Add zipformer from Dan using multi-dataset setup (#675)

* Bug fix

* Change subsamplling factor from 1 to 2

* Implement AttentionCombine as replacement for RandomCombine

* Decrease random_prob from 0.5 to 0.333

* Add print statement

* Apply single_prob mask, so sometimes we just get one layer as output.

* Introduce feature mask per frame

* Include changes from Liyong about padding conformer module.

* Reduce single_prob from 0.5 to 0.25

* Reduce feature_mask_dropout_prob from 0.25 to 0.15.

* Remove dropout from inside ConformerEncoderLayer, for adding to residuals

* Increase feature_mask_dropout_prob from 0.15 to 0.2.

* Swap random_prob and single_prob, to reduce prob of being randomized.

* Decrease feature_mask_dropout_prob back from 0.2 to 0.15, i.e. revert the 43->48 change.

* Randomize order of some modules

* Bug fix

* Stop backprop bug

* Introduce a scale dependent on the masking value

* Implement efficient layer dropout

* Simplify the learned scaling factor on the modules

* Compute valid loss on batch 0.

* Make the scaling factors more global and the randomness of dropout more random

* Bug fix

* Introduce offset in layerdrop_scaleS

* Remove final combination; implement layer drop that drops the final layers.

* Bug fices

* Fix bug RE self.training

* Fix bug setting layerdrop mask

* Fix eigs call

* Add debug info

* Remove warmup

* Remove layer dropout and model-level warmup

* Don't always apply the frame mask

* Slight code cleanup/simplification

* Various fixes, finish implementating frame masking

* Remove debug info

* Don't compute validation if printing diagnostics.

* Apply layer bypass during warmup in a new way, including 2s and 4s of layers.

* Update checkpoint.py to deal with int params

* Revert initial_scale to previous values.

* Remove the feature where it was bypassing groups of layers.

* Implement layer dropout with probability 0.075

* Fix issue with warmup in test time

* Add warmup schedule where dropout disappears from earlier layers first.

* Have warmup that gradually removes dropout from layers; multiply initialization scales by 0.1.

* Do dropout a different way

* Fix bug in warmup

* Remove debug print

* Make the warmup mask per frame.

* Implement layer dropout (in a relatively efficient way)

* Decrease initial keep_prob to 0.25.

* Make it start warming up from the very start, and increase warmup_batches to 6k

* Change warmup schedule and increase warmup_batches from 4k to 6k

* Make the bypass scale trainable.

* Change the initial keep-prob back from 0.25 to 0.5

* Bug fix

* Limit bypass scale to >= 0.1

* Revert "Change warmup schedule and increase warmup_batches from 4k to 6k"

This reverts commit 86845bd5d859ceb6f83cd83f3719c3e6641de987.

* Do warmup by dropping out whole layers.

* Decrease frequency of logging variance_proportion

* Make layerdrop different in different processes.

* For speed, drop the same num layers per job.

* Decrease initial_layerdrop_prob from 0.75 to 0.5

* Revert also the changes in scaled_adam_exp85 regarding warmup schedule

* Remove unused code LearnedScale.

* Reintroduce batching to the optimizer

* Various fixes from debugging with nvtx, but removed the NVTX annotations.

* Only apply ActivationBalancer with prob 0.25.

* Fix s -> scaling for import.

* Increase final layerdrop prob from 0.05 to 0.075

* Fix bug where fewer layers were dropped than should be; remove unnecesary print statement.

* Fix bug in choosing layers to drop

* Refactor RelPosMultiheadAttention to have 2nd forward function and introduce more modules in conformer encoder layer

* Reduce final layerdrop_prob from 0.075 to 0.05.

* Fix issue with diagnostics if stats is None

* Remove persistent attention scores.

* Make ActivationBalancer and MaxEig more efficient.

* Cosmetic improvements

* Change scale_factor_scale from 0.5 to 0.8

* Make the ActivationBalancer regress to the data mean, not zero, when enforcing abs constraint.

* Remove unused config value

* Fix bug when channel_dim < 0

* Fix bug when channel_dim < 0

* Simplify how the positional-embedding scores work in attention (thanks to Zengwei for this concept)

* Revert dropout on attention scores to 0.0.

* This should just be a cosmetic change, regularizing how we get the warmup times from the layers.

* Reduce beta from 0.75 to  0.0.

* Reduce stats period from 10 to 4.

* Reworking of ActivationBalancer code to hopefully balance speed and effectiveness.

* Add debug code for attention weihts and eigs

* Remove debug statement

* Add different debug info.

* Penalize attention-weight entropies above a limit.

* Remove debug statements

* use larger delta but only penalize if small grad norm

* Bug fixes; change debug freq

* Change cutoff for small_grad_norm

* Implement whitening of values in conformer.

* Also whiten the keys in conformer.

* Fix an issue with scaling of grad.

* Decrease whitening limit from 2.0 to 1.1.

* Fix debug stats.

* Reorganize Whiten() code; configs are not the same as before.  Also remove MaxEig for self_attn module

* Bug fix RE float16

* Revert whitening_limit from 1.1 to 2.2.

* Replace MaxEig with Whiten with limit=5.0, and move it to end of ConformerEncoderLayer

* Change LR schedule to start off higher

* Simplify the dropout mask, no non-dropped-out sequences

* Make attention dims configurable, not embed_dim//2, trying 256.

* Reduce attention_dim to 192; cherry-pick scaled_adam_exp130 which is linear_pos interacting with query

* Use half the dim for values, vs. keys and queries.

* Increase initial-lr from 0.04 to 0.05, plus changes for diagnostics

* Cosmetic changes

* Changes to avoid bug in backward hooks, affecting diagnostics.

* Random clip attention scores to -5..5.

* Add some random clamping in model.py

* Add reflect=0.1 to invocations of random_clamp()

* Remove in_balancer.

* Revert model.py so there are no constraints on the output.

* Implement randomized backprop for softmax.

* Reduce min_abs from 1e-03 to 1e-04

* Add RandomGrad with min_abs=1.0e-04

* Use full precision to do softmax and store ans.

* Fix bug in backprop of random_clamp()

* Get the randomized backprop for softmax in autocast mode working.

* Remove debug print

* Reduce min_abs from 1.0e-04 to 5.0e-06

* Add hard limit of attention weights to +- 50

* Use normal implementation of softmax.

* Remove use of RandomGrad

* Remove the use of random_clamp in conformer.py.

* Reduce the limit on attention weights from 50 to 25.

* Reduce min_prob of ActivationBalancer from 0.1 to 0.05.

* Penalize too large weights in softmax of AttentionDownsample()

* Also apply limit on logit in SimpleCombiner

* Increase limit on logit for SimpleCombiner to 25.0

* Add more diagnostics to debug gradient scale problems

* Changes to grad scale logging; increase grad scale more frequently if less than one.

* Add logging

* Remove comparison diagnostics, which were not that useful.

* Configuration changes: scores limit 5->10, min_prob 0.05->0.1, cur_grad_scale more aggressive increase

* Reset optimizer state when we change loss function definition.

* Make warmup period decrease scale on simple loss, leaving pruned loss scale constant.

* Cosmetic change

* Increase initial-lr from 0.05 to 0.06.

* Increase initial-lr from 0.06 to 0.075 and decrease lr-epochs from 3.5 to 3.

* Fixes to logging statements.

* Introduce warmup schedule in optimizer

* Increase grad_scale to Whiten module

* Add inf check hooks

* Renaming in optim.py; remove step() from scan_pessimistic_batches_for_oom in train.py

* Change base lr to 0.1, also rename from initial lr in train.py

* Adding activation balancers after simple_am_prob and simple_lm_prob

* Reduce max_abs on am_balancer

* Increase max_factor in final lm_balancer and am_balancer

* Use penalize_abs_values_gt, not ActivationBalancer.

* Trying to reduce grad_scale of Whiten() from  0.02 to 0.01.

* Add hooks.py, had negleted to  git add it.

* don't do penalize_values_gt on simple_lm_proj and simple_am_proj; reduce --base-lr from 0.1 to  0.075

* Increase probs of activation balancer and make it decay slower.

* Dont print out full non-finite tensor

* Increase default max_factor for ActivationBalancer from 0.02 to 0.04; decrease max_abs in ConvolutionModule.deriv_balancer2 from 100.0 to 20.0

* reduce initial scale in GradScaler

* Increase max_abs in ActivationBalancer of conv module from 20 to 50

* --base-lr0.075->0.5; --lr-epochs 3->3.5

* Revert 179->180 change, i.e. change max_abs for deriv_balancer2 back from 50.0 20.0

* Save some memory in the autograd of DoubleSwish.

* Change the discretization of the sigmoid to be expectation preserving.

* Fix randn to rand

* Try a more exact way to round to uint8 that should prevent ever wrapping around to zero

* Make it use float16 if in amp but use clamp to avoid wrapping error

* Store only half precision output for softmax.

* More memory efficient backprop for DoubleSwish.

* Change to warmup schedule.

* Changes to more accurately estimate OOM conditions

* Reduce cutoff from 100 to 5 for estimating OOM with warmup

* Make 20 the limit for warmup_count

* Cast to float16 in DoubleSwish forward

* Hopefully make penalize_abs_values_gt more memory efficient.

* Add logging about memory used.

* Change scalar_max in optim.py from 2.0 to 5.0

* Regularize how we apply the min and max to the eps of BasicNorm

* Fix clamping of bypass scale; remove a couple unused variables.

* Increase floor on bypass_scale from 0.1 to 0.2.

* Increase bypass_scale from 0.2 to 0.4.

* Increase bypass_scale min from 0.4 to 0.5

* Rename conformer.py to zipformer.py

* Rename Conformer to Zipformer

* Update decode.py by copying from pruned_transducer_stateless5 and changing directory name

* Remove some unused variables.

* Fix clamping of epsilon

* Refactor zipformer for more flexibility so we can change number of encoder layers.

* Have a 3rd encoder, at downsampling factor of 8.

* Refactor how the downsampling is done so that it happens later, but the 1st encoder stack still operates after a subsampling of 2.

* Fix bug RE seq lengths

* Have 4 encoder stacks

* Have 6 different encoder stacks, U-shaped network.

* Reduce dim of linear positional encoding in attention layers.

* Reduce min of bypass_scale from 0.5 to 0.3, and make it not applied in test mode.

* Tuning change to num encoder layers, inspired by relative param importance.

* Make decoder group size equal to 4.

* Add skip connections as in normal U-net

* Avoid falling off the loop for weird inputs

* Apply layer-skip dropout prob

* Have warmup schedule for layer-skipping

* Rework how warmup count is produced; should not affect results.

* Add warmup schedule for zipformer encoder layer, from 1.0 -> 0.2.

* Reduce initial clamp_min for bypass_scale from 1.0 to 0.5.

* Restore the changes from scaled_adam_219 and scaled_adam_exp220,  accidentally lost, re layer skipping

* Change to schedule of bypass_scale min: make it larger, decrease slower.

* Change schedule after initial loss not promising

* Implement pooling module, add it after initial feedforward.

* Bug fix

* Introduce dropout rate to dynamic submodules of conformer.

* Introduce minimum probs in the SimpleCombiner

* Add bias in weight module

* Remove dynamic weights in SimpleCombine

* Remove the 5th of 6 encoder stacks

* Fix some typos

* small fixes

* small fixes

* Copy files

* Update decode.py

* Add changes from the master

* Add changes from the master

* update results

* Add CI

* Small fixes

* Small fixes

Co-authored-by: Daniel Povey <dpovey@gmail.com>

.github/scripts/run-librispeech-pruned-transducer-stateless7-2022-11-11.sh

@@ -33,6 +33,7 @@ popd
 log "Export to torchscript model"
 ./pruned_transducer_stateless7/export.py \
   --exp-dir $repo/exp \
+  --use-averaged-model false \
   --bpe-model $repo/data/lang_bpe_500/bpe.model \
   --epoch 99 \
   --avg 1 \

.github/scripts/run-librispeech-pruned-transducer-stateless8-2022-11-14.sh

@@ -0,0 +1,116 @@
+#!/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/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14
+
+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/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless8/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Export to torchscript model"
+./pruned_transducer_stateless8/export.py \
+  --exp-dir $repo/exp \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --use-averaged-model false \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless8/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless8/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless8/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $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 pruned_transducer_stateless8/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless8/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless8/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless8/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless8/exp
+  done
+
+  rm pruned_transducer_stateless8/exp/*.pt
+fi

.github/workflows/run-librispeech-2022-11-14-stateless8.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-2022-11-14-stateless8
+# 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 * * *"
+
+jobs:
+  run_librispeech_2022_11_14_zipformer_stateless8:
+    if: 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2022-09-25
+
+      - 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: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless8-2022-11-14.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless8
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless8/exp
+
+          cd pruned_transducer_stateless8
+          echo "results for pruned_transducer_stateless8"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless8
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless8-2022-11-14
+          path: egs/librispeech/ASR/pruned_transducer_stateless8/exp/

egs/librispeech/ASR/README.md

@@ -23,6 +23,7 @@ The following table lists the differences among them.
 | `pruned_transducer_stateless5`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + more layers + random combiner|
 | `pruned_transducer_stateless6`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + distillation with hubert|
 | `pruned_transducer_stateless7`        | Zipformer | Embedding + Conv1d | First experiment with Zipformer from Dan|
+| `pruned_transducer_stateless8`        | Zipformer | Embedding + Conv1d | Same as pruned_transducer_stateless7, but using extra data from GigaSpeech|
 | `pruned_stateless_emformer_rnnt2`     | Emformer(from torchaudio) | Embedding + Conv1d | Using Emformer from torchaudio for streaming ASR|
 | `conv_emformer_transducer_stateless`  | ConvEmformer | Embedding + Conv1d | Using ConvEmformer for streaming ASR + mechanisms in reworked model |
 | `conv_emformer_transducer_stateless2` | ConvEmformer | Embedding + Conv1d | Using ConvEmformer with simplified memory for streaming ASR + mechanisms in reworked model |

egs/librispeech/ASR/RESULTS.md

@@ -1,5 +1,63 @@
 ## Results
 
+### pruned_transducer_stateless8 (zipformer + multidataset)
+
+See <https://github.com/k2-fsa/icefall/pull/675> for more details.
+
+[pruned_transducer_stateless8](./pruned_transducer_stateless8)
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/y6kAPnN3S3OwvQxQqKQzsQ>
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+Number of model parameters: 70369391, i.e., 70.37 M
+
+|                      | test-clean | test-other  | comment                                |
+|----------------------|------------|-------------|----------------------------------------|
+| greedy search        | 1.87       | 4.38        | --epoch 16 --avg 2 --max-duration 600  |
+| modified beam search | 1.81       | 4.34        | --epoch 16 --avg 2 --max-duration 600  |
+| fast beam search     | 1.91       | 4.33        | --epoch 16 --avg 2 --max-duration 600  |
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless8/train.py \
+  --world-size 8 \
+  --num-epochs 20 \
+  --full-libri 1 \
+  --use-fp16 1 \
+  --max-duration 750 \
+  --exp-dir pruned_transducer_stateless8/exp \
+  --feedforward-dims  "1024,1024,2048,2048,1024" \
+  --master-port 12535 \
+  --giga-prob 0.9
+```
+
+The decoding commands are:
+```bash
+for m in greedy_search fast_beam_search modified_beam_search ; do
+  for epoch in 16; do
+    for avg in 2; do
+      ./pruned_transducer_stateless8/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --use-averaged-model 1 \
+          --exp-dir ./pruned_transducer_stateless8/exp \
+          --feedforward-dims  "1024,1024,2048,2048,1024" \
+          --max-duration 600 \
+          --decoding-method $m
+    done
+  done
+done
+```
+
+
 ### pruned_transducer_stateless7 (zipformer)
 
 See <https://github.com/k2-fsa/icefall/pull/672> for more details.

egs/librispeech/ASR/pruned_transducer_stateless7/jit_pretrained.py

@@ -30,6 +30,7 @@ Usage of this script:
 
 ./pruned_transducer_stateless7/jit_pretrained.py \
   --nn-model-filename ./pruned_transducer_stateless7/exp/cpu_jit.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
   /path/to/foo.wav \
   /path/to/bar.wav
 """

egs/librispeech/ASR/pruned_transducer_stateless8/asr_datamodule.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/asr_datamodule.py

egs/librispeech/ASR/pruned_transducer_stateless8/beam_search.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py

egs/librispeech/ASR/pruned_transducer_stateless8/decode.py

@@ -0,0 +1,863 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from librispeech import LibriSpeech
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    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=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless8/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        feature_lens += params.left_context
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.left_context),
+            value=LOG_EPS,
+        )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(
+            x=feature, x_lens=feature_lens
+        )
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif (
+        params.decoding_method == "greedy_search"
+        and params.max_sym_per_frame == 1
+    ):
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(
+                f"batch {batch_str}, cuts processed until now is {num_cuts}"
+            )
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir
+        / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += (
+            f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        )
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    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),
+                strict=False,
+            )
+        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), strict=False
+            )
+    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,
+                ),
+                strict=False,
+            )
+        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,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(
+                params.vocab_size - 1, device=device
+            )
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts)
+    test_other_dl = asr_datamodule.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,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless8/decoder.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/decoder.py

egs/librispeech/ASR/pruned_transducer_stateless8/encoder_interface.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py

egs/librispeech/ASR/pruned_transducer_stateless8/export.py

@@ -0,0 +1,334 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless8/export.py \
+  --exp-dir ./pruned_transducer_stateless8/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless8/export.py \
+  --exp-dir ./pruned_transducer_stateless8/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless8/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless8/decode.py \
+        --exp-dir ./pruned_transducer_stateless8/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14
+    # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless8/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device),
+                strict=False,
+            )
+        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),
+                strict=False,
+            )
+    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,
+                ),
+                strict=False,
+            )
+        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,
+                ),
+                strict=False,
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script()")
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/librispeech/ASR/pruned_transducer_stateless8/gigaspeech.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/gigaspeech.py

egs/librispeech/ASR/pruned_transducer_stateless8/jit_pretrained.py

@@ -0,0 +1,275 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless8/export.py \
+  --exp-dir ./pruned_transducer_stateless8/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless8/jit_pretrained.py \
+  --nn-model-filename ./pruned_transducer_stateless8/exp/cpu_jit.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    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}. "
+            f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+
+    model.eval()
+
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    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 = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    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(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(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/pruned_transducer_stateless8/joiner.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/joiner.py

egs/librispeech/ASR/pruned_transducer_stateless8/librispeech.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/librispeech.py

egs/librispeech/ASR/pruned_transducer_stateless8/model.py

@@ -0,0 +1,222 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import random
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import penalize_abs_values_gt
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        decoder_giga: Optional[nn.Module] = None,
+        joiner_giga: Optional[nn.Module] = None,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.decoder_giga = decoder_giga
+        self.joiner_giga = joiner_giga
+
+        self.simple_am_proj = nn.Linear(
+            encoder_dim,
+            vocab_size,
+        )
+        self.simple_lm_proj = nn.Linear(decoder_dim, vocab_size)
+
+        if decoder_giga is not None:
+            self.simple_am_proj_giga = nn.Linear(encoder_dim, vocab_size)
+            self.simple_lm_proj_giga = nn.Linear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        libri: bool = True,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> 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.
+          libri:
+            True to use the decoder and joiner for the LibriSpeech dataset.
+            False to use the decoder and joiner for the GigaSpeech dataset.
+          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 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
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        if libri:
+            decoder = self.decoder
+            simple_lm_proj = self.simple_lm_proj
+            simple_am_proj = self.simple_am_proj
+            joiner = self.joiner
+        else:
+            decoder = self.decoder_giga
+            simple_lm_proj = self.simple_lm_proj_giga
+            simple_am_proj = self.simple_am_proj_giga
+            joiner = self.joiner_giga
+
+        # 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 = decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (x.size(0), 4), dtype=torch.int64, device=x.device
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = simple_lm_proj(decoder_out)
+        am = simple_am_proj(encoder_out)
+
+        # if self.training and random.random() < 0.25:
+        #    lm = penalize_abs_values_gt(lm, 100.0, 1.0e-04)
+        # if self.training and random.random() < 0.25:
+        #    am = penalize_abs_values_gt(am, 30.0, 1.0e-04)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=joiner.encoder_proj(encoder_out),
+            lm=joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)

egs/librispeech/ASR/pruned_transducer_stateless8/optim.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py

egs/librispeech/ASR/pruned_transducer_stateless8/pretrained.py

@@ -0,0 +1,363 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless8/export.py \
+  --exp-dir ./pruned_transducer_stateless8/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless8/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless8/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless8/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless8/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless8/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless8/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless8/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless8/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless8/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless8/exp/pretrained.pt is generated by
+./pruned_transducer_stateless8/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import str2bool
+
+
+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(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. "
+            f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    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}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    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()
+    model.device = device
+
+    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(
+        x=features, x_lens=feature_lengths
+    )
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(sp.decode(hyp).split())
+
+    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/pruned_transducer_stateless8/scaling.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py

egs/librispeech/ASR/pruned_transducer_stateless8/scaling_converter.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py

egs/librispeech/ASR/pruned_transducer_stateless8/zipformer.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless7/zipformer.py