update repo

嘉渊

2023-05-15 2db4a207d11bb4b5269def967f29f9aed1fe3ba7

update repo

 egs/aishell/paraformerbert/run.sh

@@ -111,12 +111,12 @@
world_size=$gpu_num  # run on one machine
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
    echo "stage 3: Training"
#    if ! "${skip_extract_embed}"; then
#        echo "extract embeddings..."
#        local/extract_embeds.sh \
#            --bert_model_name ${bert_model_name} \
#            --raw_dataset_path ${feats_dir}
#    fi
    if ! "${skip_extract_embed}"; then
        echo "extract embeddings..."
        local/extract_embeds.sh \
            --bert_model_name ${bert_model_name} \
            --raw_dataset_path ${feats_dir}
    fi
    mkdir -p ${exp_dir}/exp/${model_dir}
    mkdir -p ${exp_dir}/exp/${model_dir}/log
    INIT_FILE=${exp_dir}/exp/${model_dir}/ddp_init

 egs/aishell2/paraformerbert/local/extract_embeds.sh

@@ -3,20 +3,17 @@
stage=1
stop_stage=3

bert_model_root="../../huggingface_models"
bert_model_name="bert-base-chinese"
#bert_model_name="chinese-roberta-wwm-ext"
#bert_model_name="mengzi-bert-base"
raw_dataset_path="../DATA"
model_path=${bert_model_root}/${bert_model_name}
model_path=${bert_model_name}

. utils/parse_options.sh || exit 1;

nj=100
nj=32

for data_set in train dev_ios test_ios;do
    scp=$raw_dataset_path/dump/fbank/${data_set}/text
    local_scp_dir_raw=$raw_dataset_path/embeds/$bert_model_name/${data_set}
for data_set in train dev test;do
    scp=$raw_dataset_path/data/${data_set}/text
    local_scp_dir_raw=${raw_dataset_path}/data/embeds/${data_set}
    local_scp_dir=$local_scp_dir_raw/split$nj
    local_records_dir=$local_scp_dir_raw/ark

@@ -31,7 +28,7 @@
    utils/split_scp.pl $scp ${split_scps}


    for num in {0..24};do
    for num in {0..7};do
        tmp=`expr $num \* 4`

        if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
@@ -41,20 +38,9 @@
                {
                    beg=0
                    gpu=`expr $beg + $idx`
                    echo $local_scp_dir_raw/log/log.${JOB}
                    python tools/extract_embeds.py $local_scp_dir/text.$JOB.txt ${local_records_dir}/embeds.${JOB}.ark ${local_records_dir}/embeds.${JOB}.scp ${local_records_dir}/embeds.${JOB}.shape ${gpu} ${model_path} &> $local_scp_dir_raw/log/log.${JOB}
                    echo ${local_scp_dir}/log.${JOB}
                    python utils/extract_embeds.py $local_scp_dir/data.$JOB.text ${local_records_dir}/embeds.${JOB}.ark ${local_records_dir}/embeds.${JOB}.scp ${local_records_dir}/embeds.${JOB}.shape ${gpu} ${model_path} &> ${local_scp_dir}/log.${JOB}
            } &
            done
            wait
        fi

        if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
            for idx in {1..4}; do
                JOB=`expr $tmp + $idx`
                echo "upload jobid=$JOB"
                {
                    hadoop  fs -put -f ${local_records_dir}/embeds.${JOB}.ark ${odps_des_feature_dir}/embeds.${JOB}.ark
                } &
            done
            wait
        fi

 egs/aishell2/paraformerbert/local/prepare_data.sh

@@ -17,7 +17,6 @@
fi

corpus=$1
#dict_dir=$2
tmp=$2
dir=$3

@@ -35,14 +34,14 @@
# validate utt-key list, IC0803W0380 is a bad utterance
awk '{print $1}' $corpus/wav.scp | grep -v 'IC0803W0380' > $tmp/wav_utt.list
awk '{print $1}' $corpus/trans.txt > $tmp/trans_utt.list
tools/filter_scp.pl -f 1 $tmp/wav_utt.list $tmp/trans_utt.list > $tmp/utt.list
utils/filter_scp.pl -f 1 $tmp/wav_utt.list $tmp/trans_utt.list > $tmp/utt.list

# wav.scp
awk -F'\t' -v path_prefix=$corpus '{printf("%s\t%s/%s\n",$1,path_prefix,$2)}' $corpus/wav.scp > $tmp/tmp_wav.scp
tools/filter_scp.pl -f 1 $tmp/utt.list $tmp/tmp_wav.scp | sort -k 1 | uniq > $tmp/wav.scp
utils/filter_scp.pl -f 1 $tmp/utt.list $tmp/tmp_wav.scp | sort -k 1 | uniq > $tmp/wav.scp

# text
tools/filter_scp.pl -f 1 $tmp/utt.list $corpus/trans.txt | sort -k 1 | uniq > $tmp/text
utils/filter_scp.pl -f 1 $tmp/utt.list $corpus/trans.txt | sort -k 1 | uniq > $tmp/text

# copy prepared resources from tmp_dir to target dir
mkdir -p $dir

 egs/aishell2/paraformerbert/run.sh

@@ -8,36 +8,32 @@
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=tools/run.pl
njob=1
train_cmd=utils/run.pl
infer_cmd=utils/run.pl

# general configuration
feats_dir="../DATA" #feature output dictionary, for large data
feats_dir="../DATA" #feature output dictionary
exp_dir="."
lang=zh
dumpdir=dump/fbank
feats_type=fbank
token_type=char
type=sound
scp=wav.scp
speed_perturb="0.9 1.0 1.1"
dataset_type=large
scp=feats.scp
type=kaldi_ark
stage=0
stop_stage=5
stage=3
stop_stage=4

skip_extract_embed=false
bert_model_root="../../huggingface_models"
bert_model_name="bert-base-chinese"

# feature configuration
feats_dim=80
sample_frequency=16000
nj=100
speed_perturb="0.9,1.0,1.1"
nj=64

# data
tr_dir=
dev_tst_dir=
tr_dir=/nfs/wangjiaming.wjm/asr_data/aishell2/AISHELL-2/iOS/data
dev_tst_dir=/nfs/wangjiaming.wjm/asr_data/aishell2/AISHELL-DEV-TEST-SET

# exp tag
tag="exp1"
@@ -55,7 +51,7 @@
test_sets="dev_ios test_ios"

asr_config=conf/train_asr_paraformerbert_conformer_20e_6d_1280_320.yaml
model_dir="baseline_$(basename "${asr_config}" .yaml)_${feats_type}_${lang}_${token_type}_${tag}"
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"

inference_config=conf/decode_asr_transformer_noctc_1best.yaml
inference_asr_model=valid.acc.ave_10best.pb
@@ -75,86 +71,44 @@
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
    echo "stage 0: Data preparation"
    # For training set
    local/prepare_data.sh ${tr_dir} data/local/train data/train || exit 1;
    local/prepare_data.sh ${tr_dir} ${feats_dir}/data/local/train ${feats_dir}/data/train || exit 1;
    # # For dev and test set
    for x in Android iOS Mic; do
        local/prepare_data.sh ${dev_tst_dir}/${x}/dev data/local/dev_${x,,} data/dev_${x,,} || exit 1;
        local/prepare_data.sh ${dev_tst_dir}/${x}/test data/local/test_${x,,} data/test_${x,,} || exit 1;
    done 
    for x in iOS; do
        local/prepare_data.sh ${dev_tst_dir}/${x}/dev ${feats_dir}/data/local/dev_${x,,} ${feats_dir}/data/dev_${x,,} || exit 1;
        local/prepare_data.sh ${dev_tst_dir}/${x}/test ${feats_dir}/data/local/test_${x,,} ${feats_dir}/data/test_${x,,} || exit 1;
    done
    # Normalize text to capital letters
    for x in train dev_android dev_ios dev_mic test_android test_ios test_mic; do
        mv data/${x}/text data/${x}/text.org
        paste <(cut -f 1 data/${x}/text.org) <(cut -f 2 data/${x}/text.org | tr '[:lower:]' '[:upper:]') \
            > data/${x}/text
        tools/text2token.py -n 1 -s 1 data/${x}/text > data/${x}/text.org
        mv data/${x}/text.org data/${x}/text
    for x in train dev_ios test_ios; do
        mv ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
        paste -d " " <(cut -f 1 ${feats_dir}/data/${x}/text.org) <(cut -f 2- ${feats_dir}/data/${x}/text.org \
             | tr 'A-Z' 'a-z' | 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

feat_train_dir=${feats_dir}/${dumpdir}/${train_set}; mkdir -p ${feat_train_dir}
feat_dev_dir=${feats_dir}/${dumpdir}/${valid_set}; mkdir -p ${feat_dev_dir}
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
    echo "stage 1: Feature Generation"
    # compute fbank features
    fbankdir=${feats_dir}/fbank
    steps/compute_fbank.sh --cmd "$train_cmd" --nj $nj --speed_perturb ${speed_perturb} \
        data/train exp/make_fbank/train ${fbankdir}/train
    tools/fix_data_feat.sh ${fbankdir}/train
    for x in android ios mic; do
        steps/compute_fbank.sh --cmd "$train_cmd" --nj $nj \
            data/dev_${x} exp/make_fbank/dev_${x} ${fbankdir}/dev_${x}
        tools/fix_data_feat.sh ${fbankdir}/dev_${x}
        steps/compute_fbank.sh --cmd "$train_cmd" --nj $nj \
            data/test_${x} exp/make_fbank/test_${x} ${fbankdir}/test_${x}
        tools/fix_data_feat.sh ${fbankdir}/test_${x}
    done
    
    # compute global cmvn
    steps/compute_cmvn.sh --cmd "$train_cmd" --nj $nj \
        ${fbankdir}/train exp/make_fbank/train

    # apply cmvn 
    steps/apply_cmvn.sh --cmd "$train_cmd" --nj $nj \
        ${fbankdir}/${train_set} ${fbankdir}/train/cmvn.json exp/make_fbank/${train_set} ${feat_train_dir}
    steps/apply_cmvn.sh --cmd "$train_cmd" --nj $nj \
        ${fbankdir}/${valid_set} ${fbankdir}/train/cmvn.json exp/make_fbank/${valid_set} ${feat_dev_dir}
    for x in android ios mic; do
        steps/apply_cmvn.sh --cmd "$train_cmd" --nj $nj \
            ${fbankdir}/test_${x} ${fbankdir}/train/cmvn.json exp/make_fbank/test_${x} ${feats_dir}/${dumpdir}/test_${x}
    done
    
    cp ${fbankdir}/${train_set}/text ${fbankdir}/${train_set}/speech_shape ${fbankdir}/${train_set}/text_shape ${feat_train_dir}
    tools/fix_data_feat.sh ${feat_train_dir}
    cp ${fbankdir}/${valid_set}/text ${fbankdir}/${valid_set}/speech_shape ${fbankdir}/${valid_set}/text_shape ${feat_dev_dir}
    tools/fix_data_feat.sh ${feat_dev_dir}
    for x in android ios mic; do
        cp ${fbankdir}/test_${x}/text ${fbankdir}/test_${x}/speech_shape ${fbankdir}/test_${x}/text_shape ${feats_dir}/${dumpdir}/test_${x}
        tools/fix_data_feat.sh ${feats_dir}/${dumpdir}/test_${x}
    done
    echo "stage 1: Feature and CMVN Generation"
    utils/compute_cmvn.sh --cmd "$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 ] && [ ${stop_stage} -ge 2 ]; then
    echo "stage 2: Dictionary Preparation"
    mkdir -p data/${lang}_token_list/char/
   
    mkdir -p ${feats_dir}/data/${lang}_token_list/char/

    echo "make a dictionary"
    echo "<blank>" > ${token_list}
    echo "<s>" >> ${token_list}
    echo "</s>" >> ${token_list}
    tools/text2token.py -s 1 -n 1 --space "" data/${train_set}/text | cut -f 2- -d" " | tr " " "\n" \
    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}' >> ${token_list}
    num_token=$(cat ${token_list} | wc -l)
    echo "<unk>" >> ${token_list}
    vocab_size=$(cat ${token_list} | wc -l)
    awk -v v=,${vocab_size} '{print $0v}' ${feat_train_dir}/text_shape > ${feat_train_dir}/text_shape.char
    awk -v v=,${vocab_size} '{print $0v}' ${feat_dev_dir}/text_shape > ${feat_dev_dir}/text_shape.char
    mkdir -p asr_stats_fbank_zh_char/${train_set} 
    mkdir -p asr_stats_fbank_zh_char/${valid_set}
    cp ${feat_train_dir}/speech_shape ${feat_train_dir}/text_shape ${feat_train_dir}/text_shape.char asr_stats_fbank_zh_char/${train_set} 
    cp ${feat_dev_dir}/speech_shape ${feat_dev_dir}/text_shape ${feat_dev_dir}/text_shape.char asr_stats_fbank_zh_char/${valid_set}
fi
    mkdir -p ${feats_dir}/asr_stats_fbank_zh_char/${train_set}
    mkdir -p ${feats_dir}/asr_stats_fbank_zh_char/${valid_set}
 fi

# Training Stage
world_size=$gpu_num  # run on one machine

 funasr/datasets/large_datasets/utils/tokenize.py

@@ -37,7 +37,7 @@
    vad = -2

    if bpe_tokenizer is not None:
        text = bpe_tokenizer.text2tokens("".join(text))
        text = bpe_tokenizer.text2tokens(text)

    if seg_dict is not None:
        assert isinstance(seg_dict, dict)