run.sh

#!/bin/bash

# Copyright 2019 Mobvoi Inc. All Rights Reserved.
#           2022 Binbin Zhang(binbizha@qq.com)

. ./path.sh || exit 1;

# Use this to control how many gpu you use, It's 1-gpu training if you specify
# just 1gpu, otherwise it's is multiple gpu training based on DDP in pytorch
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"

stage=0 # start from 0 if you need to start from data preparation
stop_stage=5

# The num of machines(nodes) for multi-machine training, 1 is for one machine.
# NFS is required if num_nodes > 1.
num_nodes=1

# The rank of each node or machine, which ranges from 0 to `num_nodes - 1`.
# You should set the node_rank=0 on the first machine, set the node_rank=1
# on the second machine, and so on.
node_rank=0
# The aishell dataset location, please change this to your own path
# make sure of using absolute path. DO-NOT-USE relatvie path!
data=/export/data/asr-data/OpenSLR/33/
data_url=www.openslr.org/resources/33

nj=16
dict=data/dict/lang_char.txt

# data_type can be `raw` or `shard`. Typically, raw is used for small dataset,
# `shard` is used for large dataset which is over 1k hours, and `shard` is
# faster on reading data and training.
data_type=raw
num_utts_per_shard=1000

train_set=train
train_config=conf/conformer_u2pp_rnnt.yaml
cmvn=true
dir=exp/conformer_rnnt
checkpoint=

# use average_checkpoint will get better result
average_checkpoint=true
decode_checkpoint=$dir/final.pt
average_num=30
decode_modes="rnnt_beam_search"

. tools/parse_options.sh || exit 1;

if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
  echo "stage -1: Data Download"
  local/download_and_untar.sh ${data} ${data_url} data_aishell
  local/download_and_untar.sh ${data} ${data_url} resource_aishell
fi

if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
  # Data preparation
  local/aishell_data_prep.sh ${data}/data_aishell/wav \
    ${data}/data_aishell/transcript
fi


if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
  # remove the space between the text labels for Mandarin dataset
  for x in train dev test; do
    cp data/${x}/text data/${x}/text.org
    paste -d " " <(cut -f 1 -d" " data/${x}/text.org) \
      <(cut -f 2- -d" " data/${x}/text.org | tr -d " ") \
      > data/${x}/text
    rm data/${x}/text.org
  done

  tools/compute_cmvn_stats.py --num_workers 16 --train_config $train_config \
    --in_scp data/${train_set}/wav.scp \
    --out_cmvn data/$train_set/global_cmvn
fi

if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
  echo "Make a dictionary"
  mkdir -p $(dirname $dict)
  echo "<blank> 0" > ${dict}  # 0 is for "blank" in CTC
  echo "<unk> 1"  >> ${dict}  # <unk> must be 1
  tools/text2token.py -s 1 -n 1 data/train/text | cut -f 2- -d" " \
    | tr " " "\n" | sort | uniq | grep -a -v -e '^\s*$' | \
    awk '{print $0 " " NR+1}' >> ${dict}
  num_token=$(cat $dict | wc -l)
  echo "<sos/eos> $num_token" >> $dict
fi

if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
  echo "Prepare data, prepare required format"
  for x in dev test ${train_set}; do
    if [ $data_type == "shard" ]; then
      tools/make_shard_list.py --num_utts_per_shard $num_utts_per_shard \
        --num_threads 16 data/$x/wav.scp data/$x/text \
        $(realpath data/$x/shards) data/$x/data.list
    else
      tools/make_raw_list.py data/$x/wav.scp data/$x/text \
        data/$x/data.list
    fi
  done
fi

if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
  mkdir -p $dir
  # You have to rm `INIT_FILE` manually when you resume or restart a
  # multi-machine training.
  INIT_FILE=$dir/ddp_init
  init_method=file://$(readlink -f $INIT_FILE)
  echo "$0: init method is $init_method"
  num_gpus=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
  # Use "nccl" if it works, otherwise use "gloo"
  dist_backend="gloo"
  world_size=`expr $num_gpus \* $num_nodes`
  echo "total gpus is: $world_size"
  cmvn_opts=
  $cmvn && cp data/${train_set}/global_cmvn $dir
  $cmvn && cmvn_opts="--cmvn ${dir}/global_cmvn"

  # train.py rewrite $train_config to $dir/train.yaml with model input
  # and output dimension, and $dir/train.yaml will be used for inference
  # and export.
  for ((i = 0; i < $num_gpus; ++i)); do
  {
    gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
    # Rank of each gpu/process used for knowing whether it is
    # the master of a worker.
    rank=`expr $node_rank \* $num_gpus + $i`
    python wenet/bin/train.py --gpu $gpu_id \
      --config $train_config \
      --data_type $data_type \
      --symbol_table $dict \
      --train_data data/$train_set/data.list \
      --cv_data data/dev/data.list \
      ${checkpoint:+--checkpoint $checkpoint} \
      --model_dir $dir \
      --ddp.init_method $init_method \
      --ddp.world_size $world_size \
      --ddp.rank $rank \
      --ddp.dist_backend $dist_backend \
      --num_workers 1 \
      $cmvn_opts \
      --pin_memory
  } &
  done
  wait
fi

if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
  # Test model, please specify the model you want to test by --checkpoint
  if [ ${average_checkpoint} == true ]; then
    decode_checkpoint=$dir/avg_${average_num}.pt
    echo "do model average and final checkpoint is $decode_checkpoint"
    python wenet/bin/average_model.py \
      --dst_model $decode_checkpoint \
      --src_path $dir  \
      --num ${average_num} \
      --val_best
  fi
  # Please specify decoding_chunk_size for unified streaming and
  # non-streaming model. The default value is -1, which is full chunk
  # for non-streaming inference.
  decoding_chunk_size=
  # only used in rescore mode for weighting different scores
  rescore_ctc_weight=0.5
  rescore_transducer_weight=0.5
  rescore_attn_weight=0.5
  # only used in beam search, either pure beam search mode OR beam search inside rescoring
  search_ctc_weight=0.3
  search_transducer_weight=0.7

  reverse_weight=0.0
  for mode in ${decode_modes}; do
  {
    test_dir=$dir/test_${mode}
    mkdir -p $test_dir
    python wenet/bin/recognize.py --gpu 0 \
      --mode $mode \
      --config $dir/train.yaml \
      --data_type $data_type \
      --test_data data/test/data.list \
      --checkpoint $decode_checkpoint \
      --beam_size 10 \
      --batch_size 1 \
      --penalty 0.0 \
      --dict $dict \
      --ctc_weight $rescore_ctc_weight \
      --transducer_weight $rescore_transducer_weight \
      --attn_weight $rescore_attn_weight \
      --search_ctc_weight $search_ctc_weight \
      --search_transducer_weight $search_transducer_weight \
      --reverse_weight $reverse_weight \
      --result_file $test_dir/text \
      ${decoding_chunk_size:+--decoding_chunk_size $decoding_chunk_size}
    python tools/compute-wer.py --char=1 --v=1 \
      data/test/text $test_dir/text > $test_dir/wer
  } &
  done
  wait
fi