Dev bat (#701)

* boundary aware transducer

* resolve conflict

* delete type check

* add bat egs and results

---------

Co-authored-by: aky15 <ankeyu.aky@11.17.44.249>

egs/aishell/bat/README.md

@@ -0,0 +1,16 @@
+# Boundary Aware Transducer (BAT) Result
+
+## Training Config
+- 8 gpu(Tesla V100)
+- Feature info: using 80 dims fbank, global cmvn, speed perturb(0.9, 1.0, 1.1), specaugment
+- Train config: conf/train_conformer_bat.yaml
+- LM config: LM was not used
+- Model size: 90M
+
+## Results (CER)
+- Decode config: conf/decode_bat_conformer.yaml
+
+|   testset   |  CER(%) |
+|:-----------:|:-------:|
+|     dev     |  4.56   |
+|    test     |  4.97   |

egs/aishell/bat/conf/decode_bat_conformer.yaml

@@ -0,0 +1 @@
+beam_size: 10

egs/aishell/bat/conf/train_conformer_bat.yaml

@@ -0,0 +1,108 @@
+encoder: chunk_conformer
+encoder_conf:
+      activation_type: swish
+      positional_dropout_rate: 0.5
+      time_reduction_factor: 2
+      embed_vgg_like: false
+      subsampling_factor: 4
+      linear_units: 2048
+      output_size: 512
+      attention_heads: 8
+      dropout_rate: 0.5
+      positional_dropout_rate: 0.5
+      attention_dropout_rate: 0.5
+      cnn_module_kernel: 15
+      num_blocks: 12    
+
+# decoder related
+rnnt_decoder: rnnt
+rnnt_decoder_conf:
+    embed_size: 512
+    hidden_size: 512
+    embed_dropout_rate: 0.5
+    dropout_rate: 0.5
+    use_embed_mask: true
+
+predictor: bat_predictor
+predictor_conf:
+  idim: 512
+  threshold: 1.0
+  l_order: 1
+  r_order: 1
+  return_accum: true
+
+joint_network_conf:
+    joint_space_size: 512
+
+# frontend related
+frontend: wav_frontend
+frontend_conf:
+    fs: 16000
+    window: hamming
+    n_mels: 80
+    frame_length: 25
+    frame_shift: 10
+    lfr_m: 1
+    lfr_n: 1
+
+
+# Auxiliary CTC
+model: bat
+model_conf:
+    auxiliary_ctc_weight: 0.0
+    cif_weight: 1.0
+    r_d: 3
+    r_u: 5
+
+# minibatch related
+use_amp: true
+
+# optimization related
+accum_grad: 1
+grad_clip: 5
+max_epoch: 100
+val_scheduler_criterion:
+    - valid
+    - loss
+best_model_criterion:
+-   - valid
+    - cer_transducer
+    - min
+keep_nbest_models: 10
+
+optim: adam
+optim_conf:
+   lr: 0.001
+scheduler: warmuplr
+scheduler_conf:
+   warmup_steps: 25000
+
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 40
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_range:
+    - 0
+    - 50
+    num_time_mask: 5
+
+dataset_conf:
+    data_names: speech,text
+    data_types: sound,text
+    shuffle: True
+    shuffle_conf:
+        shuffle_size: 2048
+        sort_size: 500
+    batch_conf:
+        batch_type: token
+        batch_size: 25000
+    num_workers: 8
+
+log_interval: 50

egs/aishell/bat/local/aishell_data_prep.sh

@@ -0,0 +1,66 @@
+#!/bin/bash
+
+# Copyright 2017 Xingyu Na
+# Apache 2.0
+
+#. ./path.sh || exit 1;
+
+if [ $# != 3 ]; then
+  echo "Usage: $0 <audio-path> <text-path> <output-path>"
+  echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
+  exit 1;
+fi
+
+aishell_audio_dir=$1
+aishell_text=$2/aishell_transcript_v0.8.txt
+output_dir=$3
+
+train_dir=$output_dir/data/local/train
+dev_dir=$output_dir/data/local/dev
+test_dir=$output_dir/data/local/test
+tmp_dir=$output_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
+  utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
+  awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
+  utils/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 $output_dir/data/train $output_dir/data/dev $output_dir/data/test
+
+for f in wav.scp text; do
+  cp $train_dir/$f $output_dir/data/train/$f || exit 1;
+  cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
+  cp $test_dir/$f $output_dir/data/test/$f || exit 1;
+done
+
+echo "$0: AISHELL data preparation succeeded"
+exit 0;

egs/aishell/bat/path.sh

@@ -0,0 +1,5 @@
+export FUNASR_DIR=$PWD/../../..
+
+# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
+export PYTHONIOENCODING=UTF-8
+export PATH=$FUNASR_DIR/funasr/bin:$PATH

egs/aishell/bat/run.sh

@@ -0,0 +1,210 @@
+#!/usr/bin/env bash
+
+. ./path.sh || exit 1;
+
+# machines configuration
+CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+gpu_num=8
+count=1
+gpu_inference=true  # Whether to perform gpu decoding, set false for cpu decoding
+# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
+njob=5
+train_cmd=utils/run.pl
+infer_cmd=utils/run.pl
+
+# general configuration
+feats_dir="../DATA" #feature output dictionary
+exp_dir="."
+lang=zh
+token_type=char
+type=sound
+scp=wav.scp
+speed_perturb="0.9 1.0 1.1"
+stage=0
+stop_stage=5
+
+# feature configuration
+feats_dim=80
+nj=64
+
+# data
+raw_data=../raw_data
+data_url=www.openslr.org/resources/33
+
+# exp tag
+tag="exp1"
+
+. utils/parse_options.sh || exit 1;
+
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+set -u
+set -o pipefail
+
+train_set=train
+valid_set=dev
+test_sets="dev test"
+
+asr_config=conf/train_conformer_bat.yaml
+model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
+
+inference_config=conf/decode_bat_conformer.yaml
+inference_asr_model=valid.cer_transducer.ave_10best.pb
+
+# you can set gpu num for decoding here
+gpuid_list=$CUDA_VISIBLE_DEVICES  # set gpus for decoding, the same as training stage by default
+ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
+
+if ${gpu_inference}; then
+    inference_nj=$[${ngpu}*${njob}]
+    _ngpu=1
+else
+    inference_nj=$njob
+    _ngpu=0
+fi
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+    echo "stage -1: Data Download"
+    local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
+    local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    echo "stage 0: Data preparation"
+    # Data preparation
+    local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
+    for x in train dev test; do
+        cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
+        paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
+            > ${feats_dir}/data/${x}/text
+        utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
+        mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
+    done
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    echo "stage 1: Feature and CMVN Generation"
+    utils/compute_cmvn.sh --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} ${feats_dir}/data/${train_set}
+fi
+
+token_list=${feats_dir}/data/${lang}_token_list/char/tokens.txt
+echo "dictionary: ${token_list}"
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    echo "stage 2: Dictionary Preparation"
+    mkdir -p ${feats_dir}/data/${lang}_token_list/char/
+
+    echo "make a dictionary"
+    echo "<blank>" > ${token_list}
+    echo "<s>" >> ${token_list}
+    echo "</s>" >> ${token_list}
+    utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
+        | sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
+    echo "<unk>" >> ${token_list}
+fi
+
+# LM Training Stage
+world_size=$gpu_num  # run on one machine
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+    echo "stage 3: LM Training"
+fi
+
+# ASR Training Stage
+world_size=$gpu_num  # run on one machine
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    echo "stage 4: ASR Training"
+    mkdir -p ${exp_dir}/exp/${model_dir}
+    mkdir -p ${exp_dir}/exp/${model_dir}/log
+    INIT_FILE=./ddp_init
+    if [ -f $INIT_FILE ];then
+        rm -f $INIT_FILE
+    fi 
+    init_method=file://$(readlink -f $INIT_FILE)
+    echo "$0: init method is $init_method"
+    for ((i = 0; i < $gpu_num; ++i)); do
+        {
+            rank=$i
+            local_rank=$i
+            gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
+            train.py \
+                --task_name asr \
+                --gpu_id $gpu_id \
+                --use_preprocessor true \
+                --token_type char \
+                --token_list $token_list \
+                --data_dir ${feats_dir}/data \
+                --train_set ${train_set} \
+                --valid_set ${valid_set} \
+                --data_file_names "wav.scp,text" \
+                --cmvn_file ${feats_dir}/data/${train_set}/cmvn/cmvn.mvn \
+                --speed_perturb ${speed_perturb} \
+                --resume true \
+                --output_dir ${exp_dir}/exp/${model_dir} \
+                --config $asr_config \
+                --ngpu $gpu_num \
+                --num_worker_count $count \
+                --dist_init_method $init_method \
+                --dist_world_size $world_size \
+                --dist_rank $rank \
+                --local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
+        } &
+        done
+        wait
+fi
+
+# Testing Stage
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
+    echo "stage 5: Inference"
+    for dset in ${test_sets}; do
+        asr_exp=${exp_dir}/exp/${model_dir}
+        inference_tag="$(basename "${inference_config}" .yaml)"
+        _dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
+        _logdir="${_dir}/logdir"
+        if [ -d ${_dir} ]; then
+            echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
+            exit 0
+        fi
+        mkdir -p "${_logdir}"
+        _data="${feats_dir}/data/${dset}"
+        key_file=${_data}/${scp}
+        num_scp_file="$(<${key_file} wc -l)"
+        _nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
+        split_scps=
+        for n in $(seq "${_nj}"); do
+            split_scps+=" ${_logdir}/keys.${n}.scp"
+        done
+        # shellcheck disable=SC2086
+        utils/split_scp.pl "${key_file}" ${split_scps}
+        _opts=
+        if [ -n "${inference_config}" ]; then
+            _opts+="--config ${inference_config} "
+        fi
+        ${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
+            python -m funasr.bin.asr_inference_launch \
+                --batch_size 1 \
+                --ngpu "${_ngpu}" \
+                --njob ${njob} \
+                --gpuid_list ${gpuid_list} \
+                --data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
+                --cmvn_file ${feats_dir}/data/${train_set}/cmvn/cmvn.mvn \
+                --key_file "${_logdir}"/keys.JOB.scp \
+                --asr_train_config "${asr_exp}"/config.yaml \
+                --asr_model_file "${asr_exp}"/"${inference_asr_model}" \
+                --output_dir "${_logdir}"/output.JOB \
+                --mode bat \
+                ${_opts}
+
+        for f in token token_int score text; do
+            if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
+                for i in $(seq "${_nj}"); do
+                    cat "${_logdir}/output.${i}/1best_recog/${f}"
+                done | sort -k1 >"${_dir}/${f}"
+            fi
+        done
+        python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
+        python utils/proce_text.py ${_data}/text ${_data}/text.proc
+        python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
+        tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
+        cat ${_dir}/text.cer.txt
+    done
+fi

egs/aishell/bat/utils

@@ -0,0 +1 @@
+../transformer/utils