add new model

examples/aishell/s0/conf/train_u2++_efficonformer_v1.yaml

+# network architecture
+# encoder related
+encoder: efficientConformer
+encoder_conf:
+    activation_type: 'swish'
+    attention_heads: 8
+    causal: false
+    cnn_module_kernel: 15
+    cnn_module_norm: 'layer_norm'
+    dropout_rate: 0.1
+    input_layer: conv2d
+    linear_units: 2048
+    normalize_before: true
+    num_blocks: 12
+    output_size: 256
+    pos_enc_layer_type: 'rel_pos'
+    attention_dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    use_cnn_module: true
+    use_dynamic_chunk: true
+    use_dynamic_left_chunk: false
+    efficient_conf:
+        stride_layer_idx: [3]           # layer id with StrideConv
+        stride: [2]                     # stride size of each StrideConv
+        group_layer_idx: [0, 1, 2, 3]   # layer id with GroupedAttention
+        group_size: 3                   # group size of every GroupedAttention layer
+        stride_kernel: true             # true: recompute cnn kernels with stride
+
+# decoder related
+decoder: bitransformer
+decoder_conf:
+    attention_heads: 8
+    dropout_rate: 0.1
+    linear_units: 2048
+    num_blocks: 3
+    r_num_blocks: 3
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.1
+    src_attention_dropout_rate: 0.1
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+    reverse_weight: 0.3
+
+# dataset related
+dataset_conf:
+    batch_conf:
+        batch_size: 16
+        batch_type: 'static'
+    fbank_conf:
+        num_mel_bins: 80
+        frame_shift: 10
+        frame_length: 25
+        dither: 1.0
+    filter_conf:
+        max_length: 40960
+        min_length: 0
+        token_max_length: 200
+        token_min_length: 1
+    resample_conf:
+        resample_rate: 16000
+    shuffle: true
+    shuffle_conf:
+        shuffle_size: 1500
+    sort: true
+    sort_conf:
+        sort_size: 500
+    spec_aug: true
+    spec_aug_conf:
+        num_t_mask: 2
+        num_f_mask: 2
+        max_t: 50
+        max_f: 10
+    spec_sub: true
+    spec_sub_conf:
+        num_t_sub: 3
+        max_t: 30
+    spec_trim: false
+    spec_trim_conf:
+        max_t: 50
+    speed_perturb: true
+
+grad_clip: 5
+accum_grad: 1
+max_epoch: 200
+log_interval: 100
+
+optim: adam
+optim_conf:
+    lr: 0.001
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000

examples/aishell/s0/conf/train_u2++_efficonformer_v2.yaml

+# network architecture
+# encoder related
+encoder: efficientConformer
+encoder_conf:
+    activation_type: 'swish'
+    attention_heads: 8
+    causal: false
+    cnn_module_kernel: 15
+    cnn_module_norm: 'layer_norm'
+    dropout_rate: 0.1
+    input_layer: conv2d2
+    linear_units: 2048
+    normalize_before: true
+    num_blocks: 12
+    output_size: 256
+    pos_enc_layer_type: 'rel_pos'
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.1
+    use_cnn_module: true
+    use_dynamic_chunk: true
+    use_dynamic_left_chunk: false
+    efficient_conf:
+        stride_layer_idx: [3, 7]           # layer id with StrideConv
+        stride: [2, 2]                     # stride size of each StrideConv
+        group_layer_idx: [3, 7]            # layer id with GroupedAttention
+        group_size: 3                      # group size of every GroupedAttention layer
+        stride_kernel: false               # true: recompute cnn kernels with stride
+
+# decoder related
+decoder: bitransformer
+decoder_conf:
+    attention_heads: 8
+    dropout_rate: 0.1
+    linear_units: 2048
+    num_blocks: 3
+    positional_dropout_rate: 0.1
+    r_num_blocks: 3
+    self_attention_dropout_rate: 0.1
+    src_attention_dropout_rate: 0.1
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+    reverse_weight: 0.3
+
+# dataset related
+dataset_conf:
+    batch_conf:
+        batch_size: 16
+        batch_type: 'static'
+    fbank_conf:
+        dither: 1.0
+        frame_length: 25
+        frame_shift: 10
+        num_mel_bins: 80
+    filter_conf:
+        max_length: 40960
+        min_length: 0
+        max_output_input_ratio: 0.1
+        min_output_input_ratio: 0.005
+        token_max_length: 200
+        token_min_length: 1
+    resample_conf:
+        resample_rate: 16000
+    shuffle: true
+    shuffle_conf:
+        shuffle_size: 1500
+    sort: true
+    sort_conf:
+        sort_size: 500
+    spec_aug: true
+    spec_aug_conf:
+        num_t_mask: 2
+        num_f_mask: 2
+        max_t: 50
+        max_f: 10
+    spec_sub: true
+    spec_sub_conf:
+        num_t_sub: 3
+        max_t: 30
+    spec_trim: false
+    spec_trim_conf:
+        max_t: 50
+    speed_perturb: true
+
+grad_clip: 5
+accum_grad: 1
+max_epoch: 200
+log_interval: 100
+
+optim: adam
+optim_conf:
+    lr: 0.001
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000

examples/aishell/s0/conf/train_u2++_transformer.yaml

+# network architecture
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder architecture type
+    normalize_before: true
+    use_dynamic_chunk: true
+    use_dynamic_left_chunk: false
+
+# decoder related
+decoder: bitransformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 3
+    r_num_blocks: 3
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+    reverse_weight: 0.3
+
+dataset_conf:
+    filter_conf:
+        max_length: 40960
+        min_length: 0
+        token_max_length: 200
+        token_min_length: 1
+    resample_conf:
+        resample_rate: 16000
+    speed_perturb: true
+    fbank_conf:
+        num_mel_bins: 80
+        frame_shift: 10
+        frame_length: 25
+        dither: 1.0
+    spec_aug: true
+    spec_aug_conf:
+        num_t_mask: 2
+        num_f_mask: 2
+        max_t: 50
+        max_f: 10
+    spec_trim: false
+    spec_trim_conf:
+        max_t: 70
+    shuffle: true
+    shuffle_conf:
+        shuffle_size: 1500
+    sort: true
+    sort_conf:
+        sort_size: 500  # sort_size should be less than shuffle_size
+    batch_conf:
+        batch_type: 'static' # static or dynamic
+        batch_size: 26
+
+grad_clip: 5
+accum_grad: 1
+max_epoch: 360
+log_interval: 100
+
+optim: adam
+optim_conf:
+    lr: 0.001
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000

examples/aishell/s0/conf/train_unified_conformer.yaml

+# network architecture
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
+    cnn_module_kernel: 15
+    use_cnn_module: True
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
+    causal: true
+    use_dynamic_chunk: true
+    cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
+    use_dynamic_left_chunk: false
+
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+
+dataset_conf:
+    filter_conf:
+        max_length: 40960
+        min_length: 0
+        token_max_length: 200
+        token_min_length: 1
+    resample_conf:
+        resample_rate: 16000
+    speed_perturb: true
+    fbank_conf:
+        num_mel_bins: 80
+        frame_shift: 10
+        frame_length: 25
+        dither: 0.1
+    spec_aug: true
+    spec_aug_conf:
+        num_t_mask: 2
+        num_f_mask: 2
+        max_t: 50
+        max_f: 10
+    shuffle: true
+    shuffle_conf:
+        shuffle_size: 1500
+    sort: true
+    sort_conf:
+        sort_size: 500  # sort_size should be less than shuffle_size
+    batch_conf:
+        batch_type: 'static' # static or dynamic
+        batch_size: 16
+
+grad_clip: 5
+accum_grad: 1
+max_epoch: 180
+log_interval: 100
+
+optim: adam
+optim_conf:
+    lr: 0.001
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000

examples/aishell/s0/conf/train_unified_transformer.yaml

+# network architecture
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder architecture type
+    normalize_before: true
+    use_dynamic_chunk: true
+    use_dynamic_left_chunk: false
+
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+
+dataset_conf:
+    filter_conf:
+        max_length: 40960
+        min_length: 0
+        token_max_length: 200
+        token_min_length: 1
+    resample_conf:
+        resample_rate: 16000
+    speed_perturb: true
+    fbank_conf:
+        num_mel_bins: 80
+        frame_shift: 10
+        frame_length: 25
+        dither: 0.1
+    spec_aug: true
+    spec_aug_conf:
+        num_t_mask: 2
+        num_f_mask: 2
+        max_t: 50
+        max_f: 10
+    shuffle: true
+    shuffle_conf:
+        shuffle_size: 1500
+    sort: true
+    sort_conf:
+        sort_size: 500  # sort_size should be less than shuffle_size
+    batch_conf:
+        batch_type: 'static' # static or dynamic
+        batch_size: 16
+
+
+grad_clip: 5
+accum_grad: 1
+max_epoch: 180
+log_interval: 100
+
+optim: adam
+optim_conf:
+    lr: 0.002
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000

examples/aishell/s0/local/aishell_data_prep.sh

+#!/bin/bash
+
+# Copyright 2017 Xingyu Na
+# Apache 2.0
+
+. ./path.sh || exit 1;
+
+if [ $# != 2 ]; then
+  echo "Usage: $0 <audio-path> <text-path>"
+  echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript"
+  exit 1;
+fi
+
+aishell_audio_dir=$1
+aishell_text=$2/aishell_transcript_v0.8.txt
+
+train_dir=data/local/train
+dev_dir=data/local/dev
+test_dir=data/local/test
+tmp_dir=data/local/tmp
+
+mkdir -p $train_dir
+mkdir -p $dev_dir
+mkdir -p $test_dir
+mkdir -p $tmp_dir
+
+# data directory check
+if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
+  echo "Error: $0 requires two directory arguments"
+  exit 1;
+fi
+
+# find wav audio file for train, dev and test resp.
+find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
+n=`cat $tmp_dir/wav.flist | wc -l`
+[ $n -ne 141925 ] && \
+  echo Warning: expected 141925 data data files, found $n
+
+grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
+grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
+grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
+
+rm -r $tmp_dir
+
+# Transcriptions preparation
+for dir in $train_dir $dev_dir $test_dir; do
+  echo Preparing $dir transcriptions
+  sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
+  paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
+  tools/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
+  awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
+  tools/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
+  sort -u $dir/transcripts.txt > $dir/text
+done
+
+mkdir -p data/train data/dev data/test
+
+for f in wav.scp text; do
+  cp $train_dir/$f data/train/$f || exit 1;
+  cp $dev_dir/$f data/dev/$f || exit 1;
+  cp $test_dir/$f data/test/$f || exit 1;
+done
+
+echo "$0: AISHELL data preparation succeeded"
+exit 0;

examples/aishell/s0/local/aishell_train_lms.sh

+#!/bin/bash
+
+
+# To be run from one directory above this script.
+. ./path.sh
+
+text=data/local/lm/text
+lexicon=data/local/dict/lexicon.txt
+
+for f in "$text" "$lexicon"; do
+  [ ! -f $x ] && echo "$0: No such file $f" && exit 1;
+done
+
+# Check SRILM tools
+if ! which ngram-count > /dev/null; then
+    echo "srilm tools are not found, please download it and install it from: "
+    echo "http://www.speech.sri.com/projects/srilm/download.html"
+    echo "Then add the tools to your PATH"
+    exit 1
+fi
+
+# This script takes no arguments.  It assumes you have already run
+# aishell_data_prep.sh.
+# It takes as input the files
+# data/local/lm/text
+# data/local/dict/lexicon.txt
+dir=data/local/lm
+mkdir -p $dir
+
+
+cleantext=$dir/text.no_oov
+
+cat $text | awk -v lex=$lexicon 'BEGIN{while((getline<lex) >0){ seen[$1]=1; } }
+  {for(n=1; n<=NF;n++) {  if (seen[$n]) { printf("%s ", $n); } else {printf("<SPOKEN_NOISE> ");} } printf("\n");}' \
+  > $cleantext || exit 1;
+
+cat $cleantext | awk '{for(n=2;n<=NF;n++) print $n; }' | sort | uniq -c | \
+   sort -nr > $dir/word.counts || exit 1;
+
+# Get counts from acoustic training transcripts, and add  one-count
+# for each word in the lexicon (but not silence, we don't want it
+# in the LM-- we'll add it optionally later).
+cat $cleantext | awk '{for(n=2;n<=NF;n++) print $n; }' | \
+  cat - <(grep -w -v '!SIL' $lexicon | awk '{print $1}') | \
+   sort | uniq -c | sort -nr > $dir/unigram.counts || exit 1;
+
+cat $dir/unigram.counts | awk '{print $2}' | cat - <(echo "<s>"; echo "</s>" ) > $dir/wordlist
+
+heldout_sent=10000 # Don't change this if you want result to be comparable with
+    # kaldi_lm results
+mkdir -p $dir
+cat $cleantext | awk '{for(n=2;n<=NF;n++){ printf $n; if(n<NF) printf " "; else print ""; }}' | \
+  head -$heldout_sent > $dir/heldout
+cat $cleantext | awk '{for(n=2;n<=NF;n++){ printf $n; if(n<NF) printf " "; else print ""; }}' | \
+  tail -n +$heldout_sent > $dir/train
+
+ngram-count -text $dir/train -order 3 -limit-vocab -vocab $dir/wordlist -unk \
+  -map-unk "<UNK>" -kndiscount -interpolate -lm $dir/lm.arpa
+ngram -lm $dir/lm.arpa -ppl $dir/heldout

examples/aishell/s0/local/download_and_untar.sh

+#!/bin/bash
+
+# Copyright   2014  Johns Hopkins University (author: Daniel Povey)
+#             2017  Xingyu Na
+# Apache 2.0
+
+remove_archive=false
+
+if [ "$1" == --remove-archive ]; then
+  remove_archive=true
+  shift
+fi
+
+if [ $# -ne 3 ]; then
+  echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
+  echo "e.g.: $0 /export/a05/xna/data www.openslr.org/resources/33 data_aishell"
+  echo "With --remove-archive it will remove the archive after successfully un-tarring it."
+  echo "<corpus-part> can be one of: data_aishell, resource_aishell."
+fi
+
+data=$1
+url=$2
+part=$3
+
+if [ ! -d "$data" ]; then
+  echo "$0: no such directory $data"
+  exit 1;
+fi
+
+part_ok=false
+list="data_aishell resource_aishell"
+for x in $list; do
+  if [ "$part" == $x ]; then part_ok=true; fi
+done
+if ! $part_ok; then
+  echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
+  exit 1;
+fi
+
+if [ -z "$url" ]; then
+  echo "$0: empty URL base."
+  exit 1;
+fi
+
+if [ -f $data/$part/.complete ]; then
+  echo "$0: data part $part was already successfully extracted, nothing to do."
+  exit 0;
+fi
+
+# sizes of the archive files in bytes.
+sizes="15582913665 1246920"
+
+if [ -f $data/$part.tgz ]; then
+  size=$(/bin/ls -l $data/$part.tgz | awk '{print $5}')
+  size_ok=false
+  for s in $sizes; do if [ $s == $size ]; then size_ok=true; fi; done
+  if ! $size_ok; then
+    echo "$0: removing existing file $data/$part.tgz because its size in bytes $size"
+    echo "does not equal the size of one of the archives."
+    rm $data/$part.tgz
+  else
+    echo "$data/$part.tgz exists and appears to be complete."
+  fi
+fi
+
+if [ ! -f $data/$part.tgz ]; then
+  if ! which wget >/dev/null; then
+    echo "$0: wget is not installed."
+    exit 1;
+  fi
+  full_url=$url/$part.tgz
+  echo "$0: downloading data from $full_url.  This may take some time, please be patient."
+
+  cd $data
+  if ! wget --no-check-certificate $full_url; then
+    echo "$0: error executing wget $full_url"
+    exit 1;
+  fi
+fi
+
+cd $data
+
+if ! tar -xvzf $part.tgz; then
+  echo "$0: error un-tarring archive $data/$part.tgz"
+  exit 1;
+fi
+
+touch $data/$part/.complete
+
+if [ $part == "data_aishell" ]; then
+  cd $data/$part/wav
+  for wav in ./*.tar.gz; do
+    echo "Extracting wav from $wav"
+    tar -zxf $wav && rm $wav
+  done
+fi
+
+echo "$0: Successfully downloaded and un-tarred $data/$part.tgz"
+
+if $remove_archive; then
+  echo "$0: removing $data/$part.tgz file since --remove-archive option was supplied."
+  rm $data/$part.tgz
+fi
+
+exit 0;

examples/aishell/s0/path.sh

+export WENET_DIR=$PWD/../../..
+export BUILD_DIR=${WENET_DIR}/runtime/libtorch/build
+export OPENFST_BIN=${BUILD_DIR}/../fc_base/openfst-build/src
+export PATH=$PWD:${BUILD_DIR}/bin:${BUILD_DIR}/kaldi:${OPENFST_BIN}/bin:$PATH
+
+# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
+export PYTHONIOENCODING=UTF-8
+export PYTHONPATH=../../../:$PYTHONPATH

examples/aishell/s0/run.sh

+#!/bin/bash
+
+# Copyright 2019 Mobvoi Inc. All Rights Reserved.
+. ./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"
+# The NCCL_SOCKET_IFNAME variable specifies which IP interface to use for nccl
+# communication. More details can be found in
+# https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html
+# export NCCL_SOCKET_IFNAME=ens4f1
+export NCCL_DEBUG=INFO
+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=/data/conformer/train/
+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
+# Optional train_config
+# 1. conf/train_transformer.yaml: Standard transformer
+# 2. conf/train_conformer.yaml: Standard conformer
+# 3. conf/train_unified_conformer.yaml: Unified dynamic chunk causal conformer
+# 4. conf/train_unified_transformer.yaml: Unified dynamic chunk transformer
+# 5. conf/train_u2++_conformer.yaml: U2++ conformer
+# 6. conf/train_u2++_transformer.yaml: U2++ transformer
+train_config=conf/train_conformer.yaml
+cmvn=true
+dir=exp/conformer
+checkpoint=
+
+# use average_checkpoint will get better result
+average_checkpoint=true
+decode_checkpoint=$dir/final.pt
+average_num=30
+decode_modes="attention_rescoring"
+
+. 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="nccl"
+  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=
+  ctc_weight=0.5
+  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 $ctc_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
+
+
+if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
+  # Export the best model you want
+  python wenet/bin/export_jit.py \
+    --config $dir/train.yaml \
+    --checkpoint $dir/avg_${average_num}.pt \
+    --output_file $dir/final.zip \
+    --output_quant_file $dir/final_quant.zip
+fi
+
+# Optionally, you can add LM and test it with runtime.
+if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
+  # 7.1 Prepare dict
+  unit_file=$dict
+  mkdir -p data/local/dict
+  cp $unit_file data/local/dict/units.txt
+  tools/fst/prepare_dict.py $unit_file ${data}/resource_aishell/lexicon.txt \
+    data/local/dict/lexicon.txt
+  # 7.2 Train lm
+  lm=data/local/lm
+  mkdir -p $lm
+  tools/filter_scp.pl data/train/text \
+    $data/data_aishell/transcript/aishell_transcript_v0.8.txt > $lm/text
+  local/aishell_train_lms.sh
+  # 7.3 Build decoding TLG
+  tools/fst/compile_lexicon_token_fst.sh \
+    data/local/dict data/local/tmp data/local/lang
+  tools/fst/make_tlg.sh data/local/lm data/local/lang data/lang_test || exit 1;
+  # 7.4 Decoding with runtime
+  chunk_size=-1
+  ./tools/decode.sh --nj 16 \
+    --beam 15.0 --lattice_beam 7.5 --max_active 7000 \
+    --blank_skip_thresh 0.98 --ctc_weight 0.5 --rescoring_weight 1.0 \
+    --chunk_size $chunk_size \
+    --fst_path data/lang_test/TLG.fst \
+    --dict_path data/lang_test/words.txt \
+    data/test/wav.scp data/test/text $dir/final.zip \
+    data/lang_test/units.txt $dir/lm_with_runtime
+  # Please see $dir/lm_with_runtime for wer
+fi
+
+# Optionally, you can decode with k2 hlg
+if [ ${stage} -le 8 ] && [ ${stop_stage} -ge 8 ]; then
+  if [ ! -f data/local/lm/lm.arpa ]; then
+    echo "Please run prepare dict and train lm in Stage 7" || exit 1;
+  fi
+
+  # 8.1 Build decoding HLG
+  required="data/local/hlg/HLG.pt data/local/hlg/words.txt"
+  for f in $required; do
+    if [ ! -f $f ]; then
+      tools/k2/make_hlg.sh data/local/dict/ data/local/lm/ data/local/hlg
+      break
+    fi
+  done
+
+  # 8.2 Decode using HLG
+  decoding_chunk_size=
+  lm_scale=0.7
+  decoder_scale=0.1
+  r_decoder_scale=0.7
+  for mode in hlg_onebest hlg_rescore; do
+  {
+    test_dir=$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 16 \
+      --penalty 0.0 \
+      --dict $dict \
+      --word data/local/hlg/words.txt \
+      --hlg data/local/hlg/HLG.pt \
+      --lm_scale $lm_scale \
+      --decoder_scale $decoder_scale \
+      --r_decoder_scale $r_decoder_scale \
+      --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
+fi
+

examples/aishell/s0/run_train.sh

+#!/usr/bin/env bash
+
+# This training script is modified from the original run.sh
+
+# Copyright 2019 Mobvoi Inc. All Rights Reserved.
+. ./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"
+# The NCCL_SOCKET_IFNAME variable specifies which IP interface to use for nccl
+# communication. More details can be found in
+# https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html
+# export NCCL_SOCKET_IFNAME=ens4f1
+export NCCL_DEBUG=INFO
+stage=4 # start from 0 if you need to start from data preparation
+stop_stage=4
+
+# 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=/data/conformer/train/
+data_url=www.openslr.org/resources/33
+
+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
+# Optional train_config
+# 1. conf/train_transformer.yaml: Standard transformer
+# 2. conf/train_conformer.yaml: Standard conformer
+# 3. conf/train_unified_conformer.yaml: Unified dynamic chunk causal conformer
+# 4. conf/train_unified_transformer.yaml: Unified dynamic chunk transformer
+# 5. conf/train_u2++_conformer.yaml: U2++ conformer
+# 6. conf/train_u2++_transformer.yaml: U2++ transformer
+train_config=conf/train_conformer.yaml
+cmvn=true
+dir=exp/conformer
+checkpoint=
+
+# use average_checkpoint will get better result
+average_checkpoint=false
+decode_checkpoint=$dir/final.pt
+average_num=30
+decode_modes="attention_rescoring"
+log_dir=
+
+. tools/parse_options.sh || exit 1;
+
+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="nccl"
+  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 + $i`
+    test_dir=$dir/test_${decode_modes}
+    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 \
+      --val_ref_file data/test/text \
+      --val_hyp_file $test_dir/text \
+      --log_dir $log_dir
+  } &
+  done
+  wait
+fi

examples/aishell/s0/run_validate.sh

+#!/bin/bash
+
+# Copyright 2019 Mobvoi Inc. All Rights Reserved.
+. ./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"
+# The NCCL_SOCKET_IFNAME variable specifies which IP interface to use for nccl
+# communication. More details can be found in
+# https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html
+# export NCCL_SOCKET_IFNAME=ens4f1
+export NCCL_DEBUG=INFO
+stage=5 # 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=/data/conformer/train/
+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
+# Optional train_config
+# 1. conf/train_transformer.yaml: Standard transformer
+# 2. conf/train_conformer.yaml: Standard conformer
+# 3. conf/train_unified_conformer.yaml: Unified dynamic chunk causal conformer
+# 4. conf/train_unified_transformer.yaml: Unified dynamic chunk transformer
+# 5. conf/train_u2++_conformer.yaml: U2++ conformer
+# 6. conf/train_u2++_transformer.yaml: U2++ transformer
+train_config=conf/train_conformer.yaml
+cmvn=true
+dir=exp/conformer
+checkpoint=
+
+# use average_checkpoint will get better result
+average_checkpoint=false
+decode_checkpoint=$dir/final.pt
+average_num=30
+decode_modes="attention_rescoring"
+
+. tools/parse_options.sh || exit 1;
+
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
+  # 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=
+  ctc_weight=0.5
+  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 $ctc_weight \
+      --reverse_weight $reverse_weight \
+      --result_file $test_dir/text \
+      ${decoding_chunk_size:+--decoding_chunk_size $decoding_chunk_size}
+    
+    # The code of computing script below has been executed in the wenet/bin/train.py 
+    #python tools/compute-wer.py --char=1 --v=1 \
+    #  data/test/text $test_dir/text > $test_dir/wer
+  } &
+  done
+  wait
+fi

examples/aishell/s0/tools

+../../../tools/
\ No newline at end of file

examples/aishell/s0/train.sh

+#!/usr/bin/env bash
+CUR_DIR="$( cd "$(dirname "$0")" ; pwd )"
+cd $CUR_DIR
+bash run_train.sh --log_dir pid.txt 2>&1 | tee conformer_train_dcu_`date +%Y%m%d%H%M%S`.log
+

examples/aishell/s0/validate.sh

+#!/usr/bin/env bash
+CUR_DIR="$( cd "$(dirname "$0")" ; pwd )"
+cd $CUR_DIR
+bash run_validate.sh

examples/aishell/s0/wenet

+../../../wenet/
\ No newline at end of file

examples/aishell2/rnnt/README.md

+# Performance Record
+
+## U2++ Conformer Result
+
+* Feature info: using fbank feature, dither, cmvn, oneline speed perturb
+* Training info: lr 0.001, dynamic batch with max_frames_in_batch 15000, 4 gpu, acc_grad 1, 130 epochs
+* Training weight info: transducer_weight 0.75,  ctc_weight 0.1, reverse_weight 0.30, average_num 30
+* Predictor type: lstm
+
+| decoding mode/chunk size  | full  | 16    |
+|---------------------------|-------|-------|
+| rnnt greedy search        | 6.44  | 7.09  |
+

examples/aishell2/rnnt/conf/conformer_rnnt.yaml

+# network architecture
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.1
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
+    cnn_module_kernel: 15
+    use_cnn_module: true
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
+
+
+joint_conf:
+    join_dim: 512
+    prejoin_linear: True
+    postjoin_linear: false
+    joint_mode: 'add'
+    activation: 'tanh'
+
+predictor: rnn
+predictor_conf:
+    embed_size: 256
+    output_size: 256
+    embed_dropout: 0.1
+    hidden_size: 256
+    num_layers: 2
+    bias: true
+    rnn_type: 'lstm'
+    dropout: 0.1
+
+decoder: bitransformer
+decoder_conf:
+  attention_heads: 4
+  dropout_rate: 0.1
+  linear_units: 2048
+  num_blocks: 3
+  positional_dropout_rate: 0.1
+  r_num_blocks: 3
+  self_attention_dropout_rate: 0.1
+  src_attention_dropout_rate: 0.1
+
+# hybrid transducer+ctc+attention
+model_conf:
+    transducer_weight: 0.75
+    ctc_weight: 0.1
+    attention_weight: 0.15
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+    reverse_weight: 0.3
+
+dataset_conf:
+    filter_conf:
+        max_length: 40960
+        min_length: 10
+        token_max_length: 200
+        token_min_length: 1
+    resample_conf:
+        resample_rate: 16000
+    speed_perturb: true
+    fbank_conf:
+        num_mel_bins: 80
+        frame_shift: 10
+        frame_length: 25
+        dither: 0.1
+    spec_aug: true
+    spec_aug_conf:
+        num_t_mask: 2
+        num_f_mask: 2
+        max_t: 50
+        max_f: 10
+    shuffle: true
+    shuffle_conf:
+        shuffle_size: 1500
+    sort: true
+    sort_conf:
+        sort_size: 500  # sort_size should be less than shuffle_size
+    batch_conf:
+        batch_type: 'static' # static or dynamic
+        batch_size: 8
+
+grad_clip: 4
+accum_grad: 1
+max_epoch: 140
+log_interval: 100
+
+optim: adam
+optim_conf:
+    lr: 0.001
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000
+

examples/aishell2/rnnt/conf/conformer_u2pp_rnnt.yaml

+# network architecture
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.1
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
+    cnn_module_kernel: 8
+    use_cnn_module: True
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
+    causal: true
+    use_dynamic_chunk: true
+    cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
+    use_dynamic_left_chunk: false
+
+
+joint_conf:
+    join_dim: 512
+    prejoin_linear: True
+    postjoin_linear: false
+    joint_mode: 'add'
+    activation: 'tanh'
+
+predictor: rnn
+predictor_conf:
+    embed_size: 256
+    output_size: 256
+    embed_dropout: 0.1
+    hidden_size: 256
+    num_layers: 2
+    bias: true
+    rnn_type: 'lstm'
+    dropout: 0.1
+
+decoder: bitransformer
+decoder_conf:
+  attention_heads: 4
+  dropout_rate: 0.1
+  linear_units: 2048
+  num_blocks: 3
+  positional_dropout_rate: 0.1
+  r_num_blocks: 3
+  self_attention_dropout_rate: 0.1
+  src_attention_dropout_rate: 0.1
+
+# hybrid transducer+ctc+attention
+model_conf:
+    transducer_weight: 0.75
+    ctc_weight: 0.1
+    attention_weight: 0.15
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+    reverse_weight: 0.3
+
+dataset_conf:
+    filter_conf:
+        max_length: 40960
+        min_length: 10
+        token_max_length: 200
+        token_min_length: 1
+    resample_conf:
+        resample_rate: 16000
+    speed_perturb: true
+    fbank_conf:
+        num_mel_bins: 80
+        frame_shift: 10
+        frame_length: 25
+        dither: 0.1
+    spec_aug: true
+    spec_aug_conf:
+        num_t_mask: 2
+        num_f_mask: 2
+        max_t: 50
+        max_f: 10
+    shuffle: true
+    shuffle_conf:
+        shuffle_size: 1500
+    sort: true
+    sort_conf:
+        sort_size: 500  # sort_size should be less than shuffle_size
+    batch_conf:
+        batch_type: 'dynamic' # static or dynamic
+        max_frames_in_batch: 15000
+
+grad_clip: 4
+accum_grad: 1
+max_epoch: 130
+log_interval: 100
+
+optim: adam
+optim_conf:
+    lr: 0.001
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000

examples/aishell2/rnnt/local

+../s0/local
\ No newline at end of file