zipformer/pretrained_ctc.py @@ -24,7 +24,7 @@ You can generate the checkpoint with the following command: ./zipformer/export.py \ --exp-dir ./zipformer/exp \ --use-ctc 1 \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 @@ -34,7 +34,7 @@ You can generate the checkpoint with the following command: --exp-dir ./zipformer/exp \ --use-ctc 1 \ --causal 1 \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 @@ -43,7 +43,7 @@ Usage of this script: (1) ctc-decoding ./zipformer/pretrained_ctc.py \ --checkpoint ./zipformer/exp/pretrained.pt \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --method ctc-decoding \ --sample-rate 16000 \ /path/to/foo.wav \ @@ -90,12 +90,12 @@ from typing import List import k2 import kaldifeat -import sentencepiece as spm import torch import torchaudio from ctc_decode import get_decoding_params +from export import num_tokens from torch.nn.utils.rnn import pad_sequence -from train import add_model_arguments, get_params, get_model +from train import add_model_arguments, get_model, get_params from icefall.decode import ( get_lattice, @@ -144,9 +144,9 @@ def get_parser(): ) parser.add_argument( - "--bpe-model", + "--tokens", type=str, - help="""Path to bpe.model. + help="""Path to tokens.txt. Used only when method is ctc-decoding. """, ) @@ -157,8 +157,8 @@ def get_parser(): default="1best", help="""Decoding method. Possible values are: - (0) ctc-decoding - Use CTC decoding. It uses a sentence - piece model, i.e., lang_dir/bpe.model, to convert + (0) ctc-decoding - Use CTC decoding. It uses a token table, + i.e., lang_dir/tokens.txt, to convert word pieces to words. It needs neither a lexicon nor an n-gram LM. (1) 1best - Use the best path as decoding output. Only @@ -273,11 +273,10 @@ def main(): params.update(get_decoding_params()) params.update(vars(args)) - sp = spm.SentencePieceProcessor() - sp.load(params.bpe_model) - - params.vocab_size = sp.get_piece_size() - params.blank_id = 0 + token_table = k2.SymbolTable.from_file(params.tokens) + params.vocab_size = num_tokens(token_table) + params.blank_id = token_table[""] + assert params.blank_id == 0 logging.info(f"{params}") @@ -358,8 +357,7 @@ def main(): lattice=lattice, use_double_scores=params.use_double_scores ) token_ids = get_texts(best_path) - hyps = sp.decode(token_ids) - hyps = [s.split() for s in hyps] + hyps = [[token_table[i] for i in ids] for ids in token_ids] elif params.method in [ "1best", "nbest-rescoring", @@ -433,6 +431,7 @@ def main(): s = "\n" for filename, hyp in zip(params.sound_files, hyps): words = " ".join(hyp) + words = words.replace("▁", " ").strip() s += f"{filename}:\n{words}\n\n" logging.info(s) diff --git a/egs/librispeech/ASR/zipformer/scaling.py b/egs/librispeech/ASR/zipformer/scaling.py index 9f23eeead..7c98ef045 100644 --- a/egs/librispeech/ASR/zipformer/scaling.py +++ b/egs/librispeech/ASR/zipformer/scaling.py @@ -25,6 +25,11 @@ import math import torch.nn as nn from torch import Tensor +def logaddexp_onnx(x: Tensor, y: Tensor) -> Tensor: + max_value = torch.max(x, y) + diff = torch.abs(x - y) + return max_value + torch.log1p(torch.exp(-diff)) + # RuntimeError: Exporting the operator logaddexp to ONNX opset version # 14 is not supported. Please feel free to request support or submit @@ -33,10 +38,22 @@ from torch import Tensor # The following function is to solve the above error when exporting # models to ONNX via torch.jit.trace() def logaddexp(x: Tensor, y: Tensor) -> Tensor: - if not torch.jit.is_tracing(): + # Caution(fangjun): Put torch.jit.is_scripting() before + # torch.onnx.is_in_onnx_export(); + # otherwise, it will cause errors for torch.jit.script(). + # + # torch.logaddexp() works for both torch.jit.script() and + # torch.jit.trace() but it causes errors for ONNX export. + # + if torch.jit.is_scripting(): + # Note: We cannot use torch.jit.is_tracing() here as it also + # matches torch.onnx.export(). return torch.logaddexp(x, y) + elif torch.onnx.is_in_onnx_export(): + return logaddexp_onnx(x, y) else: - return (x.exp() + y.exp()).log() + # for torch.jit.trace() + return torch.logaddexp(x, y) class PiecewiseLinear(object): """ @@ -108,7 +125,7 @@ class PiecewiseLinear(object): p: 'PiecewiseLinear', include_crossings: bool = False): """ - Returns (self_mod, p_mod) which are equivalent piecewise lienar + Returns (self_mod, p_mod) which are equivalent piecewise linear functions to self and p, but with the same x values. p: the other piecewise linear function @@ -149,7 +166,7 @@ class ScheduledFloat(torch.nn.Module): in, float(parent_module.whatever), and use it as something like a dropout prob. It is a floating point value whose value changes depending on the batch count of the - training loop. It is a piecewise linear function where you specifiy the (x,y) pairs + training loop. It is a piecewise linear function where you specify the (x,y) pairs in sorted order on x; x corresponds to the batch index. For batch-index values before the first x or after the last x, we just use the first or last y value. @@ -326,7 +343,7 @@ class MaxEigLimiterFunction(torch.autograd.Function): class BiasNormFunction(torch.autograd.Function): # This computes: # scales = (torch.mean((x - bias) ** 2, keepdim=True)) ** -0.5 * log_scale.exp() - # return (x - bias) * scales + # return x * scales # (after unsqueezing the bias), but it does it in a memory-efficient way so that # it can just store the returned value (chances are, this will also be needed for # some other reason, related to the next operation, so we can save memory). @@ -383,8 +400,8 @@ class BiasNorm(torch.nn.Module): Args: num_channels: the number of channels, e.g. 512. channel_dim: the axis/dimension corresponding to the channel, - interprted as an offset from the input's ndim if negative. - shis is NOT the num_channels; it should typically be one of + interpreted as an offset from the input's ndim if negative. + This is NOT the num_channels; it should typically be one of {-2, -1, 0, 1, 2, 3}. log_scale: the initial log-scale that we multiply the output by; this is learnable. @@ -1269,7 +1286,7 @@ class Dropout3(nn.Module): class SwooshLFunction(torch.autograd.Function): """ - swoosh(x) = log(1 + exp(x-4)) - 0.08*x - 0.035 + swoosh_l(x) = log(1 + exp(x-4)) - 0.08*x - 0.035 """ @staticmethod @@ -1334,10 +1351,17 @@ class SwooshL(torch.nn.Module): return k2.swoosh_l(x) # return SwooshLFunction.apply(x) +class SwooshLOnnx(torch.nn.Module): + def forward(self, x: Tensor) -> Tensor: + """Return Swoosh-L activation. + """ + zero = torch.tensor(0.0, dtype=x.dtype, device=x.device) + return logaddexp_onnx(zero, x - 4.0) - 0.08 * x - 0.035 + class SwooshRFunction(torch.autograd.Function): """ - swoosh(x) = log(1 + exp(x-1)) - 0.08*x - 0.313261687 + swoosh_r(x) = log(1 + exp(x-1)) - 0.08*x - 0.313261687 derivatives are between -0.08 and 0.92. """ @@ -1400,6 +1424,13 @@ class SwooshR(torch.nn.Module): return k2.swoosh_r(x) # return SwooshRFunction.apply(x) +class SwooshROnnx(torch.nn.Module): + def forward(self, x: Tensor) -> Tensor: + """Return Swoosh-R activation. + """ + zero = torch.tensor(0.0, dtype=x.dtype, device=x.device) + return logaddexp_onnx(zero, x - 1.) - 0.08 * x - 0.313261687 + # simple version of SwooshL that does not redefine the backprop, used in # ActivationDropoutAndLinearFunction. diff --git a/egs/librispeech/ASR/zipformer/scaling_converter.py b/egs/librispeech/ASR/zipformer/scaling_converter.py index 54a5c2a6a..76622fa12 100644 --- a/egs/librispeech/ASR/zipformer/scaling_converter.py +++ b/egs/librispeech/ASR/zipformer/scaling_converter.py @@ -26,7 +26,16 @@ from typing import List, Tuple import torch import torch.nn as nn -from scaling import Balancer, Dropout3, ScaleGrad, Whiten +from scaling import ( + Balancer, + Dropout3, + ScaleGrad, + SwooshL, + SwooshLOnnx, + SwooshR, + SwooshROnnx, + Whiten, +) from zipformer import CompactRelPositionalEncoding @@ -75,6 +84,10 @@ def convert_scaled_to_non_scaled( for name, m in model.named_modules(): if isinstance(m, (Balancer, Dropout3, ScaleGrad, Whiten)): d[name] = nn.Identity() + elif is_onnx and isinstance(m, SwooshR): + d[name] = SwooshROnnx() + elif is_onnx and isinstance(m, SwooshL): + d[name] = SwooshLOnnx() elif is_onnx and isinstance(m, CompactRelPositionalEncoding): # We want to recreate the positional encoding vector when # the input changes, so we have to use torch.jit.script() diff --git a/egs/librispeech/ASR/zipformer/streaming_beam_search.py b/egs/librispeech/ASR/zipformer/streaming_beam_search.py index e6e0fb1c8..3c8565b33 100644 --- a/egs/librispeech/ASR/zipformer/streaming_beam_search.py +++ b/egs/librispeech/ASR/zipformer/streaming_beam_search.py @@ -31,6 +31,7 @@ def greedy_search( model: nn.Module, encoder_out: torch.Tensor, streams: List[DecodeStream], + blank_penalty: float = 0.0, ) -> None: """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. @@ -71,6 +72,9 @@ def greedy_search( # logits'shape (batch_size, vocab_size) logits = logits.squeeze(1).squeeze(1) + if blank_penalty != 0.0: + logits[:, 0] -= blank_penalty + assert logits.ndim == 2, logits.shape y = logits.argmax(dim=1).tolist() emitted = False @@ -97,6 +101,7 @@ def modified_beam_search( encoder_out: torch.Tensor, streams: List[DecodeStream], num_active_paths: int = 4, + blank_penalty: float = 0.0, ) -> None: """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded. @@ -158,6 +163,9 @@ def modified_beam_search( logits = logits.squeeze(1).squeeze(1) + if blank_penalty != 0.0: + logits[:, 0] -= blank_penalty + log_probs = logits.log_softmax(dim=-1) # (num_hyps, vocab_size) log_probs.add_(ys_log_probs) @@ -205,6 +213,7 @@ def fast_beam_search_one_best( beam: float, max_states: int, max_contexts: int, + blank_penalty: float = 0.0, ) -> None: """It limits the maximum number of symbols per frame to 1. @@ -269,6 +278,10 @@ def fast_beam_search_one_best( project_input=False, ) logits = logits.squeeze(1).squeeze(1) + + if blank_penalty != 0.0: + logits[:, 0] -= blank_penalty + log_probs = logits.log_softmax(dim=-1) decoding_streams.advance(log_probs) diff --git a/egs/librispeech/ASR/zipformer/subsampling.py b/egs/librispeech/ASR/zipformer/subsampling.py index d6bf57db4..6532ddccb 100644 --- a/egs/librispeech/ASR/zipformer/subsampling.py +++ b/egs/librispeech/ASR/zipformer/subsampling.py @@ -138,9 +138,11 @@ class ConvNeXt(nn.Module): x = bypass + x x = self.out_balancer(x) - x = x.transpose(1, 3) # (N, W, H, C); need channel dim to be last - x = self.out_whiten(x) - x = x.transpose(1, 3) # (N, C, H, W) + + if x.requires_grad: + x = x.transpose(1, 3) # (N, W, H, C); need channel dim to be last + x = self.out_whiten(x) + x = x.transpose(1, 3) # (N, C, H, W) return x @@ -266,6 +268,7 @@ class Conv2dSubsampling(nn.Module): # just one convnext layer self.convnext = ConvNeXt(layer3_channels, kernel_size=(7, 7)) + # (in_channels-3)//4 self.out_width = (((in_channels - 1) // 2) - 1) // 2 self.layer3_channels = layer3_channels @@ -299,7 +302,7 @@ class Conv2dSubsampling(nn.Module): A tensor of shape (batch_size,) containing the number of frames in Returns: - - a tensor of shape (N, ((T-1)//2 - 1)//2, odim) + - a tensor of shape (N, (T-7)//2, odim) - output lengths, of shape (batch_size,) """ # On entry, x is (N, T, idim) @@ -310,14 +313,14 @@ class Conv2dSubsampling(nn.Module): x = self.conv(x) x = self.convnext(x) - # Now x is of shape (N, odim, ((T-3)//2 - 1)//2, ((idim-1)//2 - 1)//2) + # Now x is of shape (N, odim, (T-7)//2, (idim-3)//4) b, c, t, f = x.size() x = x.transpose(1, 2).reshape(b, t, c * f) - # now x: (N, ((T-1)//2 - 1))//2, out_width * layer3_channels)) + # now x: (N, (T-7)//2, out_width * layer3_channels)) x = self.out(x) - # Now x is of shape (N, ((T-1)//2 - 1))//2, odim) + # Now x is of shape (N, (T-7)//2, odim) x = self.out_whiten(x) x = self.out_norm(x) x = self.dropout(x) @@ -328,7 +331,7 @@ class Conv2dSubsampling(nn.Module): with warnings.catch_warnings(): warnings.simplefilter("ignore") x_lens = (x_lens - 7) // 2 - assert x.size(1) == x_lens.max().item() + assert x.size(1) == x_lens.max().item() , (x.size(1), x_lens.max()) return x, x_lens @@ -347,7 +350,7 @@ class Conv2dSubsampling(nn.Module): A tensor of shape (batch_size,) containing the number of frames in Returns: - - a tensor of shape (N, ((T-1)//2 - 1)//2, odim) + - a tensor of shape (N, (T-7)//2, odim) - output lengths, of shape (batch_size,) - updated cache """ @@ -383,7 +386,7 @@ class Conv2dSubsampling(nn.Module): assert self.convnext.padding[0] == 3 x_lens = (x_lens - 7) // 2 - 3 - assert x.size(1) == x_lens.max().item() + assert x.size(1) == x_lens.max().item(), (x.shape, x_lens.max()) return x, x_lens, cached_left_pad diff --git a/egs/librispeech/ASR/zipformer/test_scaling.py b/egs/librispeech/ASR/zipformer/test_scaling.py new file mode 100755 index 000000000..5c04291e7 --- /dev/null +++ b/egs/librispeech/ASR/zipformer/test_scaling.py @@ -0,0 +1,82 @@ +#!/usr/bin/env python3 + +import matplotlib.pyplot as plt +import torch +from scaling import PiecewiseLinear, ScheduledFloat, SwooshL, SwooshR + + +def test_piecewise_linear(): + # An identity map in the range [0, 1]. + # 1 - identity map in the range [1, 2] + # x1=0, y1=0 + # x2=1, y2=1 + # x3=2, y3=0 + pl = PiecewiseLinear((0, 0), (1, 1), (2, 0)) + assert pl(0.25) == 0.25, pl(0.25) + assert pl(0.625) == 0.625, pl(0.625) + assert pl(1.25) == 0.75, pl(1.25) + + assert pl(-10) == pl(0), pl(-10) # out of range + assert pl(10) == pl(2), pl(10) # out of range + + # multiplication + pl10 = pl * 10 + assert pl10(1) == 10 * pl(1) + assert pl10(0.5) == 10 * pl(0.5) + + +def test_scheduled_float(): + # Initial value is 0.2 and it decreases linearly towards 0 at 4000 + dropout = ScheduledFloat((0, 0.2), (4000, 0.0), default=0.0) + dropout.batch_count = 0 + assert float(dropout) == 0.2, (float(dropout), dropout.batch_count) + + dropout.batch_count = 1000 + assert abs(float(dropout) - 0.15) < 1e-5, (float(dropout), dropout.batch_count) + + dropout.batch_count = 2000 + assert float(dropout) == 0.1, (float(dropout), dropout.batch_count) + + dropout.batch_count = 3000 + assert abs(float(dropout) - 0.05) < 1e-5, (float(dropout), dropout.batch_count) + + dropout.batch_count = 4000 + assert float(dropout) == 0.0, (float(dropout), dropout.batch_count) + + dropout.batch_count = 5000 # out of range + assert float(dropout) == 0.0, (float(dropout), dropout.batch_count) + + +def test_swoosh(): + x1 = torch.linspace(start=-10, end=0, steps=100, dtype=torch.float32) + x2 = torch.linspace(start=0, end=10, steps=100, dtype=torch.float32) + x = torch.cat([x1, x2[1:]]) + + left = SwooshL()(x) + r = SwooshR()(x) + + relu = torch.nn.functional.relu(x) + print(left[x == 0], r[x == 0]) + plt.plot(x, left, "k") + plt.plot(x, r, "r") + plt.plot(x, relu, "b") + plt.axis([-10, 10, -1, 10]) # [xmin, xmax, ymin, ymax] + plt.legend( + [ + "SwooshL(x) = log(1 + exp(x-4)) - 0.08x - 0.035 ", + "SwooshR(x) = log(1 + exp(x-1)) - 0.08x - 0.313261687", + "ReLU(x) = max(0, x)", + ] + ) + plt.grid() + plt.savefig("swoosh.pdf") + + +def main(): + test_piecewise_linear() + test_scheduled_float() + test_swoosh() + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/zipformer/test_subsampling.py b/egs/librispeech/ASR/zipformer/test_subsampling.py new file mode 100755 index 000000000..078227fb6 --- /dev/null +++ b/egs/librispeech/ASR/zipformer/test_subsampling.py @@ -0,0 +1,152 @@ +#!/usr/bin/env python3 + +import torch +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling + + +def test_conv2d_subsampling(): + layer1_channels = 8 + layer2_channels = 32 + layer3_channels = 128 + + out_channels = 192 + encoder_embed = Conv2dSubsampling( + in_channels=80, + out_channels=out_channels, + layer1_channels=layer1_channels, + layer2_channels=layer2_channels, + layer3_channels=layer3_channels, + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + N = 2 + T = 200 + num_features = 80 + x = torch.rand(N, T, num_features) + x_copy = x.clone() + + x = x.unsqueeze(1) # (N, 1, T, num_features) + + x = encoder_embed.conv[0](x) # conv2d, in 1, out 8, kernel 3, padding (0,1) + assert x.shape == (N, layer1_channels, T - 2, num_features) + # (2, 8, 198, 80) + + x = encoder_embed.conv[1](x) # scale grad + x = encoder_embed.conv[2](x) # balancer + x = encoder_embed.conv[3](x) # swooshR + + x = encoder_embed.conv[4](x) # conv2d, in 8, out 32, kernel 3, stride 2 + assert x.shape == ( + N, + layer2_channels, + ((T - 2) - 3) // 2 + 1, + (num_features - 3) // 2 + 1, + ) + # (2, 32, 98, 39) + + x = encoder_embed.conv[5](x) # balancer + x = encoder_embed.conv[6](x) # swooshR + + # conv2d: + # in 32, out 128, kernel 3, stride (1, 2) + x = encoder_embed.conv[7](x) + assert x.shape == ( + N, + layer3_channels, + (((T - 2) - 3) // 2 + 1) - 2, + (((num_features - 3) // 2 + 1) - 3) // 2 + 1, + ) + # (2, 128, 96, 19) + + x = encoder_embed.conv[8](x) # balancer + x = encoder_embed.conv[9](x) # swooshR + + # (((T - 2) - 3) // 2 + 1) - 2 + # = (T - 2) - 3) // 2 + 1 - 2 + # = ((T - 2) - 3) // 2 - 1 + # = (T - 2 - 3) // 2 - 1 + # = (T - 5) // 2 - 1 + # = (T - 7) // 2 + assert x.shape[2] == (x_copy.shape[1] - 7) // 2 + + # (((num_features - 3) // 2 + 1) - 3) // 2 + 1, + # = ((num_features - 3) // 2 + 1 - 3) // 2 + 1, + # = ((num_features - 3) // 2 - 2) // 2 + 1, + # = (num_features - 3 - 4) // 2 // 2 + 1, + # = (num_features - 7) // 2 // 2 + 1, + # = (num_features - 7) // 4 + 1, + # = (num_features - 3) // 4 + assert x.shape[3] == (x_copy.shape[2] - 3) // 4 + + assert x.shape == (N, layer3_channels, (T - 7) // 2, (num_features - 3) // 4) + + # Input shape to convnext is + # + # (N, layer3_channels, (T-7)//2, (num_features - 3)//4) + + # conv2d: in layer3_channels, out layer3_channels, groups layer3_channels + # kernel_size 7, padding 3 + x = encoder_embed.convnext.depthwise_conv(x) + assert x.shape == (N, layer3_channels, (T - 7) // 2, (num_features - 3) // 4) + + # conv2d: in layer3_channels, out hidden_ratio * layer3_channels, kernel_size 1 + x = encoder_embed.convnext.pointwise_conv1(x) + assert x.shape == (N, layer3_channels * 3, (T - 7) // 2, (num_features - 3) // 4) + + x = encoder_embed.convnext.hidden_balancer(x) # balancer + x = encoder_embed.convnext.activation(x) # swooshL + + # conv2d: in hidden_ratio * layer3_channels, out layer3_channels, kernel 1 + x = encoder_embed.convnext.pointwise_conv2(x) + assert x.shape == (N, layer3_channels, (T - 7) // 2, (num_features - 3) // 4) + + # bypass and layer drop, omitted here. + x = encoder_embed.convnext.out_balancer(x) + + # Note: the input and output shape of ConvNeXt are the same + + x = x.transpose(1, 2).reshape(N, (T - 7) // 2, -1) + assert x.shape == (N, (T - 7) // 2, layer3_channels * ((num_features - 3) // 4)) + + x = encoder_embed.out(x) + assert x.shape == (N, (T - 7) // 2, out_channels) + + x = encoder_embed.out_whiten(x) + x = encoder_embed.out_norm(x) + # final layer is dropout + + # test streaming forward + + subsampling_factor = 2 + cached_left_padding = encoder_embed.get_init_states(batch_size=N) + depthwise_conv_kernel_size = 7 + pad_size = (depthwise_conv_kernel_size - 1) // 2 + + assert cached_left_padding.shape == ( + N, + layer3_channels, + pad_size, + (num_features - 3) // 4, + ) + + chunk_size = 16 + right_padding = pad_size * subsampling_factor + T = chunk_size * subsampling_factor + 7 + right_padding + x = torch.rand(N, T, num_features) + x_lens = torch.tensor([T] * N) + y, y_lens, next_cached_left_padding = encoder_embed.streaming_forward( + x, x_lens, cached_left_padding + ) + + assert y.shape == (N, chunk_size, out_channels), y.shape + assert next_cached_left_padding.shape == cached_left_padding.shape + + assert y.shape[1] == y_lens[0] == y_lens[1] + + +def main(): + test_conv2d_subsampling() + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/zipformer/zipformer.py b/egs/librispeech/ASR/zipformer/zipformer.py index 7d98dbeb1..b39af02b8 100644 --- a/egs/librispeech/ASR/zipformer/zipformer.py +++ b/egs/librispeech/ASR/zipformer/zipformer.py @@ -219,7 +219,7 @@ class Zipformer2(EncoderInterface): (num_frames0, batch_size, _encoder_dims0) = x.shape - assert self.encoder_dim[0] == _encoder_dims0 + assert self.encoder_dim[0] == _encoder_dims0, (self.encoder_dim[0], _encoder_dims0) feature_mask_dropout_prob = 0.125 @@ -334,7 +334,7 @@ class Zipformer2(EncoderInterface): x = self._get_full_dim_output(outputs) x = self.downsample_output(x) # class Downsample has this rounding behavior.. - assert self.output_downsampling_factor == 2 + assert self.output_downsampling_factor == 2, self.output_downsampling_factor if torch.jit.is_scripting() or torch.jit.is_tracing(): lengths = (x_lens + 1) // 2 else: diff --git a/egs/tedlium3/ASR/RESULTS.md b/egs/tedlium3/ASR/RESULTS.md index 38eaa8f44..bd8a5b43f 100644 --- a/egs/tedlium3/ASR/RESULTS.md +++ b/egs/tedlium3/ASR/RESULTS.md @@ -1,5 +1,111 @@ ## Results +### TedLium3 BPE training results (Zipformer) + +#### 2023-06-15 (Regular transducer) + +Using the codes from this PR https://github.com/k2-fsa/icefall/pull/1125. + +Number of model parameters: 65549011, i.e., 65.5 M + +The WERs are + +| | dev | test | comment | +|------------------------------------|------------|------------|------------------------------------------| +| greedy search | 6.74 | 6.16 | --epoch 50, --avg 22, --max-duration 500 | +| beam search (beam size 4) | 6.56 | 5.95 | --epoch 50, --avg 22, --max-duration 500 | +| modified beam search (beam size 4) | 6.54 | 6.00 | --epoch 50, --avg 22, --max-duration 500 | +| fast beam search (set as default) | 6.91 | 6.28 | --epoch 50, --avg 22, --max-duration 500 | + +The training command for reproducing is given below: + +``` +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./zipformer/train.py \ + --use-fp16 true \ + --world-size 4 \ + --num-epochs 50 \ + --start-epoch 0 \ + --exp-dir zipformer/exp \ + --max-duration 1000 +``` + +The tensorboard training log can be found at +https://tensorboard.dev/experiment/AKXbJha0S9aXyfmuvG4h5A/#scalars + +The decoding command is: +``` +epoch=50 +avg=22 + +## greedy search +./zipformer/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir zipformer/exp \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --max-duration 500 + +## beam search +./zipformer/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir zipformer/exp \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --max-duration 500 \ + --decoding-method beam_search \ + --beam-size 4 + +## modified beam search +./zipformer/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir zipformer/exp \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --max-duration 500 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +## fast beam search +./zipformer/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir ./zipformer/exp \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --max-duration 1500 \ + --decoding-method fast_beam_search \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +``` + +A pre-trained model and decoding logs can be found at + +#### 2023-06-26 (Modified transducer) + +``` +./zipformer/train.py \ + --use-fp16 true \ + --world-size 4 \ + --num-epochs 50 \ + --start-epoch 0 \ + --exp-dir zipformer/exp \ + --max-duration 1000 \ + --rnnt-type modified +``` + +The tensorboard training log can be found at +https://tensorboard.dev/experiment/3d4bYmbJTGiWQQaW88CVEQ/#scalars + +| | dev | test | comment | +|------------------------------------|------------|------------|------------------------------------------| +| greedy search | 6.32 | 5.83 | --epoch 50, --avg 22, --max-duration 500 | +| modified beam search (beam size 4) | 6.16 | 5.79 | --epoch 50, --avg 22, --max-duration 500 | +| fast beam search (set as default) | 6.30 | 5.89 | --epoch 50, --avg 22, --max-duration 500 | + +A pre-trained model and decoding logs can be found at . + ### TedLium3 BPE training results (Conformer-CTC 2) #### [conformer_ctc2](./conformer_ctc2) diff --git a/egs/tedlium3/ASR/zipformer/__init__.py b/egs/tedlium3/ASR/zipformer/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/tedlium3/ASR/zipformer/asr_datamodule.py b/egs/tedlium3/ASR/zipformer/asr_datamodule.py new file mode 120000 index 000000000..49b2ee483 --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/asr_datamodule.py @@ -0,0 +1 @@ +../transducer_stateless/asr_datamodule.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/beam_search.py b/egs/tedlium3/ASR/zipformer/beam_search.py new file mode 120000 index 000000000..e24eca39f --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/decode.py b/egs/tedlium3/ASR/zipformer/decode.py new file mode 100755 index 000000000..ea1cbba1b --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/decode.py @@ -0,0 +1,833 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./zipformer/decode.py \ + --epoch 30 \ + --avg 9 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 30 \ + --avg 9 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 30 \ + --avg 9 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 30 \ + --avg 9 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./zipformer/decode.py \ + --epoch 30 \ + --avg 9 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(6) fast beam search (nbest oracle WER) +./zipformer/decode.py \ + --epoch 30 \ + --avg 9 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_oracle \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(7) fast beam search (with LG) +./zipformer/decode.py \ + --epoch 30 \ + --avg 9 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import TedLiumAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + if params.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + x, x_lens = model.encoder_embed(feature, feature_lens) + + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + hyps = [] + unk = sp.decode(sp.unk_id()).strip() + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + allow_partial=True, + ) + for hyp in sp.decode(hyp_tokens): + hyp = [w for w in hyp.split() if w != unk] + hyps.append(hyp) + elif params.decoding_method == "fast_beam_search_nbest_LG": + hyp_tokens = fast_beam_search_nbest_LG( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + allow_partial=True, + ) + for hyp in hyp_tokens: + hyp = [word_table[i] for i in hyp if word_table[i] != unk] + hyps.append(hyp) + elif params.decoding_method == "fast_beam_search_nbest": + hyp_tokens = fast_beam_search_nbest( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + allow_partial=True, + ) + for hyp in sp.decode(hyp_tokens): + hyp = [w for w in hyp.split() if w != unk] + hyps.append(hyp) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + allow_partial=True, + ) + for hyp in sp.decode(hyp_tokens): + hyp = [w for w in hyp.split() if w != unk] + hyps.append(hyp) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyp = [w for w in hyp.split() if w != unk] + hyps.append(hyp) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyp = [w for w in hyp.split() if w != unk] + hyps.append(hyp) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyp = [w for w in sp.decode(hyp).split() if w != unk] + hyps.append(hyp) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + word_table=word_table, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + TedLiumAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + tedlium = TedLiumAsrDataModule(args) + + dev_cuts = tedlium.dev_cuts() + test_cuts = tedlium.test_cuts() + + dev_dl = tedlium.test_dataloaders(dev_cuts) + test_dl = tedlium.test_dataloaders(test_cuts) + + test_sets = ["dev", "test"] + test_dls = [dev_dl, test_dl] + + for name, dl in zip(test_sets, test_dls): + results_dict = decode_dataset( + dl=dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=name, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/zipformer/decoder.py b/egs/tedlium3/ASR/zipformer/decoder.py new file mode 120000 index 000000000..5a8018680 --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/encoder_interface.py b/egs/tedlium3/ASR/zipformer/encoder_interface.py new file mode 120000 index 000000000..653c5b09a --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/transducer_stateless/encoder_interface.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/export.py b/egs/tedlium3/ASR/zipformer/export.py new file mode 120000 index 000000000..dfc1bec08 --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/export.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/joiner.py b/egs/tedlium3/ASR/zipformer/joiner.py new file mode 120000 index 000000000..5b8a36332 --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/model.py b/egs/tedlium3/ASR/zipformer/model.py new file mode 100644 index 000000000..90ec7e7aa --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/model.py @@ -0,0 +1,223 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import k2 +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface + +from icefall.utils import add_sos, make_pad_mask +from scaling import ScaledLinear + + +class Transducer(nn.Module): + """It implements https://arxiv.org/pdf/1211.3711.pdf + "Sequence Transduction with Recurrent Neural Networks" + """ + + def __init__( + self, + encoder_embed: nn.Module, + encoder: EncoderInterface, + decoder: nn.Module, + joiner: nn.Module, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + ): + """ + Args: + encoder_embed: + It is a Convolutional 2D subsampling module. It converts + an input of shape (N, T, idim) to an output of of shape + (N, T', odim), where T' = (T-3)//2-2 = (T-7)//2. + encoder: + It is the transcription network in the paper. Its accepts + two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, encoder_dim) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, decoder_dim). + It should contain one attribute: `blank_id`. + joiner: + It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim). + Its output shape is (N, T, U, vocab_size). Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + assert hasattr(decoder, "blank_id") + + self.encoder_embed = encoder_embed + self.encoder = encoder + self.decoder = decoder + self.joiner = joiner + + self.simple_am_proj = ScaledLinear( + encoder_dim, + vocab_size, + initial_scale=0.25, + ) + self.simple_lm_proj = ScaledLinear( + decoder_dim, + vocab_size, + initial_scale=0.25, + ) + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + rnnt_type: str = "regular", + ) -> torch.Tensor: + """ + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + rnnt_type: + The type of label topology to use for the transducer loss. One of "regular", + "modified", or "constrained". + Returns: + Return the transducer loss. + + Note: + Regarding am_scale & lm_scale, it will make the loss-function one of + the form: + lm_scale * lm_probs + am_scale * am_probs + + (1-lm_scale-am_scale) * combined_probs + """ + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == x_lens.size(0) == y.dim0 + + # logging.info(f"Memory allocated at entry: {torch.cuda.memory_allocated() // 1000000}M") + x, x_lens = self.encoder_embed(x, x_lens) + # logging.info(f"Memory allocated after encoder_embed: {torch.cuda.memory_allocated() // 1000000}M") + + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + encoder_out, x_lens = self.encoder(x, x_lens, src_key_padding_mask) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + assert torch.all(x_lens > 0) + + # Now for the decoder, i.e., the prediction network + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + # sos_y_padded: [B, S + 1], start with SOS. + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + # decoder_out: [B, S + 1, decoder_dim] + decoder_out = self.decoder(sos_y_padded) + + # Note: y does not start with SOS + # y_padded : [B, S] + y_padded = y.pad(mode="constant", padding_value=0) + + y_padded = y_padded.to(torch.int64) + boundary = torch.zeros( + (encoder_out.size(0), 4), + dtype=torch.int64, + device=encoder_out.device, + ) + boundary[:, 2] = y_lens + boundary[:, 3] = x_lens + + lm = self.simple_lm_proj(decoder_out) + am = self.simple_am_proj(encoder_out) + + # if self.training and random.random() < 0.25: + # lm = penalize_abs_values_gt(lm, 100.0, 1.0e-04) + # if self.training and random.random() < 0.25: + # am = penalize_abs_values_gt(am, 30.0, 1.0e-04) + + with torch.cuda.amp.autocast(enabled=False): + simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed( + lm=lm.float(), + am=am.float(), + symbols=y_padded, + termination_symbol=blank_id, + lm_only_scale=lm_scale, + am_only_scale=am_scale, + boundary=boundary, + reduction="sum", + return_grad=True, + rnnt_type=rnnt_type, + ) + + # ranges : [B, T, prune_range] + ranges = k2.get_rnnt_prune_ranges( + px_grad=px_grad, + py_grad=py_grad, + boundary=boundary, + s_range=prune_range, + ) + + # am_pruned : [B, T, prune_range, encoder_dim] + # lm_pruned : [B, T, prune_range, decoder_dim] + am_pruned, lm_pruned = k2.do_rnnt_pruning( + am=self.joiner.encoder_proj(encoder_out), + lm=self.joiner.decoder_proj(decoder_out), + ranges=ranges, + ) + + # logits : [B, T, prune_range, vocab_size] + + # project_input=False since we applied the decoder's input projections + # prior to do_rnnt_pruning (this is an optimization for speed). + logits = self.joiner(am_pruned, lm_pruned, project_input=False) + + with torch.cuda.amp.autocast(enabled=False): + pruned_loss = k2.rnnt_loss_pruned( + logits=logits.float(), + symbols=y_padded, + ranges=ranges, + termination_symbol=blank_id, + boundary=boundary, + reduction="sum", + rnnt_type=rnnt_type, + ) + + return (simple_loss, pruned_loss) diff --git a/egs/tedlium3/ASR/zipformer/optim.py b/egs/tedlium3/ASR/zipformer/optim.py new file mode 120000 index 000000000..5eaa3cffd --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/pretrained.py b/egs/tedlium3/ASR/zipformer/pretrained.py new file mode 120000 index 000000000..0bd71dde4 --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/pretrained.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/profile.py b/egs/tedlium3/ASR/zipformer/profile.py new file mode 120000 index 000000000..c93adbd14 --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/profile.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/profile.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/scaling.py b/egs/tedlium3/ASR/zipformer/scaling.py new file mode 120000 index 000000000..6f398f431 --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/scaling_converter.py b/egs/tedlium3/ASR/zipformer/scaling_converter.py new file mode 120000 index 000000000..b0ecee05e --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/subsampling.py b/egs/tedlium3/ASR/zipformer/subsampling.py new file mode 120000 index 000000000..01ae9002c --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/tedlium3/ASR/zipformer/train.py b/egs/tedlium3/ASR/zipformer/train.py new file mode 100755 index 000000000..9271c8438 --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/train.py @@ -0,0 +1,1308 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Daniel Povey) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +# For non-streaming model training: +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --full-libri 1 \ + --max-duration 1000 + +# For streaming model training: +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --causal 1 \ + --full-libri 1 \ + --max-duration 1000 + +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import TedLiumAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from local.convert_transcript_words_to_bpe_ids import convert_texts_into_ids +from model import Transducer +from optim import Eden, ScaledAdam +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer2 + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " + " Must be just -1 if --causal=False", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="Maximum left-contexts for causal training, measured in frames which will " + "be converted to a number of chunks. If splitting into chunks, " + "chunk left-context frames will be chosen randomly from this list; else not relevant.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=50, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.04, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--rnnt-type", + type=str, + default="regular", + choices=["regular", "modified", "constrained"], + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=1, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder_embed=encoder_embed, + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=int(max(params.encoder_dim.split(","))), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNNT loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = convert_texts_into_ids(texts, sp) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + rnnt_type=params.rnnt_type, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + tedlium = TedLiumAsrDataModule(args) + + train_cuts = tedlium.train_cuts() + train_cuts = train_cuts.filter(lambda c: 1.0 <= c.duration <= 20.0) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = tedlium.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = tedlium.dev_cuts() + valid_dl = tedlium.valid_dataloaders(valid_cuts) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + TedLiumAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/zipformer/zipformer.py b/egs/tedlium3/ASR/zipformer/zipformer.py new file mode 120000 index 000000000..23011dda7 --- /dev/null +++ b/egs/tedlium3/ASR/zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/RESULTS.md b/egs/wenetspeech/ASR/RESULTS.md index 658ad4a9b..1a0e0681f 100644 --- a/egs/wenetspeech/ASR/RESULTS.md +++ b/egs/wenetspeech/ASR/RESULTS.md @@ -1,5 +1,90 @@ ## Results +### WenetSpeech char-based training results (Non-streaming and streaming) on zipformer model + +This is the [pull request](https://github.com/k2-fsa/icefall/pull/1130) in icefall. + +#### Non-streaming + +Best results (num of params : ~76M): + +Type | Greedy(dev & net & meeting) | Beam search(dev & net & meeting) |   +-- | -- | -- | -- +Non-streaming | 7.36 & 7.65 & 12.43 | 7.32 & 7.61 & 12.35 | --epoch=12 + +The training command: + +``` +./zipformer/train.py \ + --world-size 6 \ + --num-epochs 12 \ + --use-fp16 1 \ + --max-duration 450 \ + --training-subset L \ + --lr-epochs 1.5 \ + --context-size 2 \ + --exp-dir zipformer/exp_L_context_2 \ + --causal 0 \ + --num-workers 8 +``` + +Listed best results for each epoch below: + +Epoch | Greedy search(dev & net & meeting) | Modified beam search(dev & net & meeting) |   +-- | -- | -- | -- +4 | 7.83 & 8.86 &13.73 | 7.75 & 8.81 & 13.67 | avg=1;blank-penalty=2 +5 | 7.75 & 8.46 & 13.38 | 7.68 & 8.41 & 13.27 | avg=1;blank-penalty=2 +6 | 7.72 & 8.19 & 13.16 | 7.62 & 8.14 & 13.06 | avg=1;blank-penalty=2 +7 | 7.59 & 8.08 & 12.97 | 7.53 & 8.01 & 12.87 | avg=2;blank-penalty=2 +8 | 7.68 & 7.87 & 12.96 | 7.61 & 7.81 & 12.88 | avg=1;blank-penalty=2 +9 | 7.57 & 7.77 & 12.87 | 7.5 & 7.71 & 12.77 | avg=1;blank-penalty=2 +10 | 7.45 & 7.7 & 12.69 | 7.39 & 7.63 & 12.59 | avg=2;blank-penalty=2 +11 | 7.35 & 7.67 & 12.46 | 7.31 & 7.63 & 12.43 | avg=3;blank-penalty=2 +12 | 7.36 & 7.65 & 12.43 | 7.32 & 7.61 & 12.35 | avg=4;blank-penalty=2 + +The pre-trained model is available here : https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615 + + +#### Streaming + +Best results (num of params : ~76M): + +Type | Greedy(dev & net & meeting) | Beam search(dev & net & meeting) |   +-- | -- | -- | -- +Streaming | 8.45 & 9.89 & 16.46 | 8.21 & 9.77 & 16.07 | --epoch=12; --chunk-size=16; --left-context-frames=256 +Streaming | 8.0 & 9.0 & 15.11 | 7.84 & 8.94 & 14.92 | --epoch=12; --chunk-size=32; --left-context-frames=256 + +The training command: + +``` +./zipformer/train.py \ + --world-size 8 \ + --num-epochs 12 \ + --use-fp16 1 \ + --max-duration 450 \ + --training-subset L \ + --lr-epochs 1.5 \ + --context-size 2 \ + --exp-dir zipformer/exp_L_causal_context_2 \ + --causal 1 \ + --num-workers 8 +``` + +Best results for each epoch (--chunk-size=16; --left-context-frames=128) + +Epoch | Greedy search(dev & net & meeting) | Modified beam search(dev & net & meeting) |   +-- | -- | -- | -- +6 | 9.14 & 10.75 & 18.15 | 8.79 & 10.54 & 17.64 | avg=1;blank-penalty=1.5 +7 | 9.11 & 10.61 & 17.86 | 8.8 & 10.42 & 17.29 | avg=1;blank-penalty=1.5 +8 | 8.89 & 10.32 & 17.44 | 8.59 & 10.09 & 16.9 | avg=1;blank-penalty=1.5 +9 | 8.86 & 10.11 & 17.35 | 8.55 & 9.87 & 16.76 | avg=1;blank-penalty=1.5 +10 | 8.66 & 10.0 & 16.94 | 8.39 & 9.83 & 16.47 | avg=2;blank-penalty=1.5 +11 | 8.58 & 9.92 & 16.67 | 8.32 & 9.77 & 16.27 | avg=3;blank-penalty=1.5 +12 | 8.45 & 9.89 & 16.46 | 8.21 & 9.77 & 16.07 | avg=4;blank-penalty=1.5 + +The pre-trained model is available here: https://huggingface.co/pkufool/icefall-asr-zipformer-streaming-wenetspeech-20230615 + + ### WenetSpeech char-based training results (offline and streaming) (Pruned Transducer 5) #### 2022-07-22 diff --git a/egs/wenetspeech/ASR/local/prepare_char_lm_training_data.py b/egs/wenetspeech/ASR/local/prepare_char_lm_training_data.py new file mode 120000 index 000000000..2374cafdd --- /dev/null +++ b/egs/wenetspeech/ASR/local/prepare_char_lm_training_data.py @@ -0,0 +1 @@ +../../../aishell/ASR/local/prepare_char_lm_training_data.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/local/sort_lm_training_data.py b/egs/wenetspeech/ASR/local/sort_lm_training_data.py new file mode 120000 index 000000000..efef2c445 --- /dev/null +++ b/egs/wenetspeech/ASR/local/sort_lm_training_data.py @@ -0,0 +1 @@ +../../../aishell/ASR/local/sort_lm_training_data.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py index 7cb2e1048..746b212ff 100644 --- a/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py +++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py @@ -292,7 +292,7 @@ class WenetSpeechAsrDataModule: max_duration=self.args.max_duration, shuffle=self.args.shuffle, num_buckets=self.args.num_buckets, - buffer_size=30000, + buffer_size=300000, drop_last=True, ) else: diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py index dc431578c..36b8a4b67 100755 --- a/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py +++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py @@ -588,7 +588,7 @@ def decode_dataset( results = defaultdict(list) for batch_idx, batch in enumerate(dl): texts = batch["supervisions"]["text"] - texts = [list(str(text)) for text in texts] + texts = [list("".join(text.split())) for text in texts] cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] hyps_dict = decode_one_batch( diff --git a/egs/wenetspeech/ASR/zipformer/__init__.py b/egs/wenetspeech/ASR/zipformer/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/wenetspeech/ASR/zipformer/asr_datamodule.py b/egs/wenetspeech/ASR/zipformer/asr_datamodule.py new file mode 120000 index 000000000..a074d6085 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/asr_datamodule.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/asr_datamodule.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/beam_search.py b/egs/wenetspeech/ASR/zipformer/beam_search.py new file mode 120000 index 000000000..8554e44cc --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/decode.py b/egs/wenetspeech/ASR/zipformer/decode.py new file mode 100755 index 000000000..0fbc8244b --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/decode.py @@ -0,0 +1,818 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao +# Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./zipformer/decode.py \ + --epoch 35 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --lang-dir data/lang_char \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) modified beam search +./zipformer/decode.py \ + --epoch 35 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --lang-dir data/lang_char \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(3) fast beam search (trivial_graph) +./zipformer/decode.py \ + --epoch 35 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --lang-dir data/lang_char \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(4) fast beam search (LG) +./zipformer/decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --lang-dir data/lang_char \ + --max-duration 600 \ + --decoding-method fast_beam_search_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest oracle WER) +./zipformer/decode.py \ + --epoch 35 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --lang-dir data/lang_char \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_oracle \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import torch +import torch.nn as nn +from asr_datamodule import WenetSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from lhotse.cut import Cut +from train import add_model_arguments, get_model, get_params + +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_char", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_LG + - fast_beam_search_nbest_oracle + If you use fast_beam_search_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search, fast_beam_search_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--ilme-scale", + type=float, + default=0.2, + help=""" + Used only when --decoding_method is fast_beam_search_LG. + It specifies the scale for the internal language model estimation. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search, fast_beam_search_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search, fast_beam_search_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--blank-penalty", + type=float, + default=0.0, + help=""" + The penalty applied on blank symbol during decoding. + Note: It is a positive value that would be applied to logits like + this `logits[:, 0] -= blank_penalty` (suppose logits.shape is + [batch_size, vocab] and blank id is 0). + """, + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + graph_compiler: CharCtcTrainingGraphCompiler, + batch: dict, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or LG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + if params.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + x, x_lens = model.encoder_embed(feature, feature_lens) + + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + blank_penalty=params.blank_penalty, + ) + for i in range(encoder_out.size(0)): + hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + elif params.decoding_method == "fast_beam_search_LG": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + blank_penalty=params.blank_penalty, + ilme_scale=params.ilme_scale, + ) + for hyp in hyp_tokens: + sentence = "".join([lexicon.word_table[i] for i in hyp]) + hyps.append(list(sentence)) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=graph_compiler.texts_to_ids(supervisions["text"]), + nbest_scale=params.nbest_scale, + blank_penalty=params.blank_penalty, + ) + for i in range(encoder_out.size(0)): + hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + blank_penalty=params.blank_penalty, + ) + for i in range(encoder_out.size(0)): + hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + blank_penalty=params.blank_penalty, + beam=params.beam_size, + ) + for i in range(encoder_out.size(0)): + hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + blank_penalty=params.blank_penalty, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + blank_penalty=params.blank_penalty, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append([lexicon.token_table[idx] for idx in hyp]) + + key = f"blank_penalty_{params.blank_penalty}" + if params.decoding_method == "greedy_search": + return {"greedy_search_" + key: hyps} + elif "fast_beam_search" in params.decoding_method: + key += f"_beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ilme_scale_{params.ilme_scale}" + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + else: + return {f"beam_size_{params.beam_size}_" + key: hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + graph_compiler: CharCtcTrainingGraphCompiler, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or LG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + texts = [list("".join(text.split())) for text in texts] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + lexicon=lexicon, + graph_compiler=graph_compiler, + decoding_graph=decoding_graph, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + this_batch.append((cut_id, ref_text, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + WenetSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "modified_beam_search", + "fast_beam_search", + "fast_beam_search_LG", + "fast_beam_search_nbest_oracle", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"_ilme_scale_{params.ilme_scale}" + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + params.suffix += f"-blank-penalty-{params.blank_penalty}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + lexicon = Lexicon(params.lang_dir) + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + graph_compiler = CharCtcTrainingGraphCompiler( + lexicon=lexicon, + device=device, + ) + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if "fast_beam_search" in params.decoding_method: + if "LG" in params.decoding_method: + lexicon = Lexicon(params.lang_dir) + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + wenetspeech = WenetSpeechAsrDataModule(args) + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}." + ) + return T > 0 + + dev_cuts = wenetspeech.valid_cuts() + dev_cuts = dev_cuts.filter(remove_short_utt) + dev_dl = wenetspeech.valid_dataloaders(dev_cuts) + + test_net_cuts = wenetspeech.test_net_cuts() + test_net_cuts = test_net_cuts.filter(remove_short_utt) + test_net_dl = wenetspeech.test_dataloaders(test_net_cuts) + + test_meeting_cuts = wenetspeech.test_meeting_cuts() + test_meeting_cuts = test_meeting_cuts.filter(remove_short_utt) + test_meeting_dl = wenetspeech.test_dataloaders(test_meeting_cuts) + + test_sets = ["DEV", "TEST_NET", "TEST_MEETING"] + test_dls = [dev_dl, test_net_dl, test_meeting_dl] + + for test_set, test_dl in zip(test_sets, test_dls): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + lexicon=lexicon, + graph_compiler=graph_compiler, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/wenetspeech/ASR/zipformer/decode_stream.py b/egs/wenetspeech/ASR/zipformer/decode_stream.py new file mode 120000 index 000000000..b8d8ddfc4 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/decode_stream.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decode_stream.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/decoder.py b/egs/wenetspeech/ASR/zipformer/decoder.py new file mode 120000 index 000000000..5a8018680 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/encoder_interface.py b/egs/wenetspeech/ASR/zipformer/encoder_interface.py new file mode 120000 index 000000000..b9aa0ae08 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/encoder_interface.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/export-onnx-streaming.py b/egs/wenetspeech/ASR/zipformer/export-onnx-streaming.py new file mode 120000 index 000000000..2962eb784 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/export-onnx-streaming.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export-onnx-streaming.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/export-onnx.py b/egs/wenetspeech/ASR/zipformer/export-onnx.py new file mode 120000 index 000000000..70a15683c --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/export-onnx.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export-onnx.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/export.py b/egs/wenetspeech/ASR/zipformer/export.py new file mode 120000 index 000000000..dfc1bec08 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/export.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/jit_pretrained.py b/egs/wenetspeech/ASR/zipformer/jit_pretrained.py new file mode 120000 index 000000000..25108391f --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/jit_pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/jit_pretrained.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/jit_pretrained_streaming.py b/egs/wenetspeech/ASR/zipformer/jit_pretrained_streaming.py new file mode 120000 index 000000000..1962351e9 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/jit_pretrained_streaming.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/joiner.py b/egs/wenetspeech/ASR/zipformer/joiner.py new file mode 120000 index 000000000..5b8a36332 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/model.py b/egs/wenetspeech/ASR/zipformer/model.py new file mode 120000 index 000000000..cd7e07d72 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/onnx_check.py b/egs/wenetspeech/ASR/zipformer/onnx_check.py new file mode 120000 index 000000000..f3dd42004 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/onnx_check.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_check.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/onnx_decode.py b/egs/wenetspeech/ASR/zipformer/onnx_decode.py new file mode 100755 index 000000000..ed5f6db08 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/onnx_decode.py @@ -0,0 +1,334 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Xiaoyu Yang, +# Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This script loads ONNX exported models and uses them to decode the test sets. + +We use the pre-trained model from +https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/wenetspeech/ASR + +repo_url=https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "data/lang_char/tokens.txt" +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-9999.pt +popd + +2. Export the model to ONNX + +./zipformer/export-onnx.py \ + --tokens $repo/data/lang_char/tokens.txt \ + --epoch 9999 \ + --avg 1 \ + --exp-dir $repo/exp/ + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-9999-avg-1.onnx + - decoder-epoch-9999-avg-1.onnx + - joiner-epoch-9999-avg-1.onnx + +2. Run this file + +./zipformer/onnx_decode.py \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \ + --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \ + --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \ +""" + + +import argparse +import logging +import time +from pathlib import Path +from typing import List, Tuple + +import k2 +import torch +import torch.nn as nn +from asr_datamodule import WenetSpeechAsrDataModule +from lhotse.cut import Cut +from onnx_pretrained import OnnxModel, greedy_search + +from icefall.utils import setup_logger, store_transcripts, write_error_stats + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--encoder-model-filename", + type=str, + required=True, + help="Path to the encoder onnx model. ", + ) + + parser.add_argument( + "--decoder-model-filename", + type=str, + required=True, + help="Path to the decoder onnx model. ", + ) + + parser.add_argument( + "--joiner-model-filename", + type=str, + required=True, + help="Path to the joiner onnx model. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--tokens", + type=str, + default="data/lang_char/tokens.txt", + help="Path to the tokens.txt", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="Valid values are greedy_search and modified_beam_search", + ) + + return parser + + +def decode_one_batch( + model: OnnxModel, token_table: k2.SymbolTable, batch: dict +) -> List[List[str]]: + """Decode one batch and return the result. + Currently it only greedy_search is supported. + + Args: + model: + The neural model. + token_table: + Mapping ids to tokens. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + + Returns: + Return the decoded results for each utterance. + """ + feature = batch["inputs"] + assert feature.ndim == 3 + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(dtype=torch.int64) + + encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens) + + hyps = greedy_search( + model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens + ) + + hyps = [[token_table[h] for h in hyp] for hyp in hyps] + return hyps + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + model: nn.Module, + token_table: k2.SymbolTable, +) -> Tuple[List[Tuple[str, List[str], List[str]]], float]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + model: + The neural model. + token_table: + Mapping ids to tokens. + + Returns: + - A list of tuples. Each tuple contains three elements: + - cut_id, + - reference transcript, + - predicted result. + - The total duration (in seconds) of the dataset. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + log_interval = 10 + total_duration = 0 + + results = [] + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]]) + + hyps = decode_one_batch(model=model, token_table=token_table, batch=batch) + + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = list(ref_text) + this_batch.append((cut_id, ref_words, hyp_words)) + + results.extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + + return results, total_duration + + +def save_results( + res_dir: Path, + test_set_name: str, + results: List[Tuple[str, List[str], List[str]]], +): + recog_path = res_dir / f"recogs-{test_set_name}.txt" + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = res_dir / f"errs-{test_set_name}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True) + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + errs_info = res_dir / f"wer-summary-{test_set_name}.txt" + with open(errs_info, "w") as f: + print("WER", file=f) + print(wer, file=f) + + s = "\nFor {}, WER is {}:\n".format(test_set_name, wer) + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + WenetSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + + assert ( + args.decoding_method == "greedy_search" + ), "Only supports greedy_search currently." + res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}" + + setup_logger(f"{res_dir}/log-decode") + logging.info("Decoding started") + + device = torch.device("cpu") + logging.info(f"Device: {device}") + + token_table = k2.SymbolTable.from_file(args.tokens) + assert token_table[0] == "" + + logging.info(vars(args)) + + logging.info("About to create model") + model = OnnxModel( + encoder_model_filename=args.encoder_model_filename, + decoder_model_filename=args.decoder_model_filename, + joiner_model_filename=args.joiner_model_filename, + ) + + # we need cut ids to display recognition results. + args.return_cuts = True + + wenetspeech = WenetSpeechAsrDataModule(args) + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}." + ) + return T > 0 + + dev_cuts = wenetspeech.valid_cuts() + dev_cuts = dev_cuts.filter(remove_short_utt) + dev_dl = wenetspeech.valid_dataloaders(dev_cuts) + + test_net_cuts = wenetspeech.test_net_cuts() + test_net_cuts = test_net_cuts.filter(remove_short_utt) + test_net_dl = wenetspeech.test_dataloaders(test_net_cuts) + + test_meeting_cuts = wenetspeech.test_meeting_cuts() + test_meeting_cuts = test_meeting_cuts.filter(remove_short_utt) + test_meeting_dl = wenetspeech.test_dataloaders(test_meeting_cuts) + + test_sets = ["DEV", "TEST_NET", "TEST_MEETING"] + test_dl = [dev_dl, test_net_dl, test_meeting_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + start_time = time.time() + results, total_duration = decode_dataset( + dl=test_dl, model=model, token_table=token_table + ) + end_time = time.time() + elapsed_seconds = end_time - start_time + rtf = elapsed_seconds / total_duration + + logging.info(f"Elapsed time: {elapsed_seconds:.3f} s") + logging.info(f"Wave duration: {total_duration:.3f} s") + logging.info( + f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}" + ) + + save_results(res_dir=res_dir, test_set_name=test_set, results=results) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/wenetspeech/ASR/zipformer/onnx_pretrained-streaming.py b/egs/wenetspeech/ASR/zipformer/onnx_pretrained-streaming.py new file mode 120000 index 000000000..cfea104c2 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/onnx_pretrained-streaming.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/onnx_pretrained.py b/egs/wenetspeech/ASR/zipformer/onnx_pretrained.py new file mode 120000 index 000000000..8f32f4ee7 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/onnx_pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_pretrained.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/optim.py b/egs/wenetspeech/ASR/zipformer/optim.py new file mode 120000 index 000000000..5eaa3cffd --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/pretrained.py b/egs/wenetspeech/ASR/zipformer/pretrained.py new file mode 120000 index 000000000..0bd71dde4 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/pretrained.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/scaling.py b/egs/wenetspeech/ASR/zipformer/scaling.py new file mode 120000 index 000000000..6f398f431 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/scaling_converter.py b/egs/wenetspeech/ASR/zipformer/scaling_converter.py new file mode 120000 index 000000000..b0ecee05e --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/streaming_beam_search.py b/egs/wenetspeech/ASR/zipformer/streaming_beam_search.py new file mode 120000 index 000000000..b1ed54557 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/streaming_beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/streaming_beam_search.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/streaming_decode.py b/egs/wenetspeech/ASR/zipformer/streaming_decode.py new file mode 100755 index 000000000..94c5fae5f --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/streaming_decode.py @@ -0,0 +1,881 @@ +#!/usr/bin/env python3 +# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang, +# Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +Usage: +./zipformer/streaming_decode.py \ + --epoch 28 \ + --avg 15 \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 256 \ + --exp-dir ./zipformer/exp \ + --decoding-method greedy_search \ + --num-decode-streams 2000 +""" + +import argparse +import logging +import math +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import numpy as np +import torch +from asr_datamodule import WenetSpeechAsrDataModule +from decode_stream import DecodeStream +from kaldifeat import Fbank, FbankOptions +from lhotse import CutSet +from streaming_beam_search import ( + fast_beam_search_one_best, + greedy_search, + modified_beam_search, +) +from torch import Tensor, nn +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_model, get_params + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="Path to the lang dir(containing lexicon, tokens, etc.)", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Supported decoding methods are: + greedy_search + modified_beam_search + fast_beam_search + """, + ) + + parser.add_argument( + "--num_active_paths", + type=int, + default=4, + help="""An interger indicating how many candidates we will keep for each + frame. Used only when --decoding-method is modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=32, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--blank-penalty", + type=float, + default=0.0, + help=""" + The penalty applied on blank symbol during decoding. + Note: It is a positive value that would be applied to logits like + this `logits[:, 0] -= blank_penalty` (suppose logits.shape is + [batch_size, vocab] and blank id is 0). + """, + ) + + parser.add_argument( + "--num-decode-streams", + type=int, + default=2000, + help="The number of streams that can be decoded parallel.", + ) + + add_model_arguments(parser) + + return parser + + +def get_init_states( + model: nn.Module, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), +) -> List[torch.Tensor]: + """ + Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + states[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + """ + states = model.encoder.get_init_states(batch_size, device) + + embed_states = model.encoder_embed.get_init_states(batch_size, device) + states.append(embed_states) + + processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device) + states.append(processed_lens) + + return states + + +def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]: + """Stack list of zipformer states that correspond to separate utterances + into a single emformer state, so that it can be used as an input for + zipformer when those utterances are formed into a batch. + + Args: + state_list: + Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. For element-n, + state_list[n] is a list of cached tensors of all encoder layers. For layer-i, + state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, + cached_val2, cached_conv1, cached_conv2). + state_list[n][-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + state_list[n][-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + + Note: + It is the inverse of :func:`unstack_states`. + """ + batch_size = len(state_list) + assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0]) + tot_num_layers = (len(state_list[0]) - 2) // 6 + + batch_states = [] + for layer in range(tot_num_layers): + layer_offset = layer * 6 + # cached_key: (left_context_len, batch_size, key_dim) + cached_key = torch.cat( + [state_list[i][layer_offset] for i in range(batch_size)], dim=1 + ) + # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim) + cached_nonlin_attn = torch.cat( + [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1 + ) + # cached_val1: (left_context_len, batch_size, value_dim) + cached_val1 = torch.cat( + [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1 + ) + # cached_val2: (left_context_len, batch_size, value_dim) + cached_val2 = torch.cat( + [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1 + ) + # cached_conv1: (#batch, channels, left_pad) + cached_conv1 = torch.cat( + [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0 + ) + # cached_conv2: (#batch, channels, left_pad) + cached_conv2 = torch.cat( + [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0 + ) + batch_states += [ + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ] + + cached_embed_left_pad = torch.cat( + [state_list[i][-2] for i in range(batch_size)], dim=0 + ) + batch_states.append(cached_embed_left_pad) + + processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0) + batch_states.append(processed_lens) + + return batch_states + + +def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]: + """Unstack the zipformer state corresponding to a batch of utterances + into a list of states, where the i-th entry is the state from the i-th + utterance in the batch. + + Note: + It is the inverse of :func:`stack_states`. + + Args: + batch_states: A list of cached tensors of all encoder layers. For layer-i, + states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, + cached_conv1, cached_conv2). + state_list[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + + Returns: + state_list: A list of list. Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. + """ + assert (len(batch_states) - 2) % 6 == 0, len(batch_states) + tot_num_layers = (len(batch_states) - 2) // 6 + + processed_lens = batch_states[-1] + batch_size = processed_lens.shape[0] + + state_list = [[] for _ in range(batch_size)] + + for layer in range(tot_num_layers): + layer_offset = layer * 6 + # cached_key: (left_context_len, batch_size, key_dim) + cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1) + # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim) + cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk( + chunks=batch_size, dim=1 + ) + # cached_val1: (left_context_len, batch_size, value_dim) + cached_val1_list = batch_states[layer_offset + 2].chunk( + chunks=batch_size, dim=1 + ) + # cached_val2: (left_context_len, batch_size, value_dim) + cached_val2_list = batch_states[layer_offset + 3].chunk( + chunks=batch_size, dim=1 + ) + # cached_conv1: (#batch, channels, left_pad) + cached_conv1_list = batch_states[layer_offset + 4].chunk( + chunks=batch_size, dim=0 + ) + # cached_conv2: (#batch, channels, left_pad) + cached_conv2_list = batch_states[layer_offset + 5].chunk( + chunks=batch_size, dim=0 + ) + for i in range(batch_size): + state_list[i] += [ + cached_key_list[i], + cached_nonlin_attn_list[i], + cached_val1_list[i], + cached_val2_list[i], + cached_conv1_list[i], + cached_conv2_list[i], + ] + + cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0) + for i in range(batch_size): + state_list[i].append(cached_embed_left_pad_list[i]) + + processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0) + for i in range(batch_size): + state_list[i].append(processed_lens_list[i]) + + return state_list + + +def streaming_forward( + features: Tensor, + feature_lens: Tensor, + model: nn.Module, + states: List[Tensor], + chunk_size: int, + left_context_len: int, +) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Returns encoder outputs, output lengths, and updated states. + """ + cached_embed_left_pad = states[-2] + (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward( + x=features, + x_lens=feature_lens, + cached_left_pad=cached_embed_left_pad, + ) + assert x.size(1) == chunk_size, (x.size(1), chunk_size) + + src_key_padding_mask = make_pad_mask(x_lens) + + # processed_mask is used to mask out initial states + processed_mask = torch.arange(left_context_len, device=x.device).expand( + x.size(0), left_context_len + ) + processed_lens = states[-1] # (batch,) + # (batch, left_context_size) + processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1) + # Update processed lengths + new_processed_lens = processed_lens + x_lens + + # (batch, left_context_size + chunk_size) + src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1) + + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + encoder_states = states[:-2] + ( + encoder_out, + encoder_out_lens, + new_encoder_states, + ) = model.encoder.streaming_forward( + x=x, + x_lens=x_lens, + states=encoder_states, + src_key_padding_mask=src_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + new_states = new_encoder_states + [ + new_cached_embed_left_pad, + new_processed_lens, + ] + return encoder_out, encoder_out_lens, new_states + + +def decode_one_chunk( + params: AttributeDict, + model: nn.Module, + decode_streams: List[DecodeStream], +) -> List[int]: + """Decode one chunk frames of features for each decode_streams and + return the indexes of finished streams in a List. + + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + decode_streams: + A List of DecodeStream, each belonging to a utterance. + Returns: + Return a List containing which DecodeStreams are finished. + """ + device = model.device + chunk_size = int(params.chunk_size) + left_context_len = int(params.left_context_frames) + + features = [] + feature_lens = [] + states = [] + processed_lens = [] # Used in fast-beam-search + + for stream in decode_streams: + feat, feat_len = stream.get_feature_frames(chunk_size * 2) + features.append(feat) + feature_lens.append(feat_len) + states.append(stream.states) + processed_lens.append(stream.done_frames) + + feature_lens = torch.tensor(feature_lens, device=device) + features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS) + + # Make sure the length after encoder_embed is at least 1. + # The encoder_embed subsample features (T - 7) // 2 + # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling + tail_length = chunk_size * 2 + 7 + 2 * 3 + if features.size(1) < tail_length: + pad_length = tail_length - features.size(1) + feature_lens += pad_length + features = torch.nn.functional.pad( + features, + (0, 0, 0, pad_length), + mode="constant", + value=LOG_EPS, + ) + + states = stack_states(states) + + encoder_out, encoder_out_lens, new_states = streaming_forward( + features=features, + feature_lens=feature_lens, + model=model, + states=states, + chunk_size=chunk_size, + left_context_len=left_context_len, + ) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + if params.decoding_method == "greedy_search": + greedy_search( + model=model, + encoder_out=encoder_out, + streams=decode_streams, + blank_penalty=params.blank_penalty, + ) + elif params.decoding_method == "fast_beam_search": + processed_lens = torch.tensor(processed_lens, device=device) + processed_lens = processed_lens + encoder_out_lens + fast_beam_search_one_best( + model=model, + encoder_out=encoder_out, + processed_lens=processed_lens, + streams=decode_streams, + beam=params.beam, + max_states=params.max_states, + max_contexts=params.max_contexts, + blank_penalty=params.blank_penalty, + ) + elif params.decoding_method == "modified_beam_search": + modified_beam_search( + model=model, + streams=decode_streams, + encoder_out=encoder_out, + num_active_paths=params.num_active_paths, + blank_penalty=params.blank_penalty, + ) + else: + raise ValueError(f"Unsupported decoding method: {params.decoding_method}") + + states = unstack_states(new_states) + + finished_streams = [] + for i in range(len(decode_streams)): + decode_streams[i].states = states[i] + decode_streams[i].done_frames += encoder_out_lens[i] + if decode_streams[i].done: + finished_streams.append(i) + + return finished_streams + + +def decode_dataset( + cuts: CutSet, + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + cuts: + Lhotse Cutset containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + lexicon: + The Lexicon. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + device = model.device + + opts = FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = 16000 + opts.mel_opts.num_bins = 80 + + log_interval = 100 + + decode_results = [] + # Contain decode streams currently running. + decode_streams = [] + for num, cut in enumerate(cuts): + # each utterance has a DecodeStream. + initial_states = get_init_states(model=model, batch_size=1, device=device) + decode_stream = DecodeStream( + params=params, + cut_id=cut.id, + initial_states=initial_states, + decoding_graph=decoding_graph, + device=device, + ) + + audio: np.ndarray = cut.load_audio() + # audio.shape: (1, num_samples) + assert len(audio.shape) == 2 + assert audio.shape[0] == 1, "Should be single channel" + assert audio.dtype == np.float32, audio.dtype + + # The trained model is using normalized samples + if audio.max() > 1: + logging.warning( + f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}." + f"Clipping to [-1, 1]." + ) + audio = np.clip(audio, -1, 1) + + samples = torch.from_numpy(audio).squeeze(0) + + fbank = Fbank(opts) + feature = fbank(samples.to(device)) + decode_stream.set_features(feature, tail_pad_len=30) + decode_stream.ground_truth = cut.supervisions[0].text + + decode_streams.append(decode_stream) + + while len(decode_streams) >= params.num_decode_streams: + finished_streams = decode_one_chunk( + params=params, model=model, decode_streams=decode_streams + ) + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + decode_streams[i].id, + list(decode_streams[i].ground_truth.strip()), + [ + lexicon.token_table[idx] + for idx in decode_streams[i].decoding_result() + ], + ) + ) + del decode_streams[i] + + if num % log_interval == 0: + logging.info(f"Cuts processed until now is {num}.") + + # decode final chunks of last sequences + while len(decode_streams): + finished_streams = decode_one_chunk( + params=params, model=model, decode_streams=decode_streams + ) + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + decode_streams[i].id, + decode_streams[i].ground_truth.split(), + [ + lexicon.token_table[idx] + for idx in decode_streams[i].decoding_result() + ], + ) + ) + del decode_streams[i] + + key = f"blank_penalty_{params.blank_penalty}" + if params.decoding_method == "greedy_search": + key = f"greedy_search_{key}" + elif params.decoding_method == "fast_beam_search": + key = ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}_{key}" + ) + elif params.decoding_method == "modified_beam_search": + key = f"num_active_paths_{params.num_active_paths}_{key}" + else: + raise ValueError(f"Unsupported decoding method: {params.decoding_method}") + return {key: decode_results} + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + WenetSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + params.res_dir = params.exp_dir / "streaming" / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + assert params.causal, params.causal + assert "," not in params.chunk_size, "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + params.suffix += f"-blank-penalty-{params.blank_penalty}" + + # for fast_beam_search + if params.decoding_method == "fast_beam_search": + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + lexicon = Lexicon(params.lang_dir) + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if start >= 0: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + model.device = device + + decoding_graph = None + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + wenetspeech = WenetSpeechAsrDataModule(args) + + dev_cuts = wenetspeech.valid_cuts() + test_net_cuts = wenetspeech.test_net_cuts() + test_meeting_cuts = wenetspeech.test_meeting_cuts() + + test_sets = ["DEV", "TEST_NET", "TEST_MEETING"] + test_cuts = [dev_cuts, test_net_cuts, test_meeting_cuts] + + for test_set, test_cut in zip(test_sets, test_cuts): + results_dict = decode_dataset( + cuts=test_cut, + params=params, + model=model, + lexicon=lexicon, + decoding_graph=decoding_graph, + ) + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/wenetspeech/ASR/zipformer/subsampling.py b/egs/wenetspeech/ASR/zipformer/subsampling.py new file mode 120000 index 000000000..01ae9002c --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/train.py b/egs/wenetspeech/ASR/zipformer/train.py new file mode 100755 index 000000000..83dbfa22f --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/train.py @@ -0,0 +1,1350 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Daniel Povey) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./zipformer/train.py \ + --world-size 8 \ + --num-epochs 12 \ + --start-epoch 1 \ + --exp-dir zipformer/exp \ + --training-subset L + --lr-epochs 1.5 \ + --max-duration 350 + +# For mix precision training: + +./zipformer/train.py \ + --world-size 8 \ + --num-epochs 12 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --training-subset L \ + --lr-epochs 1.5 \ + --max-duration 750 + +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import WenetSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from optim import Eden, ScaledAdam +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer2 + +from icefall import diagnostics +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="""Feedforward dimension of the zipformer encoder layers, per stack, comma separated.""", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="""Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="""Embedding dimension in encoder stacks: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="""Query/key dimension per head in encoder stacks: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="""Value dimension per head in encoder stacks: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="""Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="""Unmasked dimensions in the encoders, relates to augmentation during training. A single int or comma-separated list. Must be <= each corresponding encoder_dim.""", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="""Sizes of convolutional kernels in convolution modules in each encoder stack: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="""Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. Must be just -1 if --causal=False""", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="""Maximum left-contexts for causal training, measured in frames which will + be converted to a number of chunks. If splitting into chunks, + chunk left-context frames will be chosen randomly from this list; else not relevant.""", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="""Reference batch duration for purposes of adjusting batch counts for setting various schedules inside the model""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="""The context size in the decoder. 1 means bigram; 2 means tri-gram""", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="""The prune range for rnnt loss, it means how many symbols(context) + we are using to compute the loss""", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="""The scale to smooth the loss with lm + (output of prediction network) part.""", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="""The scale to smooth the loss with am (output of encoder network) part.""", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="""To get pruning ranges, we will calculate a simple version + loss(joiner is just addition), this simple loss also uses for + training (as a regularization item). We will scale the simple loss + with this parameter before adding to the final loss.""", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = AsrModel( + encoder_embed=encoder_embed, + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=int(max(params.encoder_dim.split(","))), + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + graph_compiler: CharCtcTrainingGraphCompiler, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute CTC loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = graph_compiler.texts_to_ids(texts) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, _ = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + graph_compiler: CharCtcTrainingGraphCompiler, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + graph_compiler: CharCtcTrainingGraphCompiler, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, graph_compiler=graph_compiler) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + lexicon = Lexicon(params.lang_dir) + graph_compiler = CharCtcTrainingGraphCompiler( + lexicon=lexicon, + device=device, + ) + + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + wenetspeech = WenetSpeechAsrDataModule(args) + + train_cuts = wenetspeech.train_cuts() + valid_cuts = wenetspeech.valid_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 15 seconds + # + # Caution: There is a reason to select 15.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 15.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_frames - 7) // 2 + 1) // 2 + tokens = graph_compiler.texts_to_ids([c.supervisions[0].text])[0] + + if T < len(tokens): + logging.warning( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_frames}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].text}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = wenetspeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_dl = wenetspeech.valid_dataloaders(valid_cuts) + + if False and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + graph_compiler=graph_compiler, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + graph_compiler=graph_compiler, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + graph_compiler: CharCtcTrainingGraphCompiler, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + graph_compiler: + The compiler to encode texts to ids. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + texts = supervisions["text"] + y = graph_compiler.texts_to_ids(texts) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + graph_compiler: CharCtcTrainingGraphCompiler, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, graph_compiler=graph_compiler) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + WenetSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.lang_dir = Path(args.lang_dir) + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/wenetspeech/ASR/zipformer/zipformer.py b/egs/wenetspeech/ASR/zipformer/zipformer.py new file mode 120000 index 000000000..23011dda7 --- /dev/null +++ b/egs/wenetspeech/ASR/zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file diff --git a/icefall/rnn_lm/train.py b/icefall/rnn_lm/train.py index 0f0887859..3d206d139 100755 --- a/icefall/rnn_lm/train.py +++ b/icefall/rnn_lm/train.py @@ -99,6 +99,15 @@ def get_parser(): """, ) + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + parser.add_argument( "--exp-dir", type=str, @@ -242,7 +251,9 @@ def load_checkpoint_if_available( ) -> None: """Load checkpoint from file. - If params.start_epoch is positive, it will load the checkpoint from + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from `params.start_epoch - 1`. Otherwise, this function does nothing. Apart from loading state dict for `model`, `optimizer` and `scheduler`, @@ -261,10 +272,14 @@ def load_checkpoint_if_available( Returns: Return None. """ - if params.start_epoch <= 0: - return - filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + logging.info(f"Loading checkpoint: {filename}") saved_params = load_checkpoint( filename, @@ -283,6 +298,13 @@ def load_checkpoint_if_available( for k in keys: params[k] = saved_params[k] + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + return saved_params @@ -438,7 +460,14 @@ def train_one_epoch( tot_loss = MetricsTracker() + cur_batch_idx = params.get("cur_batch_idx", 0) + for batch_idx, batch in enumerate(train_dl): + + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + params.batch_idx_train += 1 x, y, sentence_lengths = batch batch_size = x.size(0) @@ -463,6 +492,7 @@ def train_one_epoch( params.batch_idx_train > 0 and params.batch_idx_train % params.save_every_n == 0 ): + params.cur_batch_idx = batch_idx save_checkpoint_with_global_batch_idx( out_dir=params.exp_dir, global_batch_idx=params.batch_idx_train, @@ -471,6 +501,7 @@ def train_one_epoch( optimizer=optimizer, rank=rank, ) + del params.cur_batch_idx if batch_idx % params.log_interval == 0: # Note: "frames" here means "num_tokens" diff --git a/icefall/utils.py b/icefall/utils.py index dfe9a7b42..0feff9dc8 100644 --- a/icefall/utils.py +++ b/icefall/utils.py @@ -429,6 +429,8 @@ def store_transcripts( texts: An iterable of tuples. The first element is the cur_id, the second is the reference transcript and the third element is the predicted result. + If it is a multi-talker ASR system, the ref and hyp may also be lists of + strings. Returns: Return None. """ @@ -886,8 +888,167 @@ def write_error_stats_with_timestamps( hyp_count = corr + hyp_sub + ins print(f"{word} {corr} {tot_errs} {ref_count} {hyp_count}", file=f) + return float(tot_err_rate), float(mean_delay), float(var_delay) - return tot_err_rate, mean_delay, var_delay + +def write_surt_error_stats( + f: TextIO, + test_set_name: str, + results: List[Tuple[str, str]], + enable_log: bool = True, + num_channels: int = 2, +) -> float: + """Write statistics based on predicted results and reference transcripts for SURT + multi-talker ASR systems. The difference between this and the `write_error_stats` + is that this function finds the optimal speaker-agnostic WER using the ``meeteval`` + toolkit. + + Args: + f: File to write the statistics to. + test_set_name: Name of the test set. + results: List of tuples containing the utterance ID and the predicted + transcript. + enable_log: Whether to enable logging. + num_channels: Number of output channels/branches. Defaults to 2. + Returns: + Return None. + """ + from meeteval.wer import wer + + subs: Dict[Tuple[str, str], int] = defaultdict(int) + ins: Dict[str, int] = defaultdict(int) + dels: Dict[str, int] = defaultdict(int) + ref_lens: List[int] = [] + + print( + "Search below for sections starting with PER-UTT DETAILS:, " + "SUBSTITUTIONS:, DELETIONS:, INSERTIONS:, PER-WORD STATS:", + file=f, + ) + + print("", file=f) + print("PER-UTT DETAILS: corr or (ref->hyp) ", file=f) + + # `words` stores counts per word, as follows: + # corr, ref_sub, hyp_sub, ins, dels + words: Dict[str, List[int]] = defaultdict(lambda: [0, 0, 0, 0, 0]) + num_corr = 0 + ERR = "*" + for cut_id, ref, hyp in results: + # First compute the optimal assignment of references to output channels + orc_wer = wer.orc_word_error_rate(ref, hyp) + assignment = orc_wer.assignment + refs = [[] for _ in range(num_channels)] + # Assign references to channels + for i, ref_text in zip(assignment, ref): + refs[i] += ref_text.split() + hyps = [hyp_text.split() for hyp_text in hyp] + # Now compute the WER for each channel + for ref_c, hyp_c in zip(refs, hyps): + ref_lens.append(len(ref_c)) + ali = kaldialign.align(ref_c, hyp_c, ERR) + for ref_word, hyp_word in ali: + if ref_word == ERR: + ins[hyp_word] += 1 + words[hyp_word][3] += 1 + elif hyp_word == ERR: + dels[ref_word] += 1 + words[ref_word][4] += 1 + elif hyp_word != ref_word: + subs[(ref_word, hyp_word)] += 1 + words[ref_word][1] += 1 + words[hyp_word][2] += 1 + else: + words[ref_word][0] += 1 + num_corr += 1 + combine_successive_errors = True + if combine_successive_errors: + ali = [[[x], [y]] for x, y in ali] + for i in range(len(ali) - 1): + if ali[i][0] != ali[i][1] and ali[i + 1][0] != ali[i + 1][1]: + ali[i + 1][0] = ali[i][0] + ali[i + 1][0] + ali[i + 1][1] = ali[i][1] + ali[i + 1][1] + ali[i] = [[], []] + ali = [ + [ + list(filter(lambda a: a != ERR, x)), + list(filter(lambda a: a != ERR, y)), + ] + for x, y in ali + ] + ali = list(filter(lambda x: x != [[], []], ali)) + ali = [ + [ + ERR if x == [] else " ".join(x), + ERR if y == [] else " ".join(y), + ] + for x, y in ali + ] + + print( + f"{cut_id}:\t" + + " ".join( + ( + ref_word + if ref_word == hyp_word + else f"({ref_word}->{hyp_word})" + for ref_word, hyp_word in ali + ) + ), + file=f, + ) + ref_len = sum(ref_lens) + sub_errs = sum(subs.values()) + ins_errs = sum(ins.values()) + del_errs = sum(dels.values()) + tot_errs = sub_errs + ins_errs + del_errs + tot_err_rate = "%.2f" % (100.0 * tot_errs / ref_len) + + if enable_log: + logging.info( + f"[{test_set_name}] %WER {tot_errs / ref_len:.2%} " + f"[{tot_errs} / {ref_len}, {ins_errs} ins, " + f"{del_errs} del, {sub_errs} sub ]" + ) + + print(f"%WER = {tot_err_rate}", file=f) + print( + f"Errors: {ins_errs} insertions, {del_errs} deletions, " + f"{sub_errs} substitutions, over {ref_len} reference " + f"words ({num_corr} correct)", + file=f, + ) + + print("", file=f) + print("SUBSTITUTIONS: count ref -> hyp", file=f) + + for count, (ref, hyp) in sorted([(v, k) for k, v in subs.items()], reverse=True): + print(f"{count} {ref} -> {hyp}", file=f) + + print("", file=f) + print("DELETIONS: count ref", file=f) + for count, ref in sorted([(v, k) for k, v in dels.items()], reverse=True): + print(f"{count} {ref}", file=f) + + print("", file=f) + print("INSERTIONS: count hyp", file=f) + for count, hyp in sorted([(v, k) for k, v in ins.items()], reverse=True): + print(f"{count} {hyp}", file=f) + + print("", file=f) + print("PER-WORD STATS: word corr tot_errs count_in_ref count_in_hyp", file=f) + for _, word, counts in sorted( + [(sum(v[1:]), k, v) for k, v in words.items()], reverse=True + ): + (corr, ref_sub, hyp_sub, ins, dels) = counts + tot_errs = ref_sub + hyp_sub + ins + dels + ref_count = corr + ref_sub + dels + hyp_count = corr + hyp_sub + ins + + print(f"{word} {corr} {tot_errs} {ref_count} {hyp_count}", file=f) + + print(f"%WER = {tot_err_rate}", file=f) + return float(tot_err_rate) class MetricsTracker(collections.defaultdict):