Convert conv-emformer to ncnn (k2-fsa#717)

* Export conv-emformer via torch.jit.trace()
tramphero · Dec 6, 2022 · f13cf61 · f13cf61
1 parent be6e08f
commit f13cf61
Show file tree

Hide file tree

Showing 11 changed files with 4,099 additions and 1 deletion.
diff --git a/.github/scripts/run-librispeech-conv-emformer-transducer-stateless2-2022-12-05.sh b/.github/scripts/run-librispeech-conv-emformer-transducer-stateless2-2022-12-05.sh
@@ -0,0 +1,79 @@
+#!/usr/bin/env bash
+#
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+pushd $repo
+git lfs pull --include "exp/pretrained-epoch-30-avg-10-averaged.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+cd exp
+ln -s pretrained-epoch-30-avg-10-averaged.pt epoch-99.pt
+popd
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+log  "Install ncnn and pnnx"
+
+# We are using a modified ncnn here. Will try to merge it to the official repo
+# of ncnn
+git clone https://github.com/csukuangfj/ncnn
+pushd ncnn
+git submodule init
+git submodule update python/pybind11
+python3 setup.py bdist_wheel
+ls -lh dist/
+pip install dist/*.whl
+cd tools/pnnx
+mkdir build
+cd build
+cmake -D Python3_EXECUTABLE=/opt/hostedtoolcache/Python/3.8.14/x64/bin/python3 ..
+make -j4 pnnx
+
+./src/pnnx || echo "pass"
+
+popd
+
+log "Test exporting to pnnx format"
+
+./conv_emformer_transducer_stateless2/export-for-ncnn.py \
+  --exp-dir $repo/exp \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+
+./ncnn/tools/pnnx/build/src/pnnx $repo/exp/encoder_jit_trace-pnnx.pt
+./ncnn/tools/pnnx/build/src/pnnx $repo/exp/decoder_jit_trace-pnnx.pt
+./ncnn/tools/pnnx/build/src/pnnx $repo/exp/joiner_jit_trace-pnnx.pt
+
+./conv_emformer_transducer_stateless2/streaming-ncnn-decode.py \
+ --tokens $repo/data/lang_bpe_500/tokens.txt \
+ --encoder-param-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.param \
+ --encoder-bin-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.bin \
+ --decoder-param-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.param \
+ --decoder-bin-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.bin \
+ --joiner-param-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.param \
+ --joiner-bin-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.bin \
+ $repo/test_wavs/1089-134686-0001.wav
diff --git a/...lstm-transducer-stateless2-2022-09-03.yml → ...-lstm-transducer-stateless2-2022-09-03.sh b/...lstm-transducer-stateless2-2022-09-03.yml → ...-lstm-transducer-stateless2-2022-09-03.sh
diff --git a/.github/workflows/run-librispeech-conv-emformer-transducer-stateless2-2022-12-05.yml b/.github/workflows/run-librispeech-conv-emformer-transducer-stateless2-2022-12-05.yml
@@ -0,0 +1,77 @@
+name: run-librispeech-conv-emformer-transducer-stateless2-2022-12-05
+
+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_conv_emformer_transducer_stateless2_2022_12_05:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'ncnn' || 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | grep -v kaldifst | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ 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: 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:$PYTHONPATH
+
+          .github/scripts/run-librispeech-conv-emformer-transducer-stateless2-2022-12-05.sh
diff --git a/.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml b/.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml
@@ -111,7 +111,7 @@ jobs:
           export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
           export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
 
-          .github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml
+          .github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.sh
 
       - name: Display decoding results for lstm_transducer_stateless2
         if: github.event_name == 'schedule'