Merge branch 'master' of https://github.com/k2-fsa/icefall into surt

.github/scripts/run-librispeech-zipformer-ctc-2023-06-14.sh

@@ -0,0 +1,117 @@
+#!/usr/bin/env bash
+
+set -e
+
+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]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+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/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/Linv.pt"
+git lfs pull --include "data/lm/G_4_gram.pt"
+git lfs pull --include "exp/jit_script.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./zipformer/export.py \
+  --exp-dir $repo/exp \
+  --use-transducer 1 \
+  --use-ctc 1 \
+  --use-averaged-model false \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+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 \
+    --model-filename $repo/exp/jit_script.pt \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $method \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in ctc-decoding 1best; do
+  log "$method"
+
+  ./zipformer/pretrained_ctc.py \
+    --use-transducer 1 \
+    --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  \
+    --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  \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p zipformer/exp
+  ln -s $PWD/$repo/exp/pretrained.pt zipformer/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh zipformer/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in ctc-decoding 1best; do
+    log "Decoding with $method"
+
+    ./zipformer/ctc_decode.py \
+      --use-transducer 1 \
+      --use-ctc 1 \
+      --decoding-method $method \
+      --nbest-scale 1.0 \
+      --hlg-scale 0.6 \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir zipformer/exp
+  done
+
+  rm zipformer/exp/*.pt
+fi

.github/workflows/run-librispeech-zipformer-ctc-2023-06-14.yml

@@ -0,0 +1,155 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# 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.
+
+name: run-librispeech-zipformer-ctc-2023-06-14
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2023_06_14_zipformer-ctc-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2023_06_14_zipformer_ctc:
+    if: github.event.label.name == 'zipformer' ||github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-zipformer-ctc-2023-06-14.sh
+
+      - name: Display decoding results for librispeech zipformer
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./zipformer/exp
+
+          cd zipformer
+          echo "results for zipformer"
+          echo "===ctc-decoding==="
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech zipformer
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-zipformer-2022-11-11
+          path: egs/librispeech/ASR/zipformer/exp/

egs/librispeech/ASR/README.md

@@ -47,6 +47,7 @@ We place an additional Conv1d layer right after the input embedding layer.
 | `conformer-ctc`              | Conformer          | Use auxiliary attention head |
 | `conformer-ctc2`             | Reworked Conformer | Use auxiliary attention head |
 | `conformer-ctc3`             | Reworked Conformer | Streaming version + delay penalty |
+| `zipformer`                  | Upgraded Zipformer | Use auxiliary transducer head | The latest recipe |
 
 # MMI
 

egs/librispeech/ASR/RESULTS.md

@@ -1,5 +1,69 @@
 ## Results
 
+### zipformer (zipformer + pruned stateless transducer + CTC)
+
+See <https://github.com/k2-fsa/icefall/pull/1111> for more details.
+
+[zipformer](./zipformer)
+
+#### Non-streaming
+
+##### normal-scaled model, number of model parameters: 65805511, i.e., 65.81 M
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/Lo3Qlad7TP68ulM2K0ixgQ/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+Results of the CTC head:
+
+| decoding method         | test-clean | test-other | comment            |
+|-------------------------|------------|------------|--------------------|
+| ctc-decoding            | 2.40       | 5.66       | --epoch 40 --avg 16 |
+| 1best                   | 2.46       | 5.11       | --epoch 40 --avg 16 |
+| nbest                   | 2.46       | 5.11       | --epoch 40 --avg 16 |
+| nbest-rescoring         | 2.37       | 4.93       | --epoch 40 --avg 16 |
+| whole-lattice-rescoring | 2.37       | 4.88       | --epoch 40 --avg 16 |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp-ctc-rnnt \
+  --causal 0 \
+  --use-transducer 1 \
+  --use-ctc 1 \
+  --ctc-loss-scale 0.2 \
+  --full-libri 1 \
+  --max-duration 1000
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+for m in ctc-decoding 1best nbest nbest-rescoring whole-lattice-rescoring; do
+  ./zipformer/ctc_decode.py \
+      --epoch 40 \
+      --avg 16 \
+      --exp-dir zipformer/exp-ctc-rnnt \
+      --use-transducer 1 \
+      --use-ctc 1 \
+      --max-duration 300 \
+      --causal 0 \
+      --num-paths 100 \
+      --nbest-scale 1.0 \
+      --hlg-scale 0.6 \
+      --decoding-method $m
+done
+```
+
 ### zipformer (zipformer + pruned stateless transducer)
 
 See <https://github.com/k2-fsa/icefall/pull/1058> for more details.

egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py

@@ -22,7 +22,6 @@ from typing import Dict, List, Optional, Tuple, Union
 import k2
 import sentencepiece as spm
 import torch
-from model import Transducer
 
 from icefall import ContextGraph, ContextState, NgramLm, NgramLmStateCost
 from icefall.decode import Nbest, one_best_decoding
@@ -36,10 +35,11 @@ from icefall.utils import (
     get_texts,
     get_texts_with_timestamp,
 )
+from torch import nn
 
 
 def fast_beam_search_one_best(
-    model: Transducer,
+    model: nn.Module,
     decoding_graph: k2.Fsa,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
@@ -103,7 +103,7 @@ def fast_beam_search_one_best(
 
 
 def fast_beam_search_nbest_LG(
-    model: Transducer,
+    model: nn.Module,
     decoding_graph: k2.Fsa,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
@@ -229,7 +229,7 @@ def fast_beam_search_nbest_LG(
 
 
 def fast_beam_search_nbest(
-    model: Transducer,
+    model: nn.Module,
     decoding_graph: k2.Fsa,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
@@ -319,7 +319,7 @@ def fast_beam_search_nbest(
 
 
 def fast_beam_search_nbest_oracle(
-    model: Transducer,
+    model: nn.Module,
     decoding_graph: k2.Fsa,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
@@ -424,7 +424,7 @@ def fast_beam_search_nbest_oracle(
 
 
 def fast_beam_search(
-    model: Transducer,
+    model: nn.Module,
     decoding_graph: k2.Fsa,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
@@ -523,7 +523,7 @@ def fast_beam_search(
 
 
 def greedy_search(
-    model: Transducer,
+    model: nn.Module,
     encoder_out: torch.Tensor,
     max_sym_per_frame: int,
     return_timestamps: bool = False,
@@ -623,7 +623,7 @@ def greedy_search(
 
 
 def greedy_search_batch(
-    model: Transducer,
+    model: nn.Module,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
     return_timestamps: bool = False,
@@ -917,7 +917,7 @@ def get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape:
 
 
 def modified_beam_search(
-    model: Transducer,
+    model: nn.Module,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
     context_graph: Optional[ContextGraph] = None,
@@ -1119,7 +1119,7 @@ def modified_beam_search(
 
 
 def modified_beam_search_lm_rescore(
-    model: Transducer,
+    model: nn.Module,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
     LM: LmScorer,
@@ -1317,7 +1317,7 @@ def modified_beam_search_lm_rescore(
 
 
 def modified_beam_search_lm_rescore_LODR(
-    model: Transducer,
+    model: nn.Module,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
     LM: LmScorer,
@@ -1533,7 +1533,7 @@ def modified_beam_search_lm_rescore_LODR(
 
 
 def _deprecated_modified_beam_search(
-    model: Transducer,
+    model: nn.Module,
     encoder_out: torch.Tensor,
     beam: int = 4,
     return_timestamps: bool = False,
@@ -1658,7 +1658,7 @@ def _deprecated_modified_beam_search(
 
 
 def beam_search(
-    model: Transducer,
+    model: nn.Module,
     encoder_out: torch.Tensor,
     beam: int = 4,
     temperature: float = 1.0,
@@ -1818,7 +1818,7 @@ def beam_search(
 
 
 def fast_beam_search_with_nbest_rescoring(
-    model: Transducer,
+    model: nn.Module,
     decoding_graph: k2.Fsa,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
@@ -1978,7 +1978,7 @@ def fast_beam_search_with_nbest_rescoring(
 
 
 def fast_beam_search_with_nbest_rnn_rescoring(
-    model: Transducer,
+    model: nn.Module,
     decoding_graph: k2.Fsa,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
@@ -2169,7 +2169,7 @@ def fast_beam_search_with_nbest_rnn_rescoring(
 
 
 def modified_beam_search_ngram_rescoring(
-    model: Transducer,
+    model: nn.Module,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
     ngram_lm: NgramLm,
@@ -2333,7 +2333,7 @@ def modified_beam_search_ngram_rescoring(
 
 
 def modified_beam_search_LODR(
-    model: Transducer,
+    model: nn.Module,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
     LODR_lm: NgramLm,
@@ -2604,7 +2604,7 @@ def modified_beam_search_LODR(
 
 
 def modified_beam_search_lm_shallow_fusion(
-    model: Transducer,
+    model: nn.Module,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
     LM: LmScorer,

egs/librispeech/ASR/zipformer/ctc_decode.py

@@ -0,0 +1,847 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Liyong Guo,
+#                                                 Quandong Wang,
+#                                                 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) ctc-decoding
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --decoding-method ctc-decoding
+
+(2) 1best
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --decoding-method 1best
+
+(3) nbest
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --decoding-method nbest
+
+(4) nbest-rescoring
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --nbest-scale 1.0 \
+    --lm-dir data/lm \
+    --decoding-method nbest-rescoring
+
+(5) whole-lattice-rescoring
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --nbest-scale 1.0 \
+    --lm-dir data/lm \
+    --decoding-method whole-lattice-rescoring
+"""
+
+
+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 LibriSpeechAsrDataModule
+from train import add_model_arguments, get_params, get_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    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(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="ctc-decoding",
+        help="""Decoding method.
+        Supported values are:
+        - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
+          model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        - (2) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (3) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+          rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+          the highest score is the decoding result.
+        - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+          n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+          is the decoding result.
+          you have trained an RNN LM using ./rnn_lm/train.py
+        - (6) nbest-oracle. Its WER is the lower bound of any n-best
+          rescoring method can achieve. Useful for debugging n-best
+          rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--hlg-scale",
+        type=float,
+        default=0.6,
+        help="""The scale to be applied to `hlg.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    G: 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 no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - 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`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      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.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.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,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=ctc_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.decoding_method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.decoding_method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.decoding_method in ["1best", "nbest"]:
+        if params.decoding_method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.decoding_method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.decoding_method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.decoding_method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.decoding_method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    G: 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.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring 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 = "?"
+
+    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,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+        )
+
+        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 % 100 == 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}-{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}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            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}-{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()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "ctc-decoding",
+        "1best",
+        "nbest",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "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 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}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    if params.decoding_method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        HLG.scores *= params.hlg_scale
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.decoding_method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.decoding_method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    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()
+
+    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
+    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):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/zipformer/decode.py

@@ -116,7 +116,7 @@ from beam_search import (
     greedy_search_batch,
     modified_beam_search,
 )
-from train import add_model_arguments, get_params, get_transducer_model
+from train import add_model_arguments, get_params, get_model
 
 from icefall.checkpoint import (
     average_checkpoints,
@@ -366,15 +366,7 @@ def decode_one_batch(
             value=LOG_EPS,
         )
 
-    x, x_lens = model.encoder_embed(feature, feature_lens)
+    encoder_out, encoder_out_lens = model.forward_encoder(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 = []
 
@@ -694,7 +686,7 @@ def main():
     logging.info(params)
 
     logging.info("About to create model")
-    model = get_transducer_model(params)
+    model = get_model(params)
 
     if not params.use_averaged_model:
         if params.iter > 0:

egs/librispeech/ASR/zipformer/decoder.py

@@ -58,7 +58,6 @@ class Decoder(nn.Module):
         self.embedding = nn.Embedding(
             num_embeddings=vocab_size,
             embedding_dim=decoder_dim,
-            padding_idx=blank_id,
         )
         # the balancers are to avoid any drift in the magnitude of the
         # embeddings, which would interact badly with parameter averaging.

egs/librispeech/ASR/zipformer/export-onnx-streaming.py

@@ -76,7 +76,7 @@ import torch.nn as nn
 from decoder import Decoder
 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_transducer_model
+from train import add_model_arguments, get_params, get_model
 from zipformer import Zipformer2
 
 from icefall.checkpoint import (
@@ -595,7 +595,7 @@ def main():
     logging.info(params)
 
     logging.info("About to create model")
-    model = get_transducer_model(params)
+    model = get_model(params)
 
     model.to(device)
 

egs/librispeech/ASR/zipformer/export-onnx.py

@@ -74,7 +74,7 @@ import torch.nn as nn
 from decoder import Decoder
 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_transducer_model
+from train import add_model_arguments, get_params, get_model
 from zipformer import Zipformer2
 
 from icefall.checkpoint import (
@@ -444,7 +444,7 @@ def main():
     logging.info(params)
 
     logging.info("About to create model")
-    model = get_transducer_model(params)
+    model = get_model(params)
 
     model.to(device)
 

egs/librispeech/ASR/zipformer/export.py

@@ -161,7 +161,7 @@ from typing import List, Tuple
 import sentencepiece as spm
 import torch
 from torch import Tensor, nn
-from train import add_model_arguments, get_params, get_transducer_model
+from train import add_model_arguments, get_params, get_model
 
 from icefall.checkpoint import (
     average_checkpoints,
@@ -408,7 +408,7 @@ def main():
     logging.info(params)
 
     logging.info("About to create model")
-    model = get_transducer_model(params)
+    model = get_model(params)
 
     if not params.use_averaged_model:
         if params.iter > 0:

egs/librispeech/ASR/zipformer/generate_averaged_model.py

@@ -44,7 +44,7 @@ import sentencepiece as spm
 import torch
 from asr_datamodule import LibriSpeechAsrDataModule
 
-from train import add_model_arguments, get_params, get_transducer_model
+from train import add_model_arguments, get_params, get_model
 
 from icefall.checkpoint import (
     average_checkpoints_with_averaged_model,
@@ -140,7 +140,7 @@ def main():
     params.vocab_size = sp.get_piece_size()
 
     print("About to create model")
-    model = get_transducer_model(params)
+    model = get_model(params)
 
     if params.iter > 0:
         filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[

egs/librispeech/ASR/zipformer/jit_pretrained.py

@@ -97,7 +97,7 @@ def read_sound_files(
             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])
+        ans.append(wave[0].contiguous())
     return ans
 
 

egs/librispeech/ASR/zipformer/jit_pretrained_ctc.py

@@ -0,0 +1,428 @@
+#!/usr/bin/env python3
+# Copyright      2022-2023  Xiaomi Corp.        (authors: 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.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --use-ctc 1 \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --use-ctc 1 \
+  --causal 1 \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+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 \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./zipformer/jit_pretrained_ctc.py \
+  --model-filename ./zipformer/exp/jit_script.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) nbest-rescoring
+./zipformer/jit_pretrained_ctc.py \
+  --model-filename ./zipformer/exp/jit_script.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(4) whole-lattice-rescoring
+./zipformer/jit_pretrained_ctc.py \
+  --model-filename ./zipformer/exp/jit_script.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from ctc_decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model.",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        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
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + nbest n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + whole-lattice n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    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.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> 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].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.model_filename)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    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, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(features, feature_lengths)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    batch_size = ctc_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i].item() // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        max_token_id = params.vocab_size - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=ctc_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            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]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=ctc_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\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()

egs/librispeech/ASR/zipformer/model.py

@@ -1,4 +1,6 @@
-# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Zengwei Yao)
 #
 # See ../../../../LICENSE for clarification regarding multiple authors
 #
@@ -14,6 +16,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import Optional, Tuple
 
 import k2
 import torch
@@ -24,23 +27,25 @@ from icefall.utils import add_sos, make_pad_mask
 from scaling import ScaledLinear
 
 
-class Transducer(nn.Module):
+class AsrModel(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,
+        decoder: Optional[nn.Module] = None,
-        joiner: nn.Module,
+        joiner: Optional[nn.Module] = None,
-        encoder_dim: int,
+        encoder_dim: int = 384,
-        decoder_dim: int,
+        decoder_dim: int = 512,
-        joiner_dim: int,
+        vocab_size: int = 500,
-        vocab_size: int,
+        use_transducer: bool = True,
+        use_ctc: bool = False,
     ):
-        """
+        """A joint CTC & Transducer ASR model.
+
+        - Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks (http://imagine.enpc.fr/~obozinsg/teaching/mva_gm/papers/ctc.pdf)
+        - Sequence Transduction with Recurrent Neural Networks (https://arxiv.org/pdf/1211.3711.pdf)
+        - Pruned RNN-T for fast, memory-efficient ASR training (https://arxiv.org/pdf/2206.13236.pdf)
+
         Args:
           encoder_embed:
             It is a Convolutional 2D subsampling module. It converts
@@ -55,47 +60,133 @@ class Transducer(nn.Module):
             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`.
+            It is used when use_transducer is True.
           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.
+            It is used when use_transducer is True.
+          use_transducer:
+            Whether use transducer head. Default: True.
+          use_ctc:
+            Whether use CTC head. Default: False.
         """
         super().__init__()
+
+        assert (
+            use_transducer or use_ctc
+        ), f"At least one of them should be True, but got use_transducer={use_transducer}, use_ctc={use_ctc}"
+
         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(
+        self.use_transducer = use_transducer
-            encoder_dim,
+        if use_transducer:
-            vocab_size,
+            # Modules for Transducer head
-            initial_scale=0.25,
+            assert decoder is not None
-        )
+            assert hasattr(decoder, "blank_id")
-        self.simple_lm_proj = ScaledLinear(
+            assert joiner is not None
-            decoder_dim,
-            vocab_size,
-            initial_scale=0.25,
-        )
 
-    def forward(
+            self.decoder = decoder
-        self,
+            self.joiner = joiner
-        x: torch.Tensor,
+
-        x_lens: torch.Tensor,
+            self.simple_am_proj = ScaledLinear(
-        y: k2.RaggedTensor,
+                encoder_dim, vocab_size, initial_scale=0.25
-        prune_range: int = 5,
+            )
-        am_scale: float = 0.0,
+            self.simple_lm_proj = ScaledLinear(
-        lm_scale: float = 0.0,
+                decoder_dim, vocab_size, initial_scale=0.25
-    ) -> torch.Tensor:
+            )
-        """
+        else:
+            assert decoder is None
+            assert joiner is None
+
+        self.use_ctc = use_ctc
+        if use_ctc:
+            # Modules for CTC head
+            self.ctc_output = nn.Sequential(
+                nn.Dropout(p=0.1),
+                nn.Linear(encoder_dim, vocab_size),
+                nn.LogSoftmax(dim=-1),
+            )
+
+    def forward_encoder(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute encoder outputs.
         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.
+
+        Returns:
+          encoder_out:
+            Encoder output, of shape (N, T, C).
+          encoder_out_lens:
+            Encoder output lengths, of shape (N,).
+        """
+        # 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, 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)
+        assert torch.all(encoder_out_lens > 0), (x_lens, encoder_out_lens)
+
+        return encoder_out, encoder_out_lens
+
+    def forward_ctc(
+        self,
+        encoder_out: torch.Tensor,
+        encoder_out_lens: torch.Tensor,
+        targets: torch.Tensor,
+        target_lengths: torch.Tensor,
+    ) -> torch.Tensor:
+        """Compute CTC loss.
+        Args:
+          encoder_out:
+            Encoder output, of shape (N, T, C).
+          encoder_out_lens:
+            Encoder output lengths, of shape (N,).
+          targets:
+            Target Tensor of shape (sum(target_lengths)). The targets are assumed
+            to be un-padded and concatenated within 1 dimension.
+        """
+        # Compute CTC log-prob
+        ctc_output = self.ctc_output(encoder_out)  # (N, T, C)
+
+        ctc_loss = torch.nn.functional.ctc_loss(
+            log_probs=ctc_output.permute(1, 0, 2),  # (T, N, C)
+            targets=targets,
+            input_lengths=encoder_out_lens,
+            target_lengths=target_lengths,
+            reduction="sum",
+        )
+        return ctc_loss
+
+    def forward_transducer(
+        self,
+        encoder_out: torch.Tensor,
+        encoder_out_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        y_lens: torch.Tensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute Transducer loss.
+        Args:
+          encoder_out:
+            Encoder output, of shape (N, T, C).
+          encoder_out_lens:
+            Encoder output lengths, of shape (N,).
           y:
             A ragged tensor with 2 axes [utt][label]. It contains labels of each
             utterance.
@@ -108,37 +199,8 @@ class Transducer(nn.Module):
           lm_scale:
             The scale to smooth the loss with lm (output of predictor network)
             part
-        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)
 
@@ -159,7 +221,7 @@ class Transducer(nn.Module):
             device=encoder_out.device,
         )
         boundary[:, 2] = y_lens
-        boundary[:, 3] = x_lens
+        boundary[:, 3] = encoder_out_lens
 
         lm = self.simple_lm_proj(decoder_out)
         am = self.simple_am_proj(encoder_out)
@@ -214,4 +276,83 @@ class Transducer(nn.Module):
                 reduction="sum",
             )
 
-        return (simple_loss, pruned_loss)
+        return simple_loss, pruned_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,
+    ) -> 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 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
+        Returns:
+          Return the transducer losses and CTC loss,
+          in form of (simple_loss, pruned_loss, ctc_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
+
+        # Compute encoder outputs
+        encoder_out, encoder_out_lens = self.forward_encoder(x, x_lens)
+
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        if self.use_transducer:
+            # Compute transducer loss
+            simple_loss, pruned_loss = self.forward_transducer(
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+                y=y.to(x.device),
+                y_lens=y_lens,
+                prune_range=prune_range,
+                am_scale=am_scale,
+                lm_scale=lm_scale,
+            )
+        else:
+            simple_loss = torch.empty(0)
+            pruned_loss = torch.empty(0)
+
+        if self.use_ctc:
+            # Compute CTC loss
+            targets = y.values
+            ctc_loss = self.forward_ctc(
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+                targets=targets,
+                target_lengths=y_lens,
+            )
+        else:
+            ctc_loss = torch.empty(0)
+
+        return simple_loss, pruned_loss, ctc_loss

egs/librispeech/ASR/zipformer/pretrained.py

@@ -120,9 +120,8 @@ from beam_search import (
     greedy_search_batch,
     modified_beam_search,
 )
-from icefall.utils import make_pad_mask
 from torch.nn.utils.rnn import pad_sequence
-from train import add_model_arguments, get_params, get_transducer_model
+from train import add_model_arguments, get_params, get_model
 
 
 def get_parser():
@@ -246,7 +245,7 @@ def read_sound_files(
             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])
+        ans.append(wave[0].contiguous())
     return ans
 
 
@@ -284,7 +283,7 @@ def main():
         ), "left_context_frames should be one value in decoding."
 
     logging.info("Creating model")
-    model = get_transducer_model(params)
+    model = get_model(params)
 
     num_param = sum([p.numel() for p in model.parameters()])
     logging.info(f"Number of model parameters: {num_param}")
@@ -318,15 +317,7 @@ def main():
     feature_lengths = torch.tensor(feature_lengths, device=device)
 
     # model forward
-    x, x_lens = model.encoder_embed(features, feature_lengths)
+    encoder_out, encoder_out_lens = model.forward_encoder(features, feature_lengths)
-
-    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 = []
     msg = f"Using {params.method}"

egs/librispeech/ASR/zipformer/pretrained_ctc.py

@@ -0,0 +1,446 @@
+#!/usr/bin/env python3
+# Copyright      2022-2023  Xiaomi Corp.        (authors: 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.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --use-ctc 1 \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 9
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --use-ctc 1 \
+  --causal 1 \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 9
+
+Usage of this script:
+
+(1) ctc-decoding
+./zipformer/pretrained_ctc.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./zipformer/pretrained_ctc.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) nbest-rescoring
+./zipformer/pretrained_ctc.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(4) whole-lattice-rescoring
+./zipformer/pretrained_ctc.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from ctc_decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_model
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    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(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        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
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + nbest n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + whole-lattice n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    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.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> 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}. " f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    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
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    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, device=device)
+
+    encoder_out, encoder_out_lens = model.forward_encoder(features, feature_lengths)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    batch_size = ctc_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i].item() // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        max_token_id = params.vocab_size - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=ctc_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            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]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=ctc_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\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()

egs/librispeech/ASR/zipformer/streaming_decode.py

@@ -51,7 +51,7 @@ from streaming_beam_search import (
 )
 from torch import Tensor, nn
 from torch.nn.utils.rnn import pad_sequence
-from train import add_model_arguments, get_params, get_transducer_model
+from train import add_model_arguments, get_params, get_model
 
 from icefall.checkpoint import (
     average_checkpoints,
@@ -756,7 +756,7 @@ def main():
     logging.info(params)
 
     logging.info("About to create model")
-    model = get_transducer_model(params)
+    model = get_model(params)
 
     if not params.use_averaged_model:
         if params.iter > 0:

egs/librispeech/ASR/zipformer/train.py

@@ -44,6 +44,10 @@ export CUDA_VISIBLE_DEVICES="0,1,2,3"
   --full-libri 1 \
   --max-duration 1000
 
+It supports training with:
+  - transducer loss (default), with `--use-transducer True --use-ctc False`
+  - ctc loss (not recommended), with `--use-transducer False --use-ctc True`
+  - transducer loss & ctc loss, with `--use-transducer True --use-ctc True`
 """
 
 
@@ -67,7 +71,7 @@ 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 Transducer
+from model import AsrModel
 from optim import Eden, ScaledAdam
 from scaling import ScheduledFloat
 from subsampling import Conv2dSubsampling
@@ -240,6 +244,20 @@ def add_model_arguments(parser: argparse.ArgumentParser):
         "chunk left-context frames will be chosen randomly from this list; else not relevant.",
     )
 
+    parser.add_argument(
+        "--use-transducer",
+        type=str2bool,
+        default=True,
+        help="If True, use Transducer head.",
+    )
+
+    parser.add_argument(
+        "--use-ctc",
+        type=str2bool,
+        default=False,
+        help="If True, use CTC head.",
+    )
+
 
 def get_parser():
     parser = argparse.ArgumentParser(
@@ -378,6 +396,13 @@ def get_parser():
         "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,
@@ -578,21 +603,33 @@ def get_joiner_model(params: AttributeDict) -> nn.Module:
     return joiner
 
 
-def get_transducer_model(params: AttributeDict) -> nn.Module:
+def get_model(params: AttributeDict) -> nn.Module:
+    assert (
+        params.use_transducer or params.use_ctc
+    ), (f"At least one of them should be True, "
+        f"but got params.use_transducer={params.use_transducer}, "
+        f"params.use_ctc={params.use_ctc}")
+
     encoder_embed = get_encoder_embed(params)
     encoder = get_encoder_model(params)
-    decoder = get_decoder_model(params)
-    joiner = get_joiner_model(params)
 
-    model = Transducer(
+    if params.use_transducer:
+        decoder = get_decoder_model(params)
+        joiner = get_joiner_model(params)
+    else:
+        decoder = None
+        joiner = None
+
+    model = AsrModel(
         encoder_embed=encoder_embed,
         encoder=encoder,
         decoder=decoder,
         joiner=joiner,
-        encoder_dim=int(max(params.encoder_dim.split(","))),
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
         decoder_dim=params.decoder_dim,
-        joiner_dim=params.joiner_dim,
         vocab_size=params.vocab_size,
+        use_transducer=params.use_transducer,
+        use_ctc=params.use_ctc,
     )
     return model
 
@@ -721,7 +758,7 @@ def compute_loss(
     is_training: bool,
 ) -> Tuple[Tensor, MetricsTracker]:
     """
-    Compute CTC loss given the model and its inputs.
+    Compute loss given the model and its inputs.
 
     Args:
       params:
@@ -752,10 +789,10 @@ def compute_loss(
 
     texts = batch["supervisions"]["text"]
     y = sp.encode(texts, out_type=int)
-    y = k2.RaggedTensor(y).to(device)
+    y = k2.RaggedTensor(y)
 
     with torch.set_grad_enabled(is_training):
-        simple_loss, pruned_loss = model(
+        simple_loss, pruned_loss, ctc_loss = model(
             x=feature,
             x_lens=feature_lens,
             y=y,
@@ -764,21 +801,27 @@ def compute_loss(
             lm_scale=params.lm_scale,
         )
 
-        s = params.simple_loss_scale
+        loss = 0.0
-        # 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
+        if params.use_transducer:
+            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
+            )
+
+        if params.use_ctc:
+            loss += params.ctc_loss_scale * ctc_loss
 
     assert loss.requires_grad == is_training
 
@@ -789,8 +832,11 @@ def compute_loss(
 
     # Note: We use reduction=sum while computing the loss.
     info["loss"] = loss.detach().cpu().item()
-    info["simple_loss"] = simple_loss.detach().cpu().item()
+    if params.use_transducer:
-    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+        info["simple_loss"] = simple_loss.detach().cpu().item()
+        info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    if params.use_ctc:
+        info["ctc_loss"] = ctc_loss.detach().cpu().item()
 
     return loss, info
 
@@ -1071,10 +1117,13 @@ def run(rank, world_size, args):
     params.blank_id = sp.piece_to_id("<blk>")
     params.vocab_size = sp.get_piece_size()
 
+    if not params.use_transducer:
+        params.ctc_loss_scale = 1.0
+
     logging.info(params)
 
     logging.info("About to create model")
-    model = get_transducer_model(params)
+    model = get_model(params)
 
     num_param = sum([p.numel() for p in model.parameters()])
     logging.info(f"Number of model parameters: {num_param}")

egs/librispeech/ASR/zipformer/zipformer.py

@@ -2190,7 +2190,7 @@ class ConvolutionModule(nn.Module):
 
         x = self.in_proj(x)  # (time, batch, 2*channels)
 
-        x, s = x.chunk(2, dim=-1)
+        x, s = x.chunk(2, dim=2)
         s = self.sigmoid(s)
         x = x * s
         # (time, batch, channels)