| egs/librispeech/branchformer/conf/decode_asr_transformer_beam10_ctc0.3.yaml | ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史 | |
| egs/librispeech/branchformer/conf/train_asr_branchformer.yaml | ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史 | |
| egs/librispeech/branchformer/local/data_prep.sh | ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史 | |
| egs/librispeech/branchformer/local/download_and_untar.sh | ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史 | |
| egs/librispeech/branchformer/local/spm_encode.py | ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史 | |
| egs/librispeech/branchformer/local/spm_train.py | ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史 | |
| egs/librispeech/branchformer/path.sh | ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史 | |
| egs/librispeech/branchformer/run.sh | ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史 | |
| egs/librispeech/branchformer/utils | ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史 |
egs/librispeech/branchformer/conf/decode_asr_transformer_beam10_ctc0.3.yaml
New file @@ -0,0 +1,6 @@ beam_size: 10 penalty: 0.0 maxlenratio: 0.0 minlenratio: 0.0 ctc_weight: 0.3 lm_weight: 0.0 egs/librispeech/branchformer/conf/train_asr_branchformer.yaml
New file @@ -0,0 +1,104 @@ # network architecture # encoder related encoder: branchformer encoder_conf: output_size: 512 use_attn: true attention_heads: 8 attention_layer_type: rel_selfattn pos_enc_layer_type: rel_pos rel_pos_type: latest use_cgmlp: true cgmlp_linear_units: 3072 cgmlp_conv_kernel: 31 use_linear_after_conv: false gate_activation: identity merge_method: concat cgmlp_weight: 0.5 # used only if merge_method is "fixed_ave" attn_branch_drop_rate: 0.0 # used only if merge_method is "learned_ave" num_blocks: 18 dropout_rate: 0.1 positional_dropout_rate: 0.1 attention_dropout_rate: 0.1 input_layer: conv2d stochastic_depth_rate: 0.0 # decoder related decoder: transformer decoder_conf: attention_heads: 8 linear_units: 2048 num_blocks: 6 dropout_rate: 0.1 positional_dropout_rate: 0.1 self_attention_dropout_rate: 0.1 src_attention_dropout_rate: 0.1 # 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 # hybrid CTC/attention model_conf: ctc_weight: 0.3 lsm_weight: 0.1 # label smoothing option length_normalized_loss: false # optimization related accum_grad: 2 grad_clip: 5 max_epoch: 210 val_scheduler_criterion: - valid - acc best_model_criterion: - - valid - acc - max keep_nbest_models: 10 optim: adam optim_conf: lr: 0.0025 weight_decay: 0.000001 scheduler: warmuplr scheduler_conf: warmup_steps: 40000 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 - 27 num_freq_mask: 2 apply_time_mask: true time_mask_width_ratio_range: - 0. - 0.05 num_time_mask: 10 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: 30000 num_workers: 8 log_interval: 50 normalize: None egs/librispeech/branchformer/local/data_prep.sh
New file @@ -0,0 +1,58 @@ #!/usr/bin/env bash # Copyright 2014 Vassil Panayotov # 2014 Johns Hopkins University (author: Daniel Povey) # Apache 2.0 if [ "$#" -ne 2 ]; then echo "Usage: $0 <src-dir> <dst-dir>" echo "e.g.: $0 /export/a15/vpanayotov/data/LibriSpeech/dev-clean data/dev-clean" exit 1 fi src=$1 dst=$2 # all utterances are FLAC compressed if ! which flac >&/dev/null; then echo "Please install 'flac' on ALL worker nodes!" exit 1 fi spk_file=$src/../SPEAKERS.TXT mkdir -p $dst || exit 1 [ ! -d $src ] && echo "$0: no such directory $src" && exit 1 [ ! -f $spk_file ] && echo "$0: expected file $spk_file to exist" && exit 1 wav_scp=$dst/wav.scp; [[ -f "$wav_scp" ]] && rm $wav_scp trans=$dst/text; [[ -f "$trans" ]] && rm $trans for reader_dir in $(find -L $src -mindepth 1 -maxdepth 1 -type d | sort); do reader=$(basename $reader_dir) if ! [ $reader -eq $reader ]; then # not integer. echo "$0: unexpected subdirectory name $reader" exit 1 fi for chapter_dir in $(find -L $reader_dir/ -mindepth 1 -maxdepth 1 -type d | sort); do chapter=$(basename $chapter_dir) if ! [ "$chapter" -eq "$chapter" ]; then echo "$0: unexpected chapter-subdirectory name $chapter" exit 1 fi find -L $chapter_dir/ -iname "*.flac" | sort | xargs -I% basename % .flac | \ awk -v "dir=$chapter_dir" '{printf "%s %s/%s.flac \n", $0, dir, $0}' >>$wav_scp|| exit 1 chapter_trans=$chapter_dir/${reader}-${chapter}.trans.txt [ ! -f $chapter_trans ] && echo "$0: expected file $chapter_trans to exist" && exit 1 cat $chapter_trans >>$trans done done echo "$0: successfully prepared data in $dst" exit 0 egs/librispeech/branchformer/local/download_and_untar.sh
New file @@ -0,0 +1,97 @@ #!/usr/bin/env bash # Copyright 2014 Johns Hopkins University (author: Daniel Povey) # 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/a15/vpanayotov/data www.openslr.org/resources/11 dev-clean" echo "With --remove-archive it will remove the archive after successfully un-tarring it." echo "<corpus-part> can be one of: dev-clean, test-clean, dev-other, test-other," echo " train-clean-100, train-clean-360, train-other-500." exit 1 fi data=$1 url=$2 part=$3 if [ ! -d "$data" ]; then echo "$0: no such directory $data" exit 1 fi part_ok=false list="dev-clean test-clean dev-other test-other train-clean-100 train-clean-360 train-other-500" 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 $data/LibriSpeech/$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. This is some older versions. sizes_old="371012589 347390293 379743611 361838298 6420417880 23082659865 30626749128" # sizes_new is the archive file sizes of the final release. Some of these sizes are of # things we probably won't download. sizes_new="337926286 314305928 695964615 297279345 87960560420 33373768 346663984 328757843 6387309499 23049477885 30593501606" if [ -f $data/$part.tar.gz ]; then size=$(/bin/ls -l $data/$part.tar.gz | awk '{print $5}') size_ok=false for s in $sizes_old $sizes_new; do if [ $s == $size ]; then size_ok=true; fi; done if ! $size_ok; then echo "$0: removing existing file $data/$part.tar.gz because its size in bytes $size" echo "does not equal the size of one of the archives." rm $data/$part.tar.gz else echo "$data/$part.tar.gz exists and appears to be complete." fi fi if [ ! -f $data/$part.tar.gz ]; then if ! which wget >/dev/null; then echo "$0: wget is not installed." exit 1 fi full_url=$url/$part.tar.gz echo "$0: downloading data from $full_url. This may take some time, please be patient." if ! wget -P $data --no-check-certificate $full_url; then echo "$0: error executing wget $full_url" exit 1 fi fi if ! tar -C $data -xvzf $data/$part.tar.gz; then echo "$0: error un-tarring archive $data/$part.tar.gz" exit 1 fi touch $data/LibriSpeech/$part/.complete echo "$0: Successfully downloaded and un-tarred $data/$part.tar.gz" if $remove_archive; then echo "$0: removing $data/$part.tar.gz file since --remove-archive option was supplied." rm $data/$part.tar.gz fi egs/librispeech/branchformer/local/spm_encode.py
New file @@ -0,0 +1,98 @@ #!/usr/bin/env python # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in # https://github.com/pytorch/fairseq/blob/master/LICENSE import argparse import contextlib import sys import sentencepiece as spm def main(): parser = argparse.ArgumentParser() parser.add_argument("--model", required=True, help="sentencepiece model to use for encoding") parser.add_argument("--inputs", nargs="+", default=['-'], help="input files to filter/encode") parser.add_argument("--outputs", nargs="+", default=['-'], help="path to save encoded outputs") parser.add_argument("--output_format", choices=["piece", "id"], default="piece") parser.add_argument("--min-len", type=int, metavar="N", help="filter sentence pairs with fewer than N tokens") parser.add_argument("--max-len", type=int, metavar="N", help="filter sentence pairs with more than N tokens") args = parser.parse_args() assert len(args.inputs) == len(args.outputs), \ "number of input and output paths should match" sp = spm.SentencePieceProcessor() sp.Load(args.model) if args.output_format == "piece": def encode(l): return sp.EncodeAsPieces(l) elif args.output_format == "id": def encode(l): return list(map(str, sp.EncodeAsIds(l))) else: raise NotImplementedError if args.min_len is not None or args.max_len is not None: def valid(line): return ( (args.min_len is None or len(line) >= args.min_len) and (args.max_len is None or len(line) <= args.max_len) ) else: def valid(lines): return True with contextlib.ExitStack() as stack: inputs = [ stack.enter_context(open(input, "r", encoding="utf-8")) if input != "-" else sys.stdin for input in args.inputs ] outputs = [ stack.enter_context(open(output, "w", encoding="utf-8")) if output != "-" else sys.stdout for output in args.outputs ] stats = { "num_empty": 0, "num_filtered": 0, } def encode_line(line): line = line.strip() if len(line) > 0: line = encode(line) if valid(line): return line else: stats["num_filtered"] += 1 else: stats["num_empty"] += 1 return None for i, lines in enumerate(zip(*inputs), start=1): enc_lines = list(map(encode_line, lines)) if not any(enc_line is None for enc_line in enc_lines): for enc_line, output_h in zip(enc_lines, outputs): print(" ".join(enc_line), file=output_h) if i % 10000 == 0: print("processed {} lines".format(i), file=sys.stderr) print("skipped {} empty lines".format(stats["num_empty"]), file=sys.stderr) print("filtered {} lines".format(stats["num_filtered"]), file=sys.stderr) if __name__ == "__main__": main() egs/librispeech/branchformer/local/spm_train.py
New file @@ -0,0 +1,12 @@ #!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # https://github.com/pytorch/fairseq/blob/master/LICENSE import sys import sentencepiece as spm if __name__ == "__main__": spm.SentencePieceTrainer.Train(" ".join(sys.argv[1:])) egs/librispeech/branchformer/path.sh
New file @@ -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:$PATH egs/librispeech/branchformer/run.sh
New file @@ -0,0 +1,223 @@ #!/usr/bin/env bash . ./path.sh || exit 1; # machines configuration CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" gpu_num=8 count=1 gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding # for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob njob=5 train_cmd=utils/run.pl infer_cmd=utils/run.pl # general configuration feats_dir="../DATA" #feature output dictionary exp_dir="." lang=en token_type=bpe type=sound scp=wav.scp speed_perturb="0.9 1.0 1.1" stage=0 stop_stage=5 # feature configuration feats_dim=80 nj=64 # data raw_data= data_url=www.openslr.org/resources/12 # bpe model nbpe=5000 bpemode=unigram # exp tag tag="exp1" . utils/parse_options.sh || exit 1; # Set bash to 'debug' mode, it will exit on : # -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands', set -e set -u set -o pipefail train_set=train_960 valid_set=dev test_sets="test_clean test_other dev_clean dev_other" asr_config=conf/train_asr_branchformer.yaml model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}" inference_config=conf/decode_asr_transformer_beam10_ctc0.3.yaml inference_asr_model=valid.acc.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" for part in dev-clean test-clean dev-other test-other train-clean-100 train-clean-360 train-other-500; do local/download_and_untar.sh ${raw_data} ${data_url} ${part} done fi if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then echo "stage 0: Data preparation" # Data preparation for x in dev-clean dev-other test-clean test-other train-clean-100 train-clean-360 train-other-500; do local/data_prep.sh ${raw_data}/LibriSpeech/${x} ${feats_dir}/data/${x//-/_} done mkdir $feats_dir/data/$valid_set dev_sets="dev_clean dev_other" for file in wav.scp text; do ( for f in $dev_sets; do cat $feats_dir/data/$f/$file; done ) | sort -k1 > $feats_dir/data/$valid_set/$file || exit 1; done mkdir $feats_dir/data/$train_set train_sets="train_clean_100 train_clean_360 train_other_500" for file in wav.scp text; do ( for f in $train_sets; do cat $feats_dir/data/$f/$file; done ) | sort -k1 > $feats_dir/data/$train_set/$file || exit 1; done fi if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then echo "stage 1: Feature and CMVN Generation" utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 1.0 fi token_list=${feats_dir}/data/lang_char/${train_set}_${bpemode}${nbpe}_units.txt bpemodel=${feats_dir}/data/lang_char/${train_set}_${bpemode}${nbpe} echo "dictionary: ${token_list}" if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then ### Task dependent. You have to check non-linguistic symbols used in the corpus. echo "stage 2: Dictionary and Json Data Preparation" mkdir -p ${feats_dir}/data/lang_char/ echo "<blank>" > ${token_list} echo "<s>" >> ${token_list} echo "</s>" >> ${token_list} cut -f 2- -d" " ${feats_dir}/data/${train_set}/text > ${feats_dir}/data/lang_char/input.txt local/spm_train.py --input=${feats_dir}/data/lang_char/input.txt --vocab_size=${nbpe} --model_type=${bpemode} --model_prefix=${bpemodel} --input_sentence_size=100000000 local/spm_encode.py --model=${bpemodel}.model --output_format=piece < ${feats_dir}/data/lang_char/input.txt | tr ' ' '\n' | sort | uniq | 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=${exp_dir}/exp/${model_dir}/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 \ --split_with_space false \ --bpemodel ${bpemodel}.model \ --token_type $token_type \ --token_list $token_list \ --dataset_type large \ --data_dir ${feats_dir}/data \ --train_set ${train_set} \ --valid_set ${valid_set} \ --cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \ --speed_perturb ${speed_perturb} \ --resume true \ --output_dir ${exp_dir}/exp/${model_dir} \ --config $asr_config \ --ngpu $gpu_num \ --num_worker_count $count \ --multiprocessing_distributed true \ --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/am.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 asr \ ${_opts} for f in token token_int score text; do if [ -f "${_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/compute_wer.py ${_data}/text ${_dir}/text ${_dir}/text.cer tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt cat ${_dir}/text.cer.txt done fi egs/librispeech/branchformer/utils
New file @@ -0,0 +1 @@ ../../aishell/transformer/utils