add model export scripts

egs/gigaspeech/KWS/prepare.sh

@@ -0,0 +1,84 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=0
+stop_stage=100
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+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]}) $*"
+}
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Prepare gigaspeech dataset."
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.gigaspeech.done ]; then
+    pushd ../ASR
+    ./prepare.sh --stage 0 --stop-stage 9
+    ./prepare.sh --stage 11 --stop-stage 11
+    popd
+    pushd data/fbank
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_DEV.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_DEV.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_TEST.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_TEST.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_L.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_L.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_M.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_M.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_S.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_S.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_XS.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_XS.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/XL_split) .
+    ln -svf $(realpath ../ASR/data/fbank/musan_cuts.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/musan_feats) .
+    popd
+    pushd data
+    ln -svf $(realpath ../ASR/data/lang_bpe_500) .
+    popd
+  else
+    log "Gigaspeech dataset already exists, skipping."
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare open commands dataset."
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.fluent_speech_commands.done ]; then
+    pushd data
+    git clone https://github.com/pkufool/open-commands.git
+    ln -svf $(realpath ./open-commands/EN/small/commands.txt) commands_small.txt
+    ln -svf $(realpath ./open-commands/EN/large/commands.txt) commands_large.txt
+    pushd open-commands
+    ./script/prepare.sh --stage 3 --stop-stage 3 
+    ./script/prepare.sh --stage 6 --stop-stage 6
+    popd
+    popd
+    pushd data/fbank
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_cuts_large.jsonl.gz) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_feats_large) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_cuts_small.jsonl.gz) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_feats_small) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_cuts_valid.jsonl.gz) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_feats_valid) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_cuts_train.jsonl.gz) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_feats_train) .
+    popd
+    touch data/fbank/.fluent_speech_commands.done
+  else
+    log "Fluent speech commands dataset already exists, skipping."
+  fi
+fi

egs/gigaspeech/KWS/run.sh

@@ -0,0 +1,202 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+export PYTHONPATH=../../../:$PYTHONPATH
+
+stage=0
+stop_stage=100
+
+pre_trained_model_host=github
+
+. shared/parse_options.sh || exit 1
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download a pre-trained model."
+  if [ $pre_trained_model_host -eq "github" ]; then
+
+  elif [$pre_trained_model_host -eq "modelscope" ]; then
+
+  else
+    log "Pretrained model host : $pre_trained_model_host not support."
+    exit -1;
+  fi
+fi
+
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Train a model."
+  if [ ! -e data/fbank/.gigaspeech.done ]; then
+    log "You need to run the prepare.sh first."
+    exit -1
+  fi
+
+  python ./zipformer/train.py \
+      --world-size 4 \
+      --exp-dir zipformer/exp \
+      --decoder-dim 320 \
+      --joiner-dim 320 \
+      --num-encoder-layers 1,1,1,1,1,1 \
+      --feedforward-dim 192,192,192,192,192,192 \
+      --encoder-dim 128,128,128,128,128,128 \
+      --encoder-unmasked-dim 128,128,128,128,128,128 \
+      --num-epochs 15 \
+      --lr-epochs 1.5 \
+      --use-fp16 1 \
+      --start-epoch 1 \
+      --training-subset L \
+      --pinyin-type partial_with_tone \
+      --causal 1 \
+      --lang-dir data/lang_partial_tone \
+      --max-duration 1000
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Decode the model."
+  for t in small, large; do
+    python ./zipformer/decode.py \
+        --epoch 15 \
+        --avg 2 \
+        --exp-dir ./zipformer/exp \
+        --lang-dir ./data/lang_partial_tone \
+        --pinyin-type partial_with_tone \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 64 \
+        --decoder-dim 320 \
+        --joiner-dim 320 \
+        --num-encoder-layers 1,1,1,1,1,1 \
+        --feedforward-dim 192,192,192,192,192,192 \
+        --encoder-dim 128,128,128,128,128,128 \
+        --encoder-unmasked-dim 128,128,128,128,128,128 \
+        --test-set $t \
+        --keywords-score 1.0 \
+        --keywords-threshold 0.35 \
+        --keywords-file ./data/commands_${t}.txt  \
+        --max-duration 3000
+  done
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Export the model."
+
+  python ./zipformer/export.py \
+      --epoch 15 \
+      --avg 2 \
+      --exp-dir ./zipformer/exp \
+      --tokens data/lang_partial_tone/tokens.txt \
+      --causal 1 \
+      --chunk-size 16 \
+      --left-context-frames 64 \
+      --decoder-dim 320 \
+      --joiner-dim 320 \
+      --num-encoder-layers 1,1,1,1,1,1 \
+      --feedforward-dim 192,192,192,192,192,192 \
+      --encoder-dim 128,128,128,128,128,128 \
+      --encoder-unmasked-dim 128,128,128,128,128,128
+
+  python ./zipformer/export_onnx_streaming.py \
+    --exp-dir zipformer/exp \
+    --tokens data/lang_partial_tone/tokens.txt \
+    --epoch 15 \
+    --avg 2 \
+    --chunk-size 16 \
+    --left-context-frames 128 \
+    --decoder-dim 320 \
+    --joiner-dim 320 \
+    --num-encoder-layers 1,1,1,1,1,1 \
+    --feedforward-dim 192,192,192,192,192,192 \
+    --encoder-dim 128,128,128,128,128,128 \
+    --encoder-unmasked-dim 128,128,128,128,128,128 \
+    --causal 1
+fi 
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 2: Finetune the model"
+
+  # The following configuration of lr schedule should work well
+  # You may also tune the following parameters to adjust learning rate schedule
+  base_lr=0.0005
+  lr_epochs=100
+  lr_batches=100000
+
+  # We recommend to start from an averaged model
+  finetune_ckpt=zipformer/exp/pretrained.pt
+
+  ./zipformer/finetune.py \
+    --world-size 4 \
+    --num-epochs 10 \
+    --start-epoch 1 \
+    --exp-dir zipformer/exp_finetune
+    --lang-dir ./data/lang_partial_tone \
+    --pinyin-type partial_with_tone \
+    --use-fp16 1 \
+    --decoder-dim 320 \
+    --joiner-dim 320 \
+    --num-encoder-layers 1,1,1,1,1,1 \
+    --feedforward-dim 192,192,192,192,192,192 \
+    --encoder-dim 128,128,128,128,128,128 \
+    --encoder-unmasked-dim 128,128,128,128,128,128 \
+    --causal 1 \
+    --base-lr $base_lr \
+    --lr-epochs $lr_epochs \
+    --lr-batches $lr_batches \
+    --finetune-ckpt $finetune_ckpt \
+    --max-duration 1500
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 1: Decode the finetuned model."
+  for t in small, large; do
+    python ./zipformer/decode.py \
+        --epoch 15 \
+        --avg 2 \
+        --exp-dir ./zipformer/exp_finetune \
+        --lang-dir ./data/lang_partial_tone \
+        --pinyin-type partial_with_tone \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 64 \
+        --decoder-dim 320 \
+        --joiner-dim 320 \
+        --num-encoder-layers 1,1,1,1,1,1 \
+        --feedforward-dim 192,192,192,192,192,192 \
+        --encoder-dim 128,128,128,128,128,128 \
+        --encoder-unmasked-dim 128,128,128,128,128,128 \
+        --test-set $t \
+        --keywords-score 1.0 \
+        --keywords-threshold 0.35 \
+        --keywords-file ./data/commands_${t}.txt  \
+        --max-duration 3000
+  done
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 2: Export the finetuned model."
+
+  python ./zipformer/export_onnx_streaming.py \
+    --exp-dir zipformer/exp_finetune \
+    --tokens data/lang_partial_tone/tokens.txt \
+    --epoch 15 \
+    --avg 2 \
+    --chunk-size 16 \
+    --left-context-frames 128 \
+    --decoder-dim 320 \
+    --joiner-dim 320 \
+    --num-encoder-layers 1,1,1,1,1,1 \
+    --feedforward-dim 192,192,192,192,192,192 \
+    --encoder-dim 128,128,128,128,128,128 \
+    --encoder-unmasked-dim 128,128,128,128,128,128 \
+    --causal 1
+fi 

egs/gigaspeech/KWS/shared

@@ -0,0 +1 @@
+../../../icefall/shared

egs/gigaspeech/KWS/zipformer/export-onnx-streaming.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-streaming.py

egs/gigaspeech/KWS/zipformer/export.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export.py

egs/wenetspeech/KWS/prepare.sh

@@ -0,0 +1,84 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=0
+stop_stage=100
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+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]}) $*"
+}
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Prepare gigaspeech dataset."
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.gigaspeech.done ]; then
+    pushd ../ASR
+    ./prepare.sh --stage 0 --stop-stage 9
+    ./prepare.sh --stage 11 --stop-stage 11
+    popd
+    pushd data/fbank
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_DEV.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_DEV.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_TEST.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_TEST.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_L.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_L.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_M.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_M.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_S.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_S.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_cuts_XS.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/gigaspeech_feats_XS.lca) .
+    ln -svf $(realpath ../ASR/data/fbank/XL_split) .
+    ln -svf $(realpath ../ASR/data/fbank/musan_cuts.jsonl.gz) .
+    ln -svf $(realpath ../ASR/data/fbank/musan_feats) .
+    popd
+    pushd data
+    ln -svf $(realpath ../ASR/data/lang_bpe_500) .
+    popd
+  else
+    log "Gigaspeech dataset already exists, skipping."
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare open commands dataset."
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.fluent_speech_commands.done ]; then
+    pushd data
+    git clone https://github.com/pkufool/open-commands.git
+    ln -svf $(realpath ./open-commands/EN/small/commands.txt) commands_small.txt
+    ln -svf $(realpath ./open-commands/EN/large/commands.txt) commands_large.txt
+    pushd open-commands
+    ./script/prepare.sh --stage 3 --stop-stage 3 
+    ./script/prepare.sh --stage 6 --stop-stage 6
+    popd
+    popd
+    pushd data/fbank
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_cuts_large.jsonl.gz) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_feats_large) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_cuts_small.jsonl.gz) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_feats_small) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_cuts_valid.jsonl.gz) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_feats_valid) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_cuts_train.jsonl.gz) .
+    ln -svf $(realpath ../open-commands/data/fbank/fluent_speech_commands_feats_train) .
+    popd
+    touch data/fbank/.fluent_speech_commands.done
+  else
+    log "Fluent speech commands dataset already exists, skipping."
+  fi
+fi