From 8f21baf63482020397be16db846a533ad8a8731a Mon Sep 17 00:00:00 2001
From: 嘉渊 <wangjiaming.wjm@alibaba-inc.com>
Date: 星期一, 15 五月 2023 19:30:32 +0800
Subject: [PATCH] update repo
---
funasr/build_utils/build_pretrain_model.py | 2
egs/aishell2/data2vec_pretrain/conf/train_pretrain_transformer.yaml | 68 +++++++++-------
egs/aishell2/data2vec_pretrain/run.sh | 131 ++++++++++++++++----------------
3 files changed, 105 insertions(+), 96 deletions(-)
diff --git a/egs/aishell2/data2vec_pretrain/conf/train_pretrain_transformer.yaml b/egs/aishell2/data2vec_pretrain/conf/train_pretrain_transformer.yaml
index 4052774..b6e8808 100644
--- a/egs/aishell2/data2vec_pretrain/conf/train_pretrain_transformer.yaml
+++ b/egs/aishell2/data2vec_pretrain/conf/train_pretrain_transformer.yaml
@@ -2,47 +2,52 @@
# encoder related
encoder: data2vec_encoder
encoder_conf:
- extractor_mode: layer_norm
- encoder_layerdrop: 0.05
- dropout_input: 0.0
- dropout_features: 0.0
- feature_grad_mult: 1.0
- encoder_embed_dim: 768
+ extractor_mode: layer_norm
+ encoder_layerdrop: 0.05
+ dropout_input: 0.0
+ dropout_features: 0.0
+ feature_grad_mult: 1.0
+ encoder_embed_dim: 768
- mask_prob: 0.65
- mask_length: 10
+ mask_prob: 0.65
+ mask_length: 10
- loss_beta: 0
- loss_scale: null
+ loss_beta: 0
+ loss_scale: null
- instance_norm_target_layer: true
- average_top_k_layers: 8
+ instance_norm_target_layer: true
+ average_top_k_layers: 8
- pos_conv_depth: 5
- conv_pos: 95
+ pos_conv_depth: 5
+ conv_pos: 95
- ema_decay: 0.999
- ema_end_decay: 0.9999
- ema_anneal_end_step: 30000
- ema_transformer_only: true
- ema_layers_only: true
+ ema_decay: 0.999
+ ema_end_decay: 0.9999
+ ema_anneal_end_step: 30000
+ ema_transformer_only: true
+ ema_layers_only: true
- require_same_masks: true
- mask_dropout: 0
+ require_same_masks: true
+ mask_dropout: 0
-log_interval: 50
-normalize: None
+# 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
-# minibatch related
-batch_type: length
-batch_bins: 64000
-num_workers: 16
+model: data2vec
# optimization related
accum_grad: 1
grad_clip: 5
patience: none
-max_epoch: 600
+max_epoch: 1800
val_scheduler_criterion:
- valid
- acc
@@ -68,7 +73,7 @@
dataset_conf:
batch_mode: clipping
data_names: speech,none
- data_types: kaldi_ark,none
+ data_types: sound,none
shuffle: true
shuffle_conf:
shuffle_size: 12800
@@ -76,4 +81,7 @@
batch_conf:
batch_type: token
batch_size: 64000
- num_workers: 8
\ No newline at end of file
+ num_workers: 8
+
+log_interval: 50
+normalize: None
\ No newline at end of file
diff --git a/egs/aishell2/data2vec_pretrain/run.sh b/egs/aishell2/data2vec_pretrain/run.sh
index eceb183..2753f00 100755
--- a/egs/aishell2/data2vec_pretrain/run.sh
+++ b/egs/aishell2/data2vec_pretrain/run.sh
@@ -7,28 +7,25 @@
gpu_num=8
count=1
-train_cmd=tools/run.pl
+train_cmd=utils/run.pl
# general configuration
feats_dir="../DATA" #feature output dictionary
exp_dir="."
lang=zh
-dumpdir=dump/fbank
-feats_type=fbank
token_type=char
+speed_perturb="0.9 1.0 1.1"
dataset_type=large
-stage=0
-stop_stage=4
+stage=3
+stop_stage=3
# 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"
@@ -45,68 +42,31 @@
valid_set=dev_ios
asr_config=conf/train_pretrain_transformer.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}"
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
echo "stage 0: Data preparation"
# For training set
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
+ 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
+ done
# Normalize text to capital letters
- for x in train dev_android dev_ios dev_mic test_android test_ios test_mic; do
+ 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
- tools/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
+ 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} \
- ${feats_dir}/data/train ${exp_dir}/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 \
- ${feats_dir}/data/dev_${x} ${exp_dir}/exp/make_fbank/dev_${x} ${fbankdir}/dev_${x}
- tools/fix_data_feat.sh ${fbankdir}/dev_${x}
- steps/compute_fbank.sh --cmd "$train_cmd" --nj $nj \
- ${feats_dir}/data/test_${x} ${exp_dir}/exp/make_fbank/test_${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 ${exp_dir}/exp/make_fbank/train
-
- # apply cmvn
- steps/apply_cmvn.sh --cmd "$train_cmd" --nj $nj \
- ${fbankdir}/${train_set} ${fbankdir}/train/cmvn.json ${exp_dir}/exp/make_fbank/${train_set} ${feat_train_dir}
- steps/apply_cmvn.sh --cmd "$train_cmd" --nj $nj \
- ${fbankdir}/${valid_set} ${fbankdir}/train/cmvn.json ${exp_dir}/exp/make_fbank/${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 ${exp_dir}/exp/make_fbank/test_${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
@@ -114,22 +74,59 @@
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}
- tools/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/${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 ${feats_dir}/asr_stats_fbank_zh_char/${train_set}
mkdir -p ${feats_dir}/asr_stats_fbank_zh_char/${valid_set}
- cp ${feat_train_dir}/speech_shape ${feat_train_dir}/text_shape ${feat_train_dir}/text_shape.char ${feats_dir}/asr_stats_fbank_zh_char/${train_set}
- cp ${feat_dev_dir}/speech_shape ${feat_dev_dir}/text_shape ${feat_dev_dir}/text_shape.char ${feats_dir}/asr_stats_fbank_zh_char/${valid_set}
+ fi
+
+# Training Stage
+world_size=$gpu_num # run on one machine
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+ echo "stage 3: 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 \
+ --token_type char \
+ --token_list $token_list \
+ --data_dir ${feats_dir}/data \
+ --train_set ${train_set} \
+ --valid_set ${valid_set} \
+ --cmvn_file ${feats_dir}/data/${train_set}/cmvn/cmvn.mvn \
+ --speed_perturb ${speed_perturb} \
+ --dataset_type $dataset_type \
+ --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.$i 2>&1
+ } &
+ done
+ wait
fi
# Training Stage
@@ -149,12 +146,16 @@
rank=$i
local_rank=$i
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
- data2vec_train.py \
+ train.py \
+ --task_name pretrain \
--gpu_id $gpu_id \
--use_preprocessor true \
+ --data_dir ${feats_dir}/data \
+ --train_set ${train_set} \
+ --valid_set ${valid_set} \
+ --cmvn_file ${feats_dir}/data/${train_set}/cmvn/cmvn.mvn \
+ --speed_perturb ${speed_perturb} \
--dataset_type $dataset_type \
- --train_data_file $feats_dir/$dumpdir/${train_set}/data.list \
- --valid_data_file $feats_dir/$dumpdir/${valid_set}/data.list \
--resume true \
--output_dir ${exp_dir}/exp/${model_dir} \
--config $asr_config \
diff --git a/funasr/build_utils/build_pretrain_model.py b/funasr/build_utils/build_pretrain_model.py
index e514215..629937f 100644
--- a/funasr/build_utils/build_pretrain_model.py
+++ b/funasr/build_utils/build_pretrain_model.py
@@ -89,7 +89,7 @@
**args.encoder_conf,
)
- if args.model_name == "data2vec":
+ if args.model == "data2vec":
model_class = model_choices.get_class("data2vec")
model = model_class(
frontend=frontend,
--
Gitblit v1.9.1