Add pruned RNN-T for aishell. (#436)

* Add pruned RNN-T for aishell.

* support torch script.

* Update CI.

* Minor fixes.

* Add links to sherpa.

.github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh

@@ -0,0 +1,86 @@
+#!/usr/bin/env bash
+
+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/aishell/ASR
+
+git lfs install
+
+fbank_url=https://huggingface.co/csukuangfj/aishell-test-dev-manifests
+log "Downloading pre-commputed fbank from $fbank_url"
+
+git clone https://huggingface.co/csukuangfj/aishell-test-dev-manifests
+ln -s $PWD/aishell-test-dev-manifests/data .
+
+log "Downloading pre-trained model from $repo_url"
+repo_url=https://huggingface.co/csukuangfj/icefall-aishell-pruned-transducer-stateless3-2022-06-20
+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
+ln -s pretrained-epoch-29-avg-5-torch-1.10.pt pretrained.pt
+popd
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless3/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $rep/test_wavs/BAC009S0764W0123.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless3/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $rep/test_wavs/BAC009S0764W0123.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_stateless3/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless3/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_char data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless3/exp
+
+  log "Decoding test and dev"
+
+  # 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_stateless3/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless3/exp
+  done
+
+  rm pruned_transducer_stateless3/exp/*.pt
+fi

.github/workflows/run-aishell-2022-06-20.yml

@@ -0,0 +1,119 @@
+# Copyright      2022  Fangjun Kuang ([email protected])
+
+# 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-aishell-2022-06-20
+# pruned RNN-T + reworked model with random combiner
+# https://huggingface.co/csukuangfj/icefall-aishell-pruned-transducer-stateless3-2022-06-20
+
+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_aishell_2022_06_20:
+    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-18.04]
+        python-version: [3.7, 3.8, 3.9]
+
+      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 }}
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh
+
+      - name: Display decoding results for aishell pruned_transducer_stateless3
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/aishell/ASR/
+          tree ./pruned_transducer_stateless3/exp
+
+          cd pruned_transducer_stateless3
+          echo "results for pruned_transducer_stateless3"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for dev" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for dev" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for dev" {} + | sort -n -k2
+
+      - name: Upload decoding results for aishell pruned_transducer_stateless3
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: aishell-torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless3-2022-06-20
+          path: egs/aishell/ASR/pruned_transducer_stateless3/exp/

egs/aishell/ASR/README.md

@@ -4,6 +4,8 @@
 Please refer to <https://icefall.readthedocs.io/en/latest/recipes/aishell/index.html>
 for how to run models in this recipe.
 
+
+
 # Transducers
 
 There are various folders containing the name `transducer` in this folder.
@@ -14,6 +16,7 @@ The following table lists the differences among them.
 | `transducer_stateless`             | Conformer | Embedding + Conv1d | with `k2.rnnt_loss`                                                               |
 | `transducer_stateless_modified`    | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer`                     |
 | `transducer_stateless_modified-2`  | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer` + extra data      |
+| `pruned_transducer_stateless3`     | Conformer (reworked) | Embedding + Conv1d | pruned RNN-T + reworked model with random combiner + using aidatatang_20zh as extra data|
 
 The decoder in `transducer_stateless` is modified from the paper
 [Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).

egs/aishell/ASR/RESULTS.md

@@ -1,10 +1,93 @@
 ## Results
-### Aishell training result(Transducer-stateless)
+
+### Aishell training result(Stateless Transducer)
+
+#### Pruned transducer stateless 3
+
+See <https://github.com/k2-fsa/icefall/pull/436>
+
+
+[./pruned_transducer_stateless3](./pruned_transducer_stateless3)
+
+It uses pruned RNN-T.
+
+|                        | test | dev  | comment                               |
+|------------------------|------|------|---------------------------------------|
+| greedy search          | 5.39 | 5.09 | --epoch 29 --avg 5 --max-duration 600 |
+| modified beam search   | 5.05 | 4.79 | --epoch 29 --avg 5 --max-duration 600 |
+| fast beam search       | 5.13 | 4.91 | --epoch 29 --avg 5 --max-duration 600 |
+
+Training command is:
+
+```bash
+./prepare.sh
+./prepare_aidatatang_200zh.sh
+
+export CUDA_VISIBLE_DEVICES="4,5,6,7"
+
+./pruned_transducer_stateless3/train.py \
+  --exp-dir ./pruned_transducer_stateless3/exp-context-size-1 \
+  --world-size 4 \
+  --max-duration 200 \
+  --datatang-prob 0.5 \
+  --start-epoch 1 \
+  --num-epochs 30 \
+  --use-fp16 1 \
+  --num-encoder-layers 12 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --context-size 1 \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --master-port 12356
+```
+
+**Caution**: It uses `--context-size=1`.
+
+The tensorboard log is available at
+<https://tensorboard.dev/experiment/OKKacljwR6ik7rbDr5gMqQ>
+
+The decoding command is:
+
+```bash
+for epoch in 29; do
+  for avg in 5; do
+    for m in greedy_search modified_beam_search fast_beam_search; do
+      ./pruned_transducer_stateless3/decode.py \
+        --exp-dir ./pruned_transducer_stateless3/exp-context-size-1 \
+        --epoch $epoch \
+        --avg $avg \
+        --use-averaged-model 1 \
+        --max-duration 600 \
+        --decoding-method $m \
+        --num-encoder-layers 12 \
+        --dim-feedforward 2048 \
+        --nhead 8 \
+        --context-size 1 \
+        --encoder-dim 512 \
+        --decoder-dim 512 \
+        --joiner-dim 512
+    done
+  done
+done
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/csukuangfj/icefall-aishell-pruned-transducer-stateless3-2022-06-20>
+
+We have a tutorial in [sherpa](https://github.com/k2-fsa/sherpa) about how
+to use the pre-trained model for non-streaming ASR. See
+<https://k2-fsa.github.io/sherpa/offline_asr/conformer/aishell.html>
 
 #### 2022-03-01
 
 [./transducer_stateless_modified-2](./transducer_stateless_modified-2)
 
+It uses [optimized_transducer](https://github.com/csukuangfj/optimized_transducer)
+for computing RNN-T loss.
+
 Stateless transducer + modified transducer + using [aidatatang_200zh](http://www.openslr.org/62/) as extra training data.
 
 

egs/aishell/ASR/pruned_transducer_stateless3/aidatatang_200zh.py

@@ -0,0 +1 @@
+../transducer_stateless_modified-2/aidatatang_200zh.py

egs/aishell/ASR/pruned_transducer_stateless3/aishell.py

@@ -0,0 +1 @@
+../transducer_stateless_modified-2/aishell.py

egs/aishell/ASR/pruned_transducer_stateless3/asr_datamodule.py

@@ -0,0 +1 @@
+../transducer_stateless_modified-2/asr_datamodule.py

egs/aishell/ASR/pruned_transducer_stateless3/beam_search.py

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

egs/aishell/ASR/pruned_transducer_stateless3/conformer.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/conformer.py