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

egs/wenetspeech/KWS/run.sh

@@ -24,12 +24,17 @@ log() {
 
 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

egs/wenetspeech/KWS/zipformer/decode-asr.py

@@ -0,0 +1,812 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao
+#                                                 Mingshuang Luo)
+#
+# 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
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) modified beam search
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(3) fast beam search (trivial_graph)
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(4) fast beam search (LG)
+./zipformer/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest oracle WER)
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --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
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+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 lhotse.cut import Cut
+from train import add_model_arguments, get_model, get_params
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+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,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_char",
+        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
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest_oracle
+        If you use fast_beam_search_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, fast_beam_search_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_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--ilme-scale",
+        type=float,
+        default=0.2,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG.
+        It specifies the scale for the internal language model estimation.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search, fast_beam_search_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, fast_beam_search_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_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 and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--blank-penalty",
+        type=float,
+        default=0.0,
+        help="""
+        The penalty applied on blank symbol during decoding.
+        Note: It is a positive value that would be applied to logits like
+        this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
+        [batch_size, vocab] and blank id is 0).
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    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.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, 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.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    x, x_lens = model.encoder_embed(feature, feature_lens)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+    encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask)
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    hyps = []
+
+    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,
+            blank_penalty=params.blank_penalty,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "fast_beam_search_LG":
+        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,
+            blank_penalty=params.blank_penalty,
+            ilme_scale=params.ilme_scale,
+        )
+        for hyp in hyp_tokens:
+            sentence = "".join([lexicon.word_table[i] for i in hyp])
+            hyps.append(list(sentence))
+    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=graph_compiler.texts_to_ids(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+            blank_penalty=params.blank_penalty,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    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,
+            blank_penalty=params.blank_penalty,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            blank_penalty=params.blank_penalty,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    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,
+                    blank_penalty=params.blank_penalty,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                    blank_penalty=params.blank_penalty,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    key = f"blank_penalty_{params.blank_penalty}"
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search_" + key: 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"_ilme_scale_{params.ilme_scale}"
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}_" + key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[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.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, 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"]
+        texts = [list("".join(text.split())) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            graph_compiler=graph_compiler,
+            decoding_graph=decoding_graph,
+            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):
+                this_batch.append((cut_id, ref_text, 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[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, 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}-{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()
+    WenetSpeechAsrDataModule.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",
+        "modified_beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest_oracle",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "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"_ilme_scale_{params.ilme_scale}"
+            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}"
+    params.suffix += f"-blank-penalty-{params.blank_penalty}"
+
+    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}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            lexicon = Lexicon(params.lang_dir)
+            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:
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = 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
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    def remove_short_utt(c: Cut):
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        if T <= 0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
+            )
+        return T > 0
+
+    dev_cuts = wenetspeech.valid_cuts()
+    dev_cuts = dev_cuts.filter(remove_short_utt)
+    dev_dl = wenetspeech.valid_dataloaders(dev_cuts)
+
+    test_net_cuts = wenetspeech.test_net_cuts()
+    test_net_cuts = test_net_cuts.filter(remove_short_utt)
+    test_net_dl = wenetspeech.test_dataloaders(test_net_cuts)
+
+    test_meeting_cuts = wenetspeech.test_meeting_cuts()
+    test_meeting_cuts = test_meeting_cuts.filter(remove_short_utt)
+    test_meeting_dl = wenetspeech.test_dataloaders(test_meeting_cuts)
+
+    test_sets = ["DEV", "TEST_NET", "TEST_MEETING"]
+    test_dls = [dev_dl, test_net_dl, test_meeting_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            graph_compiler=graph_compiler,
+            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/wenetspeech/KWS/zipformer/export-onnx-streaming.py

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