msb-public
/
PaddleSpeech
mirror of https://github.com/PaddlePaddle/PaddleSpeech
1
0
Code Issues Projects Releases Wiki Activity

change all recipes

Browse Source
pull/1225/head
huangyuxin 3 years ago
parent 5d6494decc
commit c907a8deda
122 changed files with 2427 additions and 2359 deletions
Show Stats Download Patch File Download Diff File

28
examples/aishell/asr0/conf/deepspeech2.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.00 min_output_input_ratio: 0.00
max_output_input_ratio: .inf max_output_input_ratio: .inf
collator: ###########################################
# Dataloader #
###########################################
batch_size: 64 # one gpu batch_size: 64 # one gpu
mean_std_filepath: data/mean_std.json mean_std_filepath: data/mean_std.json
unit_type: char unit_type: char
@ -34,7 +38,9 @@ collator:
shuffle_method: batch_shuffle shuffle_method: batch_shuffle
num_workers: 2 num_workers: 2
model: ############################################
# Network Architecture #
############################################
num_conv_layers: 2 num_conv_layers: 2
num_rnn_layers: 3 num_rnn_layers: 3
rnn_layer_size: 1024 rnn_layer_size: 1024
@ -43,7 +49,9 @@ model:
blank_id: 0 blank_id: 0
ctc_grad_norm_type: instance ctc_grad_norm_type: instance
training: ###########################################
# Training #
###########################################
n_epoch: 80 n_epoch: 80
accum_grad: 1 accum_grad: 1
lr: 2e-3 lr: 2e-3
@ -54,15 +62,3 @@ training:
checkpoint: checkpoint:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 128
error_rate_type: cer
decoding_method: ctc_beam_search
lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
alpha: 1.9
beta: 5.0
beam_size: 300
cutoff_prob: 0.99
cutoff_top_n: 40
num_proc_bsearch: 10

28
examples/aishell/asr0/conf/deepspeech2_online.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.00 min_output_input_ratio: 0.00
max_output_input_ratio: .inf max_output_input_ratio: .inf
collator: ###########################################
# Dataloader #
###########################################
batch_size: 64 # one gpu batch_size: 64 # one gpu
mean_std_filepath: data/mean_std.json mean_std_filepath: data/mean_std.json
unit_type: char unit_type: char
@ -34,7 +38,9 @@ collator:
shuffle_method: batch_shuffle shuffle_method: batch_shuffle
num_workers: 0 num_workers: 0
model: ############################################
# Network Architecture #
############################################
num_conv_layers: 2 num_conv_layers: 2
num_rnn_layers: 5 num_rnn_layers: 5
rnn_layer_size: 1024 rnn_layer_size: 1024
@ -45,7 +51,9 @@ model:
blank_id: 0 blank_id: 0
training: ###########################################
# Training #
###########################################
n_epoch: 65 n_epoch: 65
accum_grad: 1 accum_grad: 1
lr: 5e-4 lr: 5e-4
@ -57,14 +65,4 @@ training:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 32
error_rate_type: cer
decoding_method: ctc_beam_search
lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
alpha: 2.2 #1.9
beta: 4.3
beam_size: 300
cutoff_prob: 0.99
cutoff_top_n: 40
num_proc_bsearch: 10

10
examples/aishell/asr0/conf/tuning/chunk_decode.yaml
Unescape View File

@ -0,0 +1,10 @@
chunk_batch_size: 32
error_rate_type: cer
decoding_method: ctc_beam_search
lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
alpha: 2.2 #1.9
beta: 4.3
beam_size: 300
cutoff_prob: 0.99
cutoff_top_n: 40
num_proc_bsearch: 10

10
examples/aishell/asr0/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,10 @@
decode_batch_size: 128
error_rate_type: cer
decoding_method: ctc_beam_search
lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
alpha: 1.9
beta: 5.0
beam_size: 300
cutoff_prob: 0.99
cutoff_top_n: 40
num_proc_bsearch: 10

10
examples/aishell/asr0/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 3 ];then if [ $# != 4 ];then
echo "usage: ${0} config_path ckpt_path_prefix model_type" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
exit -1 exit -1
fi fi
@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
model_type=$3 ckpt_prefix=$3
model_type=$4
# download language model # download language model
bash local/download_lm_ch.sh bash local/download_lm_ch.sh
@ -21,6 +22,7 @@ fi
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.rsl \ --result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--model_type ${model_type} --model_type ${model_type}

10
examples/aishell/asr0/local/test_export.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 3 ];then if [ $# != 4 ];then
echo "usage: ${0} config_path ckpt_path_prefix model_type" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
exit -1 exit -1
fi fi
@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
jit_model_export_path=$2 decode_config_path=$2
model_type=$3 jit_model_export_path=$3
model_type=$4
# download language model # download language model
bash local/download_lm_ch.sh > /dev/null 2>&1 bash local/download_lm_ch.sh > /dev/null 2>&1
@ -21,6 +22,7 @@ fi
python3 -u ${BIN_DIR}/test_export.py \ python3 -u ${BIN_DIR}/test_export.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${jit_model_export_path}.rsl \ --result_file ${jit_model_export_path}.rsl \
--export_path ${jit_model_export_path} \ --export_path ${jit_model_export_path} \
--model_type ${model_type} --model_type ${model_type}

47
examples/aishell/asr0/local/test_hub_ori
Unescape View File

@ -0,0 +1,47 @@
#!/bin/bash
if [ $# != 4 ];then
echo "usage: ${0} config_path ckpt_path_prefix model_type audio_file"
exit -1
fi
ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..."
config_path=$1
ckpt_prefix=$2
model_type=$3
audio_file=$4
mkdir -p data
wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/
if [ $? -ne 0 ]; then
exit 1
fi
if [ ! -f ${audio_file} ]; then
echo "Plase input the right audio_file path"
exit 1
fi
# download language model
bash local/download_lm_ch.sh
if [ $? -ne 0 ]; then
exit 1
fi
python3 -u ${BIN_DIR}/test_hub.py \
--nproc ${ngpu} \
--config ${config_path} \
--result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix} \
--model_type ${model_type} \
--audio_file ${audio_file}
if [ $? -ne 0 ]; then
echo "Failed in evaluation!"
exit 1
fi
exit 0

12
examples/aishell/asr0/local/test_wav.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 4 ];then if [ $# != 5 ];then
echo "usage: ${0} config_path ckpt_path_prefix model_type audio_file" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type audio_file"
exit -1 exit -1
fi fi
@ -9,9 +9,10 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
model_type=$3 ckpt_prefix=$3
audio_file=$4 model_type=$4
audio_file=$5
mkdir -p data mkdir -p data
wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/ wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/
@ -33,6 +34,7 @@ fi
python3 -u ${BIN_DIR}/test_wav.py \ python3 -u ${BIN_DIR}/test_wav.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.rsl \ --result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--model_type ${model_type} \ --model_type ${model_type} \

7
examples/aishell/asr0/run.sh
Unescape View File

@ -6,6 +6,7 @@ gpus=0,1,2,3
stage=0 stage=0
stop_stage=100 stop_stage=100
conf_path=conf/deepspeech2.yaml #conf/deepspeech2.yaml or conf/deepspeeech2_online.yaml conf_path=conf/deepspeech2.yaml #conf/deepspeech2.yaml or conf/deepspeeech2_online.yaml
decode_conf_path=conf/tuning/decode.yaml
avg_num=1 avg_num=1
model_type=offline # offline or online model_type=offline # offline or online
audio_file=data/demo_01_03.wav audio_file=data/demo_01_03.wav
@ -34,7 +35,7 @@ fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type}|| exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type}|| exit -1
fi fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
@ -44,11 +45,11 @@ fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
# test export ckpt avg_n # test export ckpt avg_n
CUDA_VISIBLE_DEVICES=0 ./local/test_export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt}.jit ${model_type}|| exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test_export.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt}.jit ${model_type}|| exit -1
fi fi
# Optionally, you can add LM and test it with runtime. # Optionally, you can add LM and test it with runtime.
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
# test a single .wav file # test a single .wav file
CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} ${audio_file} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} ${audio_file} || exit -1
fi fi

7
examples/aishell/asr1/conf/chunk_conformer.yaml
Unescape View File

@ -54,8 +54,9 @@ test_manifest: data/manifest.test
########################################### ###########################################
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
spm_model_prefix: ''
unit_type: 'char' unit_type: 'char'
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
feat_dim: 80 feat_dim: 80
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
@ -74,7 +75,7 @@ subsampling_factor: 1
num_encs: 1 num_encs: 1
########################################### ###########################################
# training # # Training #
########################################### ###########################################
n_epoch: 240 n_epoch: 240
accum_grad: 2 accum_grad: 2
@ -82,7 +83,7 @@ global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.002 lr: 0.002
weight_decay: 1e-6 weight_decay: 1.0e-6
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000

5
examples/aishell/asr1/conf/conformer.yaml
Unescape View File

@ -49,8 +49,9 @@ test_manifest: data/manifest.test
# Dataloader # # Dataloader #
########################################### ###########################################
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
spm_model_prefix: ''
unit_type: 'char' unit_type: 'char'
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
feat_dim: 80 feat_dim: 80
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
@ -69,7 +70,7 @@ subsampling_factor: 1
num_encs: 1 num_encs: 1
########################################### ###########################################
# training # # Training #
########################################### ###########################################
n_epoch: 240 n_epoch: 240
accum_grad: 2 accum_grad: 2

7
examples/aishell/asr1/conf/transformer.yaml
Unescape View File

@ -46,6 +46,7 @@ test_manifest: data/manifest.test
########################################### ###########################################
unit_type: 'char' unit_type: 'char'
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
spm_model_prefix: ''
feat_dim: 80 feat_dim: 80
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
@ -59,13 +60,13 @@ batch_bins: 0
batch_frames_in: 0 batch_frames_in: 0
batch_frames_out: 0 batch_frames_out: 0
batch_frames_inout: 0 batch_frames_inout: 0
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
num_workers: 0 num_workers: 0
subsampling_factor: 1 subsampling_factor: 1
num_encs: 1 num_encs: 1
########################################### ###########################################
# training # # Training #
########################################### ###########################################
n_epoch: 240 n_epoch: 240
accum_grad: 2 accum_grad: 2
@ -73,7 +74,7 @@ global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.002 lr: 0.002
weight_decay: 1e-6 weight_decay: 1.0e-6
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000

2
examples/aishell/asr1/local/align.sh
Unescape View File

@ -21,7 +21,7 @@ mkdir -p ${output_dir}
python3 -u ${BIN_DIR}/alignment.py \ python3 -u ${BIN_DIR}/alignment.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_config ${decode_config_path} \ --decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.align \ --result_file ${output_dir}/${type}.align \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decode.decode_batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}

6
examples/aishell/asr1/local/test.sh
Unescape View File

@ -30,14 +30,14 @@ for type in attention ctc_greedy_search; do
# stream decoding only support batchsize=1 # stream decoding only support batchsize=1
batch_size=1 batch_size=1
else else
batch_size=1 batch_size=64
fi fi
output_dir=${ckpt_prefix} output_dir=${ckpt_prefix}
mkdir -p ${output_dir} mkdir -p ${output_dir}
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_config ${decode_config_path} \ --decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.rsl \ --result_file ${output_dir}/${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decode.decoding_method ${type} \ --opts decode.decoding_method ${type} \
@ -57,7 +57,7 @@ for type in ctc_prefix_beam_search attention_rescoring; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_config ${decode_config_path} \ --decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.rsl \ --result_file ${output_dir}/${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decode.decoding_method ${type} \ --opts decode.decoding_method ${type} \

2
examples/aishell/asr1/local/test_wav.sh
Unescape View File

@ -43,7 +43,7 @@ for type in attention_rescoring; do
python3 -u ${BIN_DIR}/test_wav.py \ python3 -u ${BIN_DIR}/test_wav.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_config ${decode_config_path} \ --decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.rsl \ --result_file ${output_dir}/${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decode.decoding_method ${type} \ --opts decode.decoding_method ${type} \

50
examples/callcenter/asr1/conf/chunk_conformer.yaml
Unescape View File

@ -1,21 +1,19 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
min_input_len: 0.5
max_input_len: 20.0 # second
min_output_len: 0.0
max_output_len: 400.0
min_output_input_ratio: 0.05
max_output_input_ratio: 10.0
collator: ###########################################
# Dataloader #
###########################################
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
unit_type: 'char' unit_type: 'char'
spm_model_prefix: '' spm_model_prefix: ''
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
batch_size: 32 batch_size: 32
raw_wav: True # use raw_wav or kaldi feature raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank spectrum_type: fbank #linear, mfcc, fbank
@ -36,8 +34,9 @@ collator:
num_workers: 2 num_workers: 2
# network architecture ############################################
model: # Network Architecture #
############################################
cmvn_file: cmvn_file:
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -79,15 +78,16 @@ model:
lsm_weight: 0.1 # label smoothing option lsm_weight: 0.1 # label smoothing option
length_normalized_loss: false length_normalized_loss: false
###########################################
training: # Training #
###########################################
n_epoch: 240 n_epoch: 240
accum_grad: 4 accum_grad: 4
global_grad_clip: 5.0 global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.001 lr: 0.001
weight_decay: 1e-6 weight_decay: 1.0e-6
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000
@ -96,25 +96,3 @@ training:
checkpoint: checkpoint:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 128
error_rate_type: cer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 10
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: true # simulate streaming inference. Defaults to False.

79
examples/callcenter/asr1/conf/conformer.yaml
Unescape View File

@ -1,42 +1,39 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
min_input_len: 0.5
max_input_len: 20.0 # second
min_output_len: 0.0
max_output_len: 400.0
min_output_input_ratio: 0.0
max_output_input_ratio: .inf
collator: ###########################################
# Dataloader #
###########################################
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
unit_type: 'char' unit_type: 'char'
spm_model_prefix: '' spm_model_prefix: ''
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
batch_size: 32
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80 feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 8000
max_freq: None
n_fft: None
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
use_dB_normalization: True sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
target_dB: -20 batch_size: 64
random_seed: 0 maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
keep_transcription_text: False maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
sortagrad: True minibatches: 0 # for debug
shuffle_method: batch_shuffle batch_count: auto
num_workers: 2 batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
num_workers: 0
subsampling_factor: 1
num_encs: 1
# network architecture ############################################
model: # Network Architecture #
############################################
cmvn_file: cmvn_file:
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -75,14 +72,16 @@ model:
length_normalized_loss: false length_normalized_loss: false
training: ###########################################
# Training #
###########################################
n_epoch: 100 # 50 will be lowest n_epoch: 100 # 50 will be lowest
accum_grad: 4 accum_grad: 4
global_grad_clip: 5.0 global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.002 lr: 0.002
weight_decay: 1e-6 weight_decay: 1.0e-6
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000
@ -91,27 +90,3 @@ training:
checkpoint: checkpoint:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 128
error_rate_type: cer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 10
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

2
examples/callcenter/asr1/conf/preprocess.yaml
Unescape View File

@ -1,7 +1,7 @@
process: process:
# extract kaldi fbank from PCM # extract kaldi fbank from PCM
- type: fbank_kaldi - type: fbank_kaldi
fs: 16000 fs: 8000
n_mels: 80 n_mels: 80
n_shift: 160 n_shift: 160
win_length: 400 win_length: 400

11
examples/callcenter/asr1/conf/tuning/chunk_decode.yaml
Unescape View File

@ -0,0 +1,11 @@
decode_batch_size: 128
error_rate_type: cer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
beam_size: 10
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: true # simulate streaming inference. Defaults to False.

13
examples/callcenter/asr1/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,13 @@
decode_batch_size: 128
error_rate_type: cer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
beam_size: 10
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

10
examples/callcenter/asr1/local/align.sh
Unescape View File

@ -1,7 +1,7 @@
#! /usr/bin/env bash #! /usr/bin/env bash
if [ $# != 2 ];then if [ $# != 3 ];then
echo "usage: ${0} config_path ckpt_path_prefix" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
exit -1 exit -1
fi fi
@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
ckpt_prefix=$3
ckpt_name=$(basename ${ckpt_prefxi}) ckpt_name=$(basename ${ckpt_prefxi})
@ -25,9 +26,10 @@ mkdir -p ${output_dir}
python3 -u ${BIN_DIR}/alignment.py \ python3 -u ${BIN_DIR}/alignment.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.align \ --result_file ${output_dir}/${type}.align \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in ctc alignment!" echo "Failed in ctc alignment!"

18
examples/callcenter/asr1/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#! /usr/bin/env bash #! /usr/bin/env bash
if [ $# != 2 ];then if [ $# != 3 ];then
echo "usage: ${0} config_path ckpt_path_prefix" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
exit -1 exit -1
fi fi
@ -9,7 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
ckpt_prefix=$3
ckpt_name=$(basename ${ckpt_prefxi}) ckpt_name=$(basename ${ckpt_prefxi})
@ -30,10 +32,11 @@ for type in attention ctc_greedy_search; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.rsl \ --result_file ${output_dir}/${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"
@ -49,10 +52,11 @@ for type in ctc_prefix_beam_search attention_rescoring; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.rsl \ --result_file ${output_dir}/${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"

5
examples/callcenter/asr1/run.sh
Unescape View File

@ -6,6 +6,7 @@ gpus=0,1,2,3
stage=0 stage=0
stop_stage=100 stop_stage=100
conf_path=conf/conformer.yaml conf_path=conf/conformer.yaml
decode_conf_path=conf/tuning/decode.yaml
avg_num=20 avg_num=20
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1; source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
@ -31,12 +32,12 @@ fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
# ctc alignment of test data # ctc alignment of test data
CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then

29
examples/librispeech/asr0/conf/deepspeech2.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev-clean dev_manifest: data/manifest.dev-clean
test_manifest: data/manifest.test-clean test_manifest: data/manifest.test-clean
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.00 min_output_input_ratio: 0.00
max_output_input_ratio: .inf max_output_input_ratio: .inf
collator: ###########################################
# Dataloader #
###########################################
batch_size: 20 batch_size: 20
mean_std_filepath: data/mean_std.json mean_std_filepath: data/mean_std.json
unit_type: char unit_type: char
@ -19,6 +23,7 @@ collator:
random_seed: 0 random_seed: 0
spm_model_prefix: spm_model_prefix:
spectrum_type: linear spectrum_type: linear
feat_dim:
target_sample_rate: 16000 target_sample_rate: 16000
max_freq: None max_freq: None
n_fft: None n_fft: None
@ -34,7 +39,9 @@ collator:
shuffle_method: batch_shuffle shuffle_method: batch_shuffle
num_workers: 2 num_workers: 2
model: ############################################
# Network Architecture #
############################################
num_conv_layers: 2 num_conv_layers: 2
num_rnn_layers: 3 num_rnn_layers: 3
rnn_layer_size: 2048 rnn_layer_size: 2048
@ -43,7 +50,9 @@ model:
blank_id: 0 blank_id: 0
training: ###########################################
# Training #
###########################################
n_epoch: 50 n_epoch: 50
accum_grad: 1 accum_grad: 1
lr: 1e-3 lr: 1e-3
@ -54,15 +63,3 @@ training:
checkpoint: checkpoint:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 128
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 1.9
beta: 0.3
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

29
examples/librispeech/asr0/conf/deepspeech2_online.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev-clean dev_manifest: data/manifest.dev-clean
test_manifest: data/manifest.test-clean test_manifest: data/manifest.test-clean
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.00 min_output_input_ratio: 0.00
max_output_input_ratio: .inf max_output_input_ratio: .inf
collator: ###########################################
# Dataloader #
###########################################
batch_size: 15 batch_size: 15
mean_std_filepath: data/mean_std.json mean_std_filepath: data/mean_std.json
unit_type: char unit_type: char
@ -19,6 +23,7 @@ collator:
random_seed: 0 random_seed: 0
spm_model_prefix: spm_model_prefix:
spectrum_type: linear spectrum_type: linear
feat_dim:
target_sample_rate: 16000 target_sample_rate: 16000
max_freq: None max_freq: None
n_fft: None n_fft: None
@ -34,7 +39,9 @@ collator:
shuffle_method: batch_shuffle shuffle_method: batch_shuffle
num_workers: 0 num_workers: 0
model: ############################################
# Network Architecture #
############################################
num_conv_layers: 2 num_conv_layers: 2
num_rnn_layers: 3 num_rnn_layers: 3
rnn_layer_size: 2048 rnn_layer_size: 2048
@ -45,7 +52,9 @@ model:
blank_id: 0 blank_id: 0
training: ###########################################
# Training #
###########################################
n_epoch: 50 n_epoch: 50
accum_grad: 4 accum_grad: 4
lr: 1e-3 lr: 1e-3
@ -56,15 +65,3 @@ training:
checkpoint: checkpoint:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 128
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 1.9
beta: 0.3
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/librispeech/asr0/conf/tuning/chunk_decode.yaml
Unescape View File

@ -0,0 +1,10 @@
decode_batch_size: 128
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 1.9
beta: 0.3
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/librispeech/asr0/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,10 @@
decode_batch_size: 128
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 1.9
beta: 0.3
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/librispeech/asr0/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 3 ];then if [ $# != 4 ];then
echo "usage: ${0} config_path ckpt_path_prefix model_type" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
exit -1 exit -1
fi fi
@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
model_type=$3 ckpt_prefix=$3
model_type=$4
# download language model # download language model
bash local/download_lm_en.sh bash local/download_lm_en.sh
@ -21,6 +22,7 @@ fi
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.rsl \ --result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--model_type ${model_type} --model_type ${model_type}

12
examples/librispeech/asr0/local/test_wav.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 4 ];then if [ $# != 5 ];then
echo "usage: ${0} config_path ckpt_path_prefix model_type audio_file" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type audio_file"
exit -1 exit -1
fi fi
@ -9,9 +9,10 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
model_type=$3 ckpt_prefix=$3
audio_file=$4 model_type=$4
audio_file=$5
mkdir -p data mkdir -p data
wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/en/demo_002_en.wav -P data/ wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/en/demo_002_en.wav -P data/
@ -33,6 +34,7 @@ fi
python3 -u ${BIN_DIR}/test_wav.py \ python3 -u ${BIN_DIR}/test_wav.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.rsl \ --result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--model_type ${model_type} \ --model_type ${model_type} \

5
examples/librispeech/asr0/run.sh
Unescape View File

@ -6,6 +6,7 @@ gpus=0,1,2,3,4,5,6,7
stage=0 stage=0
stop_stage=100 stop_stage=100
conf_path=conf/deepspeech2.yaml conf_path=conf/deepspeech2.yaml
decode_conf_path=conf/tuning/decode.yaml
avg_num=30 avg_num=30
model_type=offline model_type=offline
audio_file=data/demo_002_en.wav audio_file=data/demo_002_en.wav
@ -33,7 +34,7 @@ fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} || exit -1
fi fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
@ -43,5 +44,5 @@ fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
# test a single .wav file # test a single .wav file
CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} ${audio_file} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} ${audio_file} || exit -1
fi fi

6
examples/librispeech/asr1/conf/chunk_conformer.yaml
Unescape View File

@ -57,7 +57,7 @@ vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm' unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_5000' spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
mean_std_filepath: "" mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
feat_dim: 80 feat_dim: 80
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
@ -71,7 +71,6 @@ batch_bins: 0
batch_frames_in: 0 batch_frames_in: 0
batch_frames_out: 0 batch_frames_out: 0
batch_frames_inout: 0 batch_frames_inout: 0
augmentation_config: conf/preprocess.yaml
num_workers: 0 num_workers: 0
subsampling_factor: 1 subsampling_factor: 1
num_encs: 1 num_encs: 1
@ -85,10 +84,11 @@ global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.001 lr: 0.001
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000
lr_decay: 1.0
log_interval: 100 log_interval: 100
checkpoint: checkpoint:
kbest_n: 50 kbest_n: 50

5
examples/librispeech/asr1/conf/chunk_transformer.yaml
Unescape View File

@ -50,7 +50,7 @@ vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm' unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_5000' spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
mean_std_filepath: "" mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
feat_dim: 80 feat_dim: 80
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
@ -64,7 +64,6 @@ batch_bins: 0
batch_frames_in: 0 batch_frames_in: 0
batch_frames_out: 0 batch_frames_out: 0
batch_frames_inout: 0 batch_frames_inout: 0
augmentation_config: conf/preprocess.yaml
num_workers: 0 num_workers: 0
subsampling_factor: 1 subsampling_factor: 1
num_encs: 1 num_encs: 1
@ -79,7 +78,7 @@ global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.001 lr: 0.001
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000

5
examples/librispeech/asr1/conf/conformer.yaml
Unescape View File

@ -55,7 +55,7 @@ vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm' unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_5000' spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
mean_std_filepath: "" mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
feat_dim: 80 feat_dim: 80
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
@ -69,7 +69,6 @@ batch_bins: 0
batch_frames_in: 0 batch_frames_in: 0
batch_frames_out: 0 batch_frames_out: 0
batch_frames_inout: 0 batch_frames_inout: 0
augmentation_config: conf/preprocess.yaml
num_workers: 0 num_workers: 0
subsampling_factor: 1 subsampling_factor: 1
num_encs: 1 num_encs: 1
@ -84,7 +83,7 @@ global_grad_clip: 3.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.004 lr: 0.004
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000

5
examples/librispeech/asr1/conf/transformer.yaml
Unescape View File

@ -49,7 +49,7 @@ vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm' unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_5000' spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
mean_std_filepath: "" mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
feat_dim: 80 feat_dim: 80
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
@ -63,7 +63,6 @@ batch_bins: 0
batch_frames_in: 0 batch_frames_in: 0
batch_frames_out: 0 batch_frames_out: 0
batch_frames_inout: 0 batch_frames_inout: 0
augmentation_config: conf/preprocess.yaml
num_workers: 0 num_workers: 0
subsampling_factor: 1 subsampling_factor: 1
num_encs: 1 num_encs: 1
@ -78,7 +77,7 @@ global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.004 lr: 0.004
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000

2
examples/librispeech/asr1/local/align.sh
Unescape View File

@ -21,7 +21,7 @@ mkdir -p ${output_dir}
python3 -u ${BIN_DIR}/alignment.py \ python3 -u ${BIN_DIR}/alignment.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_config ${decode_config_path} \ --decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.align \ --result_file ${output_dir}/${type}.align \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decode.decode_batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}

6
examples/librispeech/asr1/local/test.sh
Unescape View File

@ -53,7 +53,7 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_config ${decode_config_path} \ --decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.${type}.rsl \ --result_file ${ckpt_prefix}.${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decode.decoding_method ${type} \ --opts decode.decoding_method ${type} \
@ -78,7 +78,7 @@ for type in ctc_greedy_search; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_config ${decode_config_path} \ --decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.${type}.rsl \ --result_file ${ckpt_prefix}.${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decode.decoding_method ${type} \ --opts decode.decoding_method ${type} \
@ -99,7 +99,7 @@ for type in ctc_prefix_beam_search attention_rescoring; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_config ${decode_config_path} \ --decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.${type}.rsl \ --result_file ${ckpt_prefix}.${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decode.decoding_method ${type} \ --opts decode.decoding_method ${type} \

2
examples/librispeech/asr1/local/test_wav.sh
Unescape View File

@ -50,7 +50,7 @@ for type in attention_rescoring; do
python3 -u ${BIN_DIR}/test_wav.py \ python3 -u ${BIN_DIR}/test_wav.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_config ${decode_config_path} \ --decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.rsl \ --result_file ${output_dir}/${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decode.decoding_method ${type} \ --opts decode.decoding_method ${type} \

11
examples/librispeech/asr2/conf/decode/decode_base.yaml
Unescape View File

@ -0,0 +1,11 @@
decode_batch_size: 1
error_rate_type: wer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
beam_size: 10
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

39
examples/librispeech/asr2/conf/transformer.yaml
Unescape View File

@ -1,6 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
# network architecture ############################################
model: # Network Architecture #
############################################
cmvn_file: cmvn_file:
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -33,12 +34,16 @@ model:
lsm_weight: 0.1 # label smoothing option lsm_weight: 0.1 # label smoothing option
length_normalized_loss: false length_normalized_loss: false
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test-clean test_manifest: data/manifest.test-clean
collator: ###########################################
# Dataloader #
###########################################
vocab_filepath: data/lang_char/train_960_unigram5000_units.txt vocab_filepath: data/lang_char/train_960_unigram5000_units.txt
unit_type: spm unit_type: spm
spm_model_prefix: data/lang_char/train_960_unigram5000 spm_model_prefix: data/lang_char/train_960_unigram5000
@ -55,16 +60,18 @@ collator:
batch_frames_in: 0 batch_frames_in: 0
batch_frames_out: 0 batch_frames_out: 0
batch_frames_inout: 0 batch_frames_inout: 0
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
num_workers: 0 num_workers: 0
subsampling_factor: 1 subsampling_factor: 1
num_encs: 1 num_encs: 1
training: ###########################################
# Training #
###########################################
n_epoch: 120 n_epoch: 120
accum_grad: 2 accum_grad: 2
log_interval: 100 log_interval: 1
checkpoint: checkpoint:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
@ -79,23 +86,5 @@ scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000
lr_decay: 1.0 lr_decay: 1.0
decoding:
batch_size: 1
error_rate_type: wer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 10
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

12
examples/librispeech/asr2/local/align.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 3 ];then if [ $# != 4 ];then
echo "usage: ${0} config_path dict_path ckpt_path_prefix" echo "usage: ${0} config_path decode_config_path dict_path ckpt_path_prefix"
exit -1 exit -1
fi fi
@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
dict_path=$2 decode_config_path=$2
ckpt_prefix=$3 dict_path=$3
ckpt_prefix=$4
batch_size=1 batch_size=1
output_dir=${ckpt_prefix} output_dir=${ckpt_prefix}
@ -24,9 +25,10 @@ python3 -u ${BIN_DIR}/test.py \
--dict-path ${dict_path} \ --dict-path ${dict_path} \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result-file ${output_dir}/${type}.align \ --result-file ${output_dir}/${type}.align \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in ctc alignment!" echo "Failed in ctc alignment!"

10
examples/librispeech/asr2/local/test.sh
Unescape View File

@ -19,8 +19,9 @@ bpeprefix=data/lang_char/${train_set}_${bpemode}${nbpe}
bpemodel=${bpeprefix}.model bpemodel=${bpeprefix}.model
config_path=conf/transformer.yaml config_path=conf/transformer.yaml
decode_config_path=conf/decode/decode_base.yaml
dict=data/lang_char/${train_set}_${bpemode}${nbpe}_units.txt dict=data/lang_char/${train_set}_${bpemode}${nbpe}_units.txt
ckpt_prefix= ckpt_prefix=exp/transformer/checkpoints/init
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1; source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
@ -79,11 +80,12 @@ for dmethd in attention ctc_greedy_search ctc_prefix_beam_search attention_resco
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--dict-path ${dict} \ --dict-path ${dict} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--result-file ${decode_dir}/data.JOB.json \ --result-file ${decode_dir}/data.JOB.json \
--opts decoding.decoding_method ${dmethd} \ --opts decode.decoding_method ${dmethd} \
--opts decoding.batch_size ${batch_size} \ --opts decode.decode_batch_size ${batch_size} \
--opts data.test_manifest ${feat_recog_dir}/split${nj}/JOB/manifest.${rtask} --opts test_manifest ${feat_recog_dir}/split${nj}/JOB/manifest.${rtask}
score_sclite.sh --bpe ${nbpe} --bpemodel ${bpemodel} --wer false ${decode_dir} ${dict} score_sclite.sh --bpe ${nbpe} --bpemodel ${bpemodel} --wer false ${decode_dir} ${dict}

8
examples/librispeech/asr2/run.sh
Unescape View File

@ -9,12 +9,14 @@ gpus=0,1,2,3,4,5,6,7
stage=0 stage=0
stop_stage=50 stop_stage=50
conf_path=conf/transformer.yaml conf_path=conf/transformer.yaml
dict_path=lang_char/train_960_unigram5000_units.txt decode_conf_path=conf/decode/decode_base.yaml
dict_path=data/lang_char/train_960_unigram5000_units.txt
avg_num=10 avg_num=10
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1; source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
avg_ckpt=avg_${avg_num} avg_ckpt=avg_${avg_num}
avg_ckpt=init
ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}') ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}')
echo "checkpoint name ${ckpt}" echo "checkpoint name ${ckpt}"
@ -35,7 +37,7 @@ fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# attetion resocre decoder # attetion resocre decoder
./local/test.sh ${conf_path} ${dict_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 ./local/test.sh ${conf_path} ${decode_conf_path} ${dict_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
@ -45,7 +47,7 @@ fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
# ctc alignment of test data # ctc alignment of test data
CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${dict_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${decode_conf_path} ${dict_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then

28
examples/other/1xt2x/aishell/conf/deepspeech2.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.00 min_output_input_ratio: 0.00
max_output_input_ratio: .inf max_output_input_ratio: .inf
collator: ###########################################
# Dataloader #
###########################################
batch_size: 64 # one gpu batch_size: 64 # one gpu
mean_std_filepath: data/mean_std.npz mean_std_filepath: data/mean_std.npz
unit_type: char unit_type: char
@ -34,7 +38,9 @@ collator:
shuffle_method: batch_shuffle shuffle_method: batch_shuffle
num_workers: 2 num_workers: 2
model: ############################################
# Network Architecture #
############################################
num_conv_layers: 2 num_conv_layers: 2
num_rnn_layers: 3 num_rnn_layers: 3
rnn_layer_size: 1024 rnn_layer_size: 1024
@ -42,7 +48,9 @@ model:
share_rnn_weights: False share_rnn_weights: False
blank_id: 4333 blank_id: 4333
training: ###########################################
# Training #
###########################################
n_epoch: 80 n_epoch: 80
accum_grad: 1 accum_grad: 1
lr: 2e-3 lr: 2e-3
@ -54,14 +62,4 @@ training:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 32
error_rate_type: cer
decoding_method: ctc_beam_search
lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
alpha: 2.6
beta: 5.0
beam_size: 300
cutoff_prob: 0.99
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/other/1xt2x/aishell/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,10 @@
decode_batch_size: 32
error_rate_type: cer
decoding_method: ctc_beam_search
lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
alpha: 2.6
beta: 5.0
beam_size: 300
cutoff_prob: 0.99
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/other/1xt2x/aishell/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 3 ];then if [ $# != 4 ];then
echo "usage: ${0} config_path ckpt_path_prefix model_type" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
exit -1 exit -1
fi fi
@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
model_type=$3 ckpt_prefix=$3
model_type=$4
# download language model # download language model
bash local/download_lm_ch.sh bash local/download_lm_ch.sh
@ -21,6 +22,7 @@ fi
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.rsl \ --result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--model_type ${model_type} --model_type ${model_type}

3
examples/other/1xt2x/aishell/run.sh
Unescape View File

@ -5,6 +5,7 @@ source path.sh
stage=0 stage=0
stop_stage=100 stop_stage=100
conf_path=conf/deepspeech2.yaml conf_path=conf/deepspeech2.yaml
decode_conf_path=conf/tuning/decode.yaml
avg_num=1 avg_num=1
model_type=offline model_type=offline
gpus=2 gpus=2
@ -23,6 +24,6 @@ fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1 CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
fi fi

27
examples/other/1xt2x/baidu_en8k/conf/deepspeech2.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test-clean test_manifest: data/manifest.test-clean
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.00 min_output_input_ratio: 0.00
max_output_input_ratio: .inf max_output_input_ratio: .inf
collator: ###########################################
# Dataloader #
###########################################
batch_size: 64 # one gpu batch_size: 64 # one gpu
mean_std_filepath: data/mean_std.npz mean_std_filepath: data/mean_std.npz
unit_type: char unit_type: char
@ -34,7 +38,9 @@ collator:
shuffle_method: batch_shuffle shuffle_method: batch_shuffle
num_workers: 2 num_workers: 2
model: ############################################
# Network Architecture #
############################################
num_conv_layers: 2 num_conv_layers: 2
num_rnn_layers: 3 num_rnn_layers: 3
rnn_layer_size: 1024 rnn_layer_size: 1024
@ -42,7 +48,9 @@ model:
share_rnn_weights: False share_rnn_weights: False
blank_id: 28 blank_id: 28
training: ###########################################
# Training #
###########################################
n_epoch: 80 n_epoch: 80
accum_grad: 1 accum_grad: 1
lr: 2e-3 lr: 2e-3
@ -54,14 +62,3 @@ training:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 32
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 1.4
beta: 0.35
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/other/1xt2x/baidu_en8k/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,10 @@
decode_batch_size: 32
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 1.4
beta: 0.35
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/other/1xt2x/baidu_en8k/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 3 ];then if [ $# != 4 ];then
echo "usage: ${0} config_path ckpt_path_prefix model_type" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
exit -1 exit -1
fi fi
@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
model_type=$3 ckpt_prefix=$3
model_type=$4
# download language model # download language model
bash local/download_lm_en.sh bash local/download_lm_en.sh
@ -21,6 +22,7 @@ fi
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.rsl \ --result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--model_type ${model_type} --model_type ${model_type}

3
examples/other/1xt2x/baidu_en8k/run.sh
Unescape View File

@ -5,6 +5,7 @@ source path.sh
stage=0 stage=0
stop_stage=100 stop_stage=100
conf_path=conf/deepspeech2.yaml conf_path=conf/deepspeech2.yaml
decode_conf_path=conf/tuning/decode.yaml
avg_num=1 avg_num=1
model_type=offline model_type=offline
gpus=0 gpus=0
@ -23,6 +24,6 @@ fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1 CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
fi fi

27
examples/other/1xt2x/librispeech/conf/deepspeech2.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test-clean test_manifest: data/manifest.test-clean
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.00 min_output_input_ratio: 0.00
max_output_input_ratio: .inf max_output_input_ratio: .inf
collator: ###########################################
# Dataloader #
###########################################
batch_size: 64 # one gpu batch_size: 64 # one gpu
mean_std_filepath: data/mean_std.npz mean_std_filepath: data/mean_std.npz
unit_type: char unit_type: char
@ -34,7 +38,9 @@ collator:
shuffle_method: batch_shuffle shuffle_method: batch_shuffle
num_workers: 2 num_workers: 2
model: ############################################
# Network Architecture #
############################################
num_conv_layers: 2 num_conv_layers: 2
num_rnn_layers: 3 num_rnn_layers: 3
rnn_layer_size: 2048 rnn_layer_size: 2048
@ -42,7 +48,9 @@ model:
share_rnn_weights: True share_rnn_weights: True
blank_id: 28 blank_id: 28
training: ###########################################
# Training #
###########################################
n_epoch: 80 n_epoch: 80
accum_grad: 1 accum_grad: 1
lr: 2e-3 lr: 2e-3
@ -54,14 +62,3 @@ training:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 32
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/other/1xt2x/librispeech/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,10 @@
decode_batch_size: 32
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/other/1xt2x/librispeech/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 3 ];then if [ $# != 4 ];then
echo "usage: ${0} config_path ckpt_path_prefix model_type" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
exit -1 exit -1
fi fi
@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
model_type=$3 ckpt_prefix=$3
model_type=$4
# download language model # download language model
bash local/download_lm_en.sh bash local/download_lm_en.sh
@ -21,6 +22,7 @@ fi
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.rsl \ --result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--model_type ${model_type} --model_type ${model_type}

3
examples/other/1xt2x/librispeech/run.sh
Unescape View File

@ -5,6 +5,7 @@ source path.sh
stage=0 stage=0
stop_stage=100 stop_stage=100
conf_path=conf/deepspeech2.yaml conf_path=conf/deepspeech2.yaml
decode_conf_path=conf/tuning/decode.yaml
avg_num=1 avg_num=1
model_type=offline model_type=offline
gpus=1 gpus=1
@ -23,5 +24,5 @@ fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1 CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
fi fi

5
examples/other/1xt2x/src_deepspeech2x/bin/test.py
Unescape View File

@ -13,6 +13,7 @@
# limitations under the License. # limitations under the License.
"""Evaluation for DeepSpeech2 model.""" """Evaluation for DeepSpeech2 model."""
from src_deepspeech2x.test_model import DeepSpeech2Tester as Tester from src_deepspeech2x.test_model import DeepSpeech2Tester as Tester
from yacs.config import CfgNode
from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults
from paddlespeech.s2t.training.cli import default_argument_parser from paddlespeech.s2t.training.cli import default_argument_parser
@ -44,6 +45,10 @@ if __name__ == "__main__":
config = get_cfg_defaults(args.model_type) config = get_cfg_defaults(args.model_type)
if args.config: if args.config:
config.merge_from_file(args.config) config.merge_from_file(args.config)
if args.decode_cfg:
decode_confs = CfgNode(new_allowed=True)
decode_confs.merge_from_file(args.decode_cfg)
config.decode = decode_confs
if args.opts: if args.opts:
config.merge_from_list(args.opts) config.merge_from_list(args.opts)
config.freeze() config.freeze()

12
examples/other/1xt2x/src_deepspeech2x/models/ds2/deepspeech2.py
Unescape View File

@ -233,11 +233,11 @@ class DeepSpeech2Model(nn.Layer):
""" """
model = cls(feat_size=dataloader.collate_fn.feature_size, model = cls(feat_size=dataloader.collate_fn.feature_size,
dict_size=len(dataloader.collate_fn.vocab_list), dict_size=len(dataloader.collate_fn.vocab_list),
num_conv_layers=config.model.num_conv_layers, num_conv_layers=config.num_conv_layers,
num_rnn_layers=config.model.num_rnn_layers, num_rnn_layers=config.num_rnn_layers,
rnn_size=config.model.rnn_layer_size, rnn_size=config.rnn_layer_size,
use_gru=config.model.use_gru, use_gru=config.use_gru,
share_rnn_weights=config.model.share_rnn_weights) share_rnn_weights=config.share_rnn_weights)
infos = Checkpoint().load_parameters( infos = Checkpoint().load_parameters(
model, checkpoint_path=checkpoint_path) model, checkpoint_path=checkpoint_path)
logger.info(f"checkpoint info: {infos}") logger.info(f"checkpoint info: {infos}")
@ -250,7 +250,7 @@ class DeepSpeech2Model(nn.Layer):
Parameters Parameters
config: yacs.config.CfgNode config: yacs.config.CfgNode
config.model config
Returns Returns
------- -------
DeepSpeech2Model DeepSpeech2Model

90
examples/other/1xt2x/src_deepspeech2x/test_model.py
Unescape View File

@ -64,7 +64,7 @@ class DeepSpeech2Trainer(Trainer):
super().__init__(config, args) super().__init__(config, args)
def train_batch(self, batch_index, batch_data, msg): def train_batch(self, batch_index, batch_data, msg):
train_conf = self.config.training train_conf = self.config
start = time.time() start = time.time()
# forward # forward
@ -98,7 +98,7 @@ class DeepSpeech2Trainer(Trainer):
iteration_time = time.time() - start iteration_time = time.time() - start
msg += "train time: {:>.3f}s, ".format(iteration_time) msg += "train time: {:>.3f}s, ".format(iteration_time)
msg += "batch size: {}, ".format(self.config.collator.batch_size) msg += "batch size: {}, ".format(self.config.batch_size)
msg += "accum: {}, ".format(train_conf.accum_grad) msg += "accum: {}, ".format(train_conf.accum_grad)
msg += ', '.join('{}: {:>.6f}'.format(k, v) msg += ', '.join('{}: {:>.6f}'.format(k, v)
for k, v in losses_np.items()) for k, v in losses_np.items())
@ -126,7 +126,7 @@ class DeepSpeech2Trainer(Trainer):
total_loss += float(loss) * num_utts total_loss += float(loss) * num_utts
valid_losses['val_loss'].append(float(loss)) valid_losses['val_loss'].append(float(loss))
if (i + 1) % self.config.training.log_interval == 0: if (i + 1) % self.config.log_interval == 0:
valid_dump = {k: np.mean(v) for k, v in valid_losses.items()} valid_dump = {k: np.mean(v) for k, v in valid_losses.items()}
valid_dump['val_history_loss'] = total_loss / num_seen_utts valid_dump['val_history_loss'] = total_loss / num_seen_utts
@ -146,15 +146,15 @@ class DeepSpeech2Trainer(Trainer):
def setup_model(self): def setup_model(self):
config = self.config.clone() config = self.config.clone()
config.defrost() config.defrost()
config.model.feat_size = self.train_loader.collate_fn.feature_size config.feat_size = self.train_loader.collate_fn.feature_size
#config.model.dict_size = self.train_loader.collate_fn.vocab_size #config.dict_size = self.train_loader.collate_fn.vocab_size
config.model.dict_size = len(self.train_loader.collate_fn.vocab_list) config.dict_size = len(self.train_loader.collate_fn.vocab_list)
config.freeze() config.freeze()
if self.args.model_type == 'offline': if self.args.model_type == 'offline':
model = DeepSpeech2Model.from_config(config.model) model = DeepSpeech2Model.from_config(config)
elif self.args.model_type == 'online': elif self.args.model_type == 'online':
model = DeepSpeech2ModelOnline.from_config(config.model) model = DeepSpeech2ModelOnline.from_config(config)
else: else:
raise Exception("wrong model type") raise Exception("wrong model type")
if self.parallel: if self.parallel:
@ -163,17 +163,13 @@ class DeepSpeech2Trainer(Trainer):
logger.info(f"{model}") logger.info(f"{model}")
layer_tools.print_params(model, logger.info) layer_tools.print_params(model, logger.info)
grad_clip = ClipGradByGlobalNormWithLog( grad_clip = ClipGradByGlobalNormWithLog(config.global_grad_clip)
config.training.global_grad_clip)
lr_scheduler = paddle.optimizer.lr.ExponentialDecay( lr_scheduler = paddle.optimizer.lr.ExponentialDecay(
learning_rate=config.training.lr, learning_rate=config.lr, gamma=config.lr_decay, verbose=True)
gamma=config.training.lr_decay,
verbose=True)
optimizer = paddle.optimizer.Adam( optimizer = paddle.optimizer.Adam(
learning_rate=lr_scheduler, learning_rate=lr_scheduler,
parameters=model.parameters(), parameters=model.parameters(),
weight_decay=paddle.regularizer.L2Decay( weight_decay=paddle.regularizer.L2Decay(config.weight_decay),
config.training.weight_decay),
grad_clip=grad_clip) grad_clip=grad_clip)
self.model = model self.model = model
@ -184,59 +180,59 @@ class DeepSpeech2Trainer(Trainer):
def setup_dataloader(self): def setup_dataloader(self):
config = self.config.clone() config = self.config.clone()
config.defrost() config.defrost()
config.collator.keep_transcription_text = False config.keep_transcription_text = False
config.data.manifest = config.data.train_manifest config.manifest = config.train_manifest
train_dataset = ManifestDataset.from_config(config) train_dataset = ManifestDataset.from_config(config)
config.data.manifest = config.data.dev_manifest config.manifest = config.dev_manifest
dev_dataset = ManifestDataset.from_config(config) dev_dataset = ManifestDataset.from_config(config)
config.data.manifest = config.data.test_manifest config.manifest = config.test_manifest
test_dataset = ManifestDataset.from_config(config) test_dataset = ManifestDataset.from_config(config)
if self.parallel: if self.parallel:
batch_sampler = SortagradDistributedBatchSampler( batch_sampler = SortagradDistributedBatchSampler(
train_dataset, train_dataset,
batch_size=config.collator.batch_size, batch_size=config.batch_size,
num_replicas=None, num_replicas=None,
rank=None, rank=None,
shuffle=True, shuffle=True,
drop_last=True, drop_last=True,
sortagrad=config.collator.sortagrad, sortagrad=config.sortagrad,
shuffle_method=config.collator.shuffle_method) shuffle_method=config.shuffle_method)
else: else:
batch_sampler = SortagradBatchSampler( batch_sampler = SortagradBatchSampler(
train_dataset, train_dataset,
shuffle=True, shuffle=True,
batch_size=config.collator.batch_size, batch_size=config.batch_size,
drop_last=True, drop_last=True,
sortagrad=config.collator.sortagrad, sortagrad=config.sortagrad,
shuffle_method=config.collator.shuffle_method) shuffle_method=config.shuffle_method)
collate_fn_train = SpeechCollator.from_config(config) collate_fn_train = SpeechCollator.from_config(config)
config.collator.augmentation_config = "" config.augmentation_config = ""
collate_fn_dev = SpeechCollator.from_config(config) collate_fn_dev = SpeechCollator.from_config(config)
config.collator.keep_transcription_text = True config.keep_transcription_text = True
config.collator.augmentation_config = "" config.augmentation_config = ""
collate_fn_test = SpeechCollator.from_config(config) collate_fn_test = SpeechCollator.from_config(config)
self.train_loader = DataLoader( self.train_loader = DataLoader(
train_dataset, train_dataset,
batch_sampler=batch_sampler, batch_sampler=batch_sampler,
collate_fn=collate_fn_train, collate_fn=collate_fn_train,
num_workers=config.collator.num_workers) num_workers=config.num_workers)
self.valid_loader = DataLoader( self.valid_loader = DataLoader(
dev_dataset, dev_dataset,
batch_size=config.collator.batch_size, batch_size=config.batch_size,
shuffle=False, shuffle=False,
drop_last=False, drop_last=False,
collate_fn=collate_fn_dev) collate_fn=collate_fn_dev)
self.test_loader = DataLoader( self.test_loader = DataLoader(
test_dataset, test_dataset,
batch_size=config.decoding.batch_size, batch_size=config.decode.decode_batch_size,
shuffle=False, shuffle=False,
drop_last=False, drop_last=False,
collate_fn=collate_fn_test) collate_fn=collate_fn_test)
@ -274,7 +270,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
def __init__(self, config, args): def __init__(self, config, args):
self._text_featurizer = TextFeaturizer( self._text_featurizer = TextFeaturizer(
unit_type=config.collator.unit_type, vocab_filepath=None) unit_type=config.unit_type, vocab=None)
super().__init__(config, args) super().__init__(config, args)
def ordid2token(self, texts, texts_len): def ordid2token(self, texts, texts_len):
@ -293,7 +289,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
texts, texts,
texts_len, texts_len,
fout=None): fout=None):
cfg = self.config.decoding cfg = self.config.decode
errors_sum, len_refs, num_ins = 0.0, 0, 0 errors_sum, len_refs, num_ins = 0.0, 0, 0
errors_func = error_rate.char_errors if cfg.error_rate_type == 'cer' else error_rate.word_errors errors_func = error_rate.char_errors if cfg.error_rate_type == 'cer' else error_rate.word_errors
error_rate_func = error_rate.cer if cfg.error_rate_type == 'cer' else error_rate.wer error_rate_func = error_rate.cer if cfg.error_rate_type == 'cer' else error_rate.wer
@ -399,31 +395,3 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
self.export() self.export()
except KeyboardInterrupt: except KeyboardInterrupt:
exit(-1) exit(-1)
def setup(self):
"""Setup the experiment.
"""
paddle.set_device('gpu' if self.args.ngpu > 0 else 'cpu')
self.setup_output_dir()
self.setup_checkpointer()
self.setup_dataloader()
self.setup_model()
self.iteration = 0
self.epoch = 0
def setup_output_dir(self):
"""Create a directory used for output.
"""
# output dir
if self.args.output:
output_dir = Path(self.args.output).expanduser()
output_dir.mkdir(parents=True, exist_ok=True)
else:
output_dir = Path(
self.args.checkpoint_path).expanduser().parent.parent
output_dir.mkdir(parents=True, exist_ok=True)
self.output_dir = output_dir

39
examples/ted_en_zh/st0/conf/transformer.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train.tiny train_manifest: data/manifest.train.tiny
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.01 min_output_input_ratio: 0.01
max_output_input_ratio: 20.0 max_output_input_ratio: 20.0
collator: ###########################################
# Dataloader #
###########################################
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm' unit_type: 'spm'
spm_model_prefix: data/lang_char/bpe_unigram_8000 spm_model_prefix: data/lang_char/bpe_unigram_8000
@ -36,8 +40,9 @@ collator:
num_workers: 2 num_workers: 2
# network architecture ############################################
model: # Network Architecture #
############################################
cmvn_file: "data/mean_std.json" cmvn_file: "data/mean_std.json"
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -71,15 +76,16 @@ model:
lsm_weight: 0.1 # label smoothing option lsm_weight: 0.1 # label smoothing option
length_normalized_loss: false length_normalized_loss: false
###########################################
training: # Training #
###########################################
n_epoch: 120 n_epoch: 120
accum_grad: 2 accum_grad: 2
global_grad_clip: 5.0 global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.004 lr: 0.004
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000
@ -88,22 +94,3 @@ training:
checkpoint: checkpoint:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 5
error_rate_type: char-bleu
decoding_method: fullsentence # 'fullsentence', 'simultaneous'
alpha: 2.5
beta: 0.3
beam_size: 10
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

38
examples/ted_en_zh/st0/conf/transformer_mtl_noam.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.01 min_output_input_ratio: 0.01
max_output_input_ratio: 20.0 max_output_input_ratio: 20.0
collator: ###########################################
# Dataloader #
###########################################
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm' unit_type: 'spm'
spm_model_prefix: data/lang_char/bpe_unigram_8000 spm_model_prefix: data/lang_char/bpe_unigram_8000
@ -36,8 +40,9 @@ collator:
num_workers: 2 num_workers: 2
# network architecture ############################################
model: # Network Architecture #
############################################
cmvn_file: "data/mean_std.json" cmvn_file: "data/mean_std.json"
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -72,14 +77,16 @@ model:
length_normalized_loss: false length_normalized_loss: false
training: ###########################################
# Training #
###########################################
n_epoch: 120 n_epoch: 120
accum_grad: 2 accum_grad: 2
global_grad_clip: 5.0 global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 2.5 lr: 2.5
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: noam scheduler: noam
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000
@ -90,23 +97,4 @@ training:
latest_n: 5 latest_n: 5
decoding:
batch_size: 5
error_rate_type: char-bleu
decoding_method: fullsentence # 'fullsentence', 'simultaneous'
alpha: 2.5
beta: 0.3
beam_size: 10
word_reward: 0.7
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

11
examples/ted_en_zh/st0/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,11 @@
batch_size: 5
error_rate_type: char-bleu
decoding_method: fullsentence # 'fullsentence', 'simultaneous'
beam_size: 10
word_reward: 0.7
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

12
examples/ted_en_zh/st0/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#! /usr/bin/env bash #! /usr/bin/env bash
if [ $# != 2 ];then if [ $# != 3 ];then
echo "usage: ${0} config_path ckpt_path_prefix" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
exit -1 exit -1
fi fi
@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
ckpt_prefix=$3
for type in fullsentence; do for type in fullsentence; do
echo "decoding ${type}" echo "decoding ${type}"
@ -17,10 +18,11 @@ for type in fullsentence; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.${type}.rsl \ --result_file ${ckpt_prefix}.${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"

3
examples/ted_en_zh/st0/run.sh
Unescape View File

@ -6,6 +6,7 @@ gpus=0,1,2,3
stage=0 stage=0
stop_stage=50 stop_stage=50
conf_path=conf/transformer_mtl_noam.yaml conf_path=conf/transformer_mtl_noam.yaml
decode_conf_path=conf/tuning/decode.yaml
avg_num=5 avg_num=5
data_path=./TED_EnZh # path to unzipped data data_path=./TED_EnZh # path to unzipped data
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1; source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
@ -32,7 +33,7 @@ fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then

39
examples/ted_en_zh/st1/conf/transformer.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train.tiny train_manifest: data/manifest.train.tiny
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.01 min_output_input_ratio: 0.01
max_output_input_ratio: 20.0 max_output_input_ratio: 20.0
collator: ###########################################
# Dataloader #
###########################################
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm' unit_type: 'spm'
spm_model_prefix: data/lang_char/bpe_unigram_8000 spm_model_prefix: data/lang_char/bpe_unigram_8000
@ -36,8 +40,9 @@ collator:
num_workers: 2 num_workers: 2
# network architecture ############################################
model: # Network Architecture #
############################################
cmvn_file: None cmvn_file: None
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -72,14 +77,16 @@ model:
length_normalized_loss: false length_normalized_loss: false
training: ###########################################
# Training #
###########################################
n_epoch: 20 n_epoch: 20
accum_grad: 2 accum_grad: 2
global_grad_clip: 5.0 global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.004 lr: 0.004
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000
@ -88,23 +95,3 @@ training:
checkpoint: checkpoint:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 5
error_rate_type: char-bleu
decoding_method: fullsentence # 'fullsentence', 'simultaneous'
alpha: 2.5
beta: 0.3
beam_size: 10
word_reward: 0.7
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

39
examples/ted_en_zh/st1/conf/transformer_mtl_noam.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
@ -10,7 +12,9 @@ data:
min_output_input_ratio: 0.01 min_output_input_ratio: 0.01
max_output_input_ratio: 20.0 max_output_input_ratio: 20.0
collator: ###########################################
# Dataloader #
###########################################
vocab_filepath: data/lang_char/ted_en_zh_bpe8000.txt vocab_filepath: data/lang_char/ted_en_zh_bpe8000.txt
unit_type: 'spm' unit_type: 'spm'
spm_model_prefix: data/lang_char/ted_en_zh_bpe8000 spm_model_prefix: data/lang_char/ted_en_zh_bpe8000
@ -36,8 +40,9 @@ collator:
num_workers: 2 num_workers: 2
# network architecture ############################################
model: # Network Architecture #
############################################
cmvn_file: None cmvn_file: None
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -72,14 +77,16 @@ model:
length_normalized_loss: false length_normalized_loss: false
training: ###########################################
# Training #
###########################################
n_epoch: 20 n_epoch: 20
accum_grad: 2 accum_grad: 2
global_grad_clip: 5.0 global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 2.5 lr: 2.5
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: noam scheduler: noam
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000
@ -88,23 +95,3 @@ training:
checkpoint: checkpoint:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 5
error_rate_type: char-bleu
decoding_method: fullsentence # 'fullsentence', 'simultaneous'
alpha: 2.5
beta: 0.3
beam_size: 10
word_reward: 0.7
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

12
examples/ted_en_zh/st1/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,12 @@
batch_size: 5
error_rate_type: char-bleu
decoding_method: fullsentence # 'fullsentence', 'simultaneous'
beam_size: 10
word_reward: 0.7
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

12
examples/ted_en_zh/st1/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#! /usr/bin/env bash #! /usr/bin/env bash
if [ $# != 2 ];then if [ $# != 3 ];then
echo "usage: ${0} config_path ckpt_path_prefix" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
exit -1 exit -1
fi fi
@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
ckpt_prefix=$3
for type in fullsentence; do for type in fullsentence; do
echo "decoding ${type}" echo "decoding ${type}"
@ -17,10 +18,11 @@ for type in fullsentence; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.${type}.rsl \ --result_file ${ckpt_prefix}.${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"

3
examples/ted_en_zh/st1/run.sh
Unescape View File

@ -7,6 +7,7 @@ gpus=0,1,2,3
stage=1 stage=1
stop_stage=4 stop_stage=4
conf_path=conf/transformer_mtl_noam.yaml conf_path=conf/transformer_mtl_noam.yaml
decode_conf_path=conf/tuning/decode.yaml
ckpt_path= # paddle.98 # (finetune from FAT-ST pretrained model) ckpt_path= # paddle.98 # (finetune from FAT-ST pretrained model)
avg_num=5 avg_num=5
data_path=./TED_EnZh # path to unzipped data data_path=./TED_EnZh # path to unzipped data
@ -38,5 +39,5 @@ fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_pat} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi

77
examples/timit/asr1/conf/transformer.yaml
Unescape View File

@ -1,42 +1,40 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
min_input_len: 0.0 # second
max_input_len: 10.0 # second
min_output_len: 0.0 # tokens
max_output_len: 150.0 # tokens
min_output_input_ratio: 0.005
max_output_input_ratio: 1000.0
collator: ###########################################
# Dataloader #
###########################################
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
spm_model_prefix: ''
unit_type: "word" unit_type: "word"
mean_std_filepath: "" mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml preprocess_config: conf/preprocess.yaml
batch_size: 64
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80 feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
use_dB_normalization: True sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
target_dB: -20 batch_size: 64
random_seed: 0 maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
keep_transcription_text: False maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
sortagrad: True minibatches: 0 # for debug
shuffle_method: batch_shuffle batch_count: auto
num_workers: 2 batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
num_workers: 0
subsampling_factor: 1
num_encs: 1
# network architecture ############################################
model: # Network Architecture #
############################################
cmvn_file: cmvn_file:
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -70,14 +68,16 @@ model:
length_normalized_loss: false length_normalized_loss: false
training: ###########################################
# Training #
###########################################
n_epoch: 50 n_epoch: 50
accum_grad: 1 accum_grad: 1
global_grad_clip: 5.0 global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.004 lr: 0.004
weight_decay: 1e-06 weight_decay: 1.0e-6
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 1200 warmup_steps: 1200
@ -87,24 +87,3 @@ training:
kbest_n: 50 kbest_n: 50
latest_n: 5 latest_n: 5
decoding:
batch_size: 64
error_rate_type: wer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 10
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

11
examples/timit/asr1/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,11 @@
decode_batch_size: 64
error_rate_type: wer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
beam_size: 10
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

10
examples/timit/asr1/local/align.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 2 ];then if [ $# != 3 ];then
echo "usage: ${0} config_path ckpt_path_prefix" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
exit -1 exit -1
fi fi
@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
ckpt_prefix=$3
batch_size=1 batch_size=1
output_dir=${ckpt_prefix} output_dir=${ckpt_prefix}
@ -20,9 +21,10 @@ mkdir -p ${output_dir}
python3 -u ${BIN_DIR}/alignment.py \ python3 -u ${BIN_DIR}/alignment.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.align \ --result_file ${output_dir}/${type}.align \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in ctc alignment!" echo "Failed in ctc alignment!"

22
examples/timit/asr1/local/test.sh
Unescape View File

@ -7,8 +7,8 @@ stop_stage=50
. ${MAIN_ROOT}/utils/parse_options.sh || exit 1; . ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
if [ $# != 2 ];then if [ $# != 3 ];then
echo "usage: ${0} config_path ckpt_path_prefix" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
exit -1 exit -1
fi fi
@ -17,7 +17,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
ckpt_prefix=$3
chunk_mode=false chunk_mode=false
if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
@ -43,10 +44,11 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.${type}.rsl \ --result_file ${ckpt_prefix}.${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"
@ -63,10 +65,11 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.${type}.rsl \ --result_file ${ckpt_prefix}.${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"
@ -82,10 +85,11 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.${type}.rsl \ --result_file ${ckpt_prefix}.${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"

13
examples/timit/asr1/run.sh
Unescape View File

@ -7,6 +7,7 @@ gpus=0,1,2,3
stage=0 stage=0
stop_stage=50 stop_stage=50
conf_path=conf/transformer.yaml conf_path=conf/transformer.yaml
decode_conf_path=conf/tuning/decode.yaml
avg_num=10 avg_num=10
TIMIT_path=/path/to/TIMIT TIMIT_path=/path/to/TIMIT
@ -34,15 +35,15 @@ fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
# ctc alignment of test data # ctc alignment of test data
CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
# if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then
# # export ckpt avg_n # export ckpt avg_n
# CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
# fi fi

27
examples/tiny/asr0/conf/deepspeech2.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.tiny train_manifest: data/manifest.tiny
dev_manifest: data/manifest.tiny dev_manifest: data/manifest.tiny
test_manifest: data/manifest.tiny test_manifest: data/manifest.tiny
@ -11,7 +13,9 @@ data:
max_output_input_ratio: 10.0 max_output_input_ratio: 10.0
collator: ###########################################
# Dataloader #
###########################################
mean_std_filepath: data/mean_std.json mean_std_filepath: data/mean_std.json
unit_type: char unit_type: char
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
@ -35,7 +39,9 @@ collator:
num_workers: 2 num_workers: 2
batch_size: 4 batch_size: 4
model: ############################################
# Network Architecture #
############################################
num_conv_layers: 2 num_conv_layers: 2
num_rnn_layers: 3 num_rnn_layers: 3
rnn_layer_size: 2048 rnn_layer_size: 2048
@ -44,7 +50,9 @@ model:
blank_id: 0 blank_id: 0
training: ###########################################
# Training #
###########################################
n_epoch: 5 n_epoch: 5
accum_grad: 1 accum_grad: 1
lr: 1e-5 lr: 1e-5
@ -57,14 +65,3 @@ training:
latest_n: 2 latest_n: 2
decoding:
batch_size: 128
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

28
examples/tiny/asr0/conf/deepspeech2_online.yaml
Unescape View File

@ -1,5 +1,7 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.tiny train_manifest: data/manifest.tiny
dev_manifest: data/manifest.tiny dev_manifest: data/manifest.tiny
test_manifest: data/manifest.tiny test_manifest: data/manifest.tiny
@ -11,7 +13,9 @@ data:
max_output_input_ratio: 10.0 max_output_input_ratio: 10.0
collator: ###########################################
# Dataloader #
###########################################
mean_std_filepath: data/mean_std.json mean_std_filepath: data/mean_std.json
unit_type: char unit_type: char
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
@ -35,7 +39,9 @@ collator:
num_workers: 0 num_workers: 0
batch_size: 4 batch_size: 4
model: ############################################
# Network Architecture #
############################################
num_conv_layers: 2 num_conv_layers: 2
num_rnn_layers: 4 num_rnn_layers: 4
rnn_layer_size: 2048 rnn_layer_size: 2048
@ -46,7 +52,9 @@ model:
blank_id: 0 blank_id: 0
training: ###########################################
# Training #
###########################################
n_epoch: 5 n_epoch: 5
accum_grad: 1 accum_grad: 1
lr: 1e-5 lr: 1e-5
@ -58,15 +66,3 @@ training:
kbest_n: 3 kbest_n: 3
latest_n: 2 latest_n: 2
decoding:
batch_size: 128
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/tiny/asr0/conf/tuning/chunk_decode.yaml
Unescape View File

@ -0,0 +1,10 @@
decode_batch_size: 128
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/tiny/asr0/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,10 @@
decode_batch_size: 128
error_rate_type: wer
decoding_method: ctc_beam_search
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 500
cutoff_prob: 1.0
cutoff_top_n: 40
num_proc_bsearch: 8

10
examples/tiny/asr0/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 3 ];then if [ $# != 4 ];then
echo "usage: ${0} config_path ckpt_path_prefix model_type" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
exit -1 exit -1
fi fi
@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
model_type=$3 ckpt_prefix=$3
model_type=$4
# download language model # download language model
bash local/download_lm_en.sh bash local/download_lm_en.sh
@ -21,6 +22,7 @@ fi
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.rsl \ --result_file ${ckpt_prefix}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--model_type ${model_type} --model_type ${model_type}

3
examples/tiny/asr0/run.sh
Unescape View File

@ -6,6 +6,7 @@ gpus=0
stage=0 stage=0
stop_stage=100 stop_stage=100
conf_path=conf/deepspeech2.yaml conf_path=conf/deepspeech2.yaml
decode_conf_path=conf/tuning/decode.yaml
avg_num=1 avg_num=1
model_type=offline model_type=offline
@ -32,7 +33,7 @@ fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} || exit -1 CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} || exit -1
fi fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then

106
examples/tiny/asr1/conf/chunk_confermer.yaml
Unescape View File

@ -1,43 +1,6 @@
# https://yaml.org/type/float.html ############################################
data: # Network Architecture #
train_manifest: data/manifest.tiny ############################################
dev_manifest: data/manifest.tiny
test_manifest: data/manifest.tiny
min_input_len: 0.5 # second
max_input_len: 30.0 # second
min_output_len: 0.0 # tokens
max_output_len: 400.0 # tokens
min_output_input_ratio: 0.05
max_output_input_ratio: 10.0
collator:
mean_std_filepath: ""
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_200'
augmentation_config: conf/preprocess.yaml
batch_size: 4
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0
window_ms: 25.0
use_dB_normalization: True
target_dB: -20
random_seed: 0
keep_transcription_text: False
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 2
# network architecture
model:
cmvn_file: "data/mean_std.json" cmvn_file: "data/mean_std.json"
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -80,14 +43,51 @@ model:
length_normalized_loss: false length_normalized_loss: false
training: ###########################################
# Data #
###########################################
train_manifest: data/manifest.tiny
dev_manifest: data/manifest.tiny
test_manifest: data/manifest.tiny
###########################################
# Dataloader #
###########################################
mean_std_filepath: ""
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_200'
preprocess_config: conf/preprocess.yaml
feat_dim: 80
stride_ms: 10.0
window_ms: 25.0
sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
batch_size: 4
maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
minibatches: 0 # for debug
batch_count: auto
batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
augmentation_config: conf/preprocess.yaml
num_workers: 0
subsampling_factor: 1
num_encs: 1
###########################################
# Training #
###########################################
n_epoch: 5 n_epoch: 5
accum_grad: 1 accum_grad: 1
global_grad_clip: 5.0 global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.001 lr: 0.001
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000
@ -96,25 +96,3 @@ training:
checkpoint: checkpoint:
kbest_n: 10 kbest_n: 10
latest_n: 1 latest_n: 1
decoding:
batch_size: 64
error_rate_type: wer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 10
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

104
examples/tiny/asr1/conf/chunk_transformer.yaml
Unescape View File

@ -1,43 +1,6 @@
# https://yaml.org/type/float.html ############################################
data: # Network Architecture #
train_manifest: data/manifest.tiny ############################################
dev_manifest: data/manifest.tiny
test_manifest: data/manifest.tiny
min_input_len: 0.5 # second
max_input_len: 20.0 # second
min_output_len: 0.0 # tokens
max_output_len: 400.0 # tokens
min_output_input_ratio: 0.05
max_output_input_ratio: 10.0
collator:
mean_std_filepath: ""
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_200'
augmentation_config: conf/preprocess.yaml
batch_size: 4
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0
window_ms: 25.0
use_dB_normalization: True
target_dB: -20
random_seed: 0
keep_transcription_text: False
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 2
# network architecture
model:
cmvn_file: "data/mean_std.json" cmvn_file: "data/mean_std.json"
cmvn_file_type: "json" cmvn_file_type: "json"
# encoder related # encoder related
@ -73,14 +36,50 @@ model:
length_normalized_loss: false length_normalized_loss: false
training: # https://yaml.org/type/float.html
###########################################
# Data #
###########################################
train_manifest: data/manifest.tiny
dev_manifest: data/manifest.tiny
test_manifest: data/manifest.tiny
###########################################
# Dataloader #
###########################################
mean_std_filepath: ""
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_200'
preprocess_config: conf/preprocess.yaml
feat_dim: 80
stride_ms: 10.0
window_ms: 25.0
sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
batch_size: 4
maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
minibatches: 0 # for debug
batch_count: auto
batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
num_workers: 0
subsampling_factor: 1
num_encs: 1
###########################################
# Training #
###########################################
n_epoch: 5 n_epoch: 5
accum_grad: 1 accum_grad: 1
global_grad_clip: 5.0 global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.002 lr: 0.002
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000
@ -90,24 +89,3 @@ training:
kbest_n: 10 kbest_n: 10
latest_n: 1 latest_n: 1
decoding:
batch_size: 64
error_rate_type: wer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 10
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

80
examples/tiny/asr1/conf/conformer.yaml
Unescape View File

@ -1,46 +1,4 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
###########################################
# Data #
###########################################
train_manifest: data/manifest.tiny
dev_manifest: data/manifest.tiny
test_manifest: data/manifest.tiny
min_input_len: 0.5 # second
max_input_len: 20.0 # second
min_output_len: 0.0 # tokens
max_output_len: 400.0 # tokens
min_output_input_ratio: 0.05
max_output_input_ratio: 10.0
###########################################
# Dataloader #
###########################################
mean_std_filepath: ""
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_200'
augmentation_config: conf/preprocess.yaml
batch_size: 4
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0
window_ms: 25.0
use_dB_normalization: True
target_dB: -20
random_seed: 0
keep_transcription_text: False
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 2
############################################ ############################################
# Network Architecture # # Network Architecture #
############################################ ############################################
@ -83,7 +41,41 @@ model_conf:
########################################### ###########################################
# training # # Data #
###########################################
train_manifest: data/manifest.tiny
dev_manifest: data/manifest.tiny
test_manifest: data/manifest.tiny
###########################################
# Dataloader #
###########################################
mean_std_filepath: ""
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_200'
preprocess_config: conf/preprocess.yaml
feat_dim: 80
stride_ms: 10.0
window_ms: 25.0
sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
batch_size: 4
maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
minibatches: 0 # for debug
batch_count: auto
batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
num_workers: 0
subsampling_factor: 1
num_encs: 1
###########################################
# Training #
########################################### ###########################################
n_epoch: 5 n_epoch: 5
accum_grad: 4 accum_grad: 4
@ -91,7 +83,7 @@ global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.002 lr: 0.002
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000

76
examples/tiny/asr1/conf/transformer.yaml
Unescape View File

@ -1,44 +1,4 @@
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
###########################################
# Data #
###########################################
train_manifest: data/manifest.tiny
dev_manifest: data/manifest.tiny
test_manifest: data/manifest.tiny
min_input_len: 0.5 # second
max_input_len: 20.0 # second
min_output_len: 0.0 # tokens
max_output_len: 400.0 # tokens
min_output_input_ratio: 0.05
max_output_input_ratio: 10.0
###########################################
# Dataloader #
###########################################
mean_std_filepath: data/mean_std.json
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_200'
augmentation_config: conf/preprocess.yaml
batch_size: 4
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0
window_ms: 25.0
use_dB_normalization: True
target_dB: -20
random_seed: 0
keep_transcription_text: False
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 2
############################################ ############################################
# Network Architecture # # Network Architecture #
############################################ ############################################
@ -74,9 +34,41 @@ model_conf:
lsm_weight: 0.1 # label smoothing option lsm_weight: 0.1 # label smoothing option
length_normalized_loss: false length_normalized_loss: false
###########################################
# Data #
###########################################
train_manifest: data/manifest.tiny
dev_manifest: data/manifest.tiny
test_manifest: data/manifest.tiny
###########################################
# Dataloader #
###########################################
mean_std_filepath: data/mean_std.json
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_200'
preprocess_config: conf/preprocess.yaml
feat_dim: 80
stride_ms: 10.0
window_ms: 25.0
sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
batch_size: 4
maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
minibatches: 0 # for debug
batch_count: auto
batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
num_workers: 0
subsampling_factor: 1
num_encs: 1
########################################### ###########################################
# training # # Training #
########################################### ###########################################
n_epoch: 5 n_epoch: 5
accum_grad: 1 accum_grad: 1
@ -84,7 +76,7 @@ global_grad_clip: 5.0
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.002 lr: 0.002
weight_decay: 1e-06 weight_decay: 1.0e-06
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 25000 warmup_steps: 25000

11
examples/tiny/asr1/conf/tuning/chunk_decode.yaml
Unescape View File

@ -0,0 +1,11 @@
decode_batch_size: 8 #64
error_rate_type: wer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
beam_size: 10
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

11
examples/tiny/asr1/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,11 @@
decode_batch_size: 8 #64
error_rate_type: wer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
beam_size: 10
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

10
examples/tiny/asr1/local/align.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 2 ];then if [ $# != 3 ];then
echo "usage: ${0} config_path ckpt_path_prefix" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
exit -1 exit -1
fi fi
@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
ckpt_prefix=$3
batch_size=1 batch_size=1
output_dir=${ckpt_prefix} output_dir=${ckpt_prefix}
@ -20,9 +21,10 @@ mkdir -p ${output_dir}
python3 -u ${BIN_DIR}/alignment.py \ python3 -u ${BIN_DIR}/alignment.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.align \ --result_file ${output_dir}/${type}.align \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in ctc alignment!" echo "Failed in ctc alignment!"

17
examples/tiny/asr1/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 2 ];then if [ $# != 3 ];then
echo "usage: ${0} config_path ckpt_path_prefix" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
exit -1 exit -1
fi fi
@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
ckpt_prefix=$3
chunk_mode=false chunk_mode=false
if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
@ -33,10 +34,11 @@ for type in attention ctc_greedy_search; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.${type}.rsl \ --result_file ${ckpt_prefix}.${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"
@ -50,10 +52,11 @@ for type in ctc_prefix_beam_search attention_rescoring; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${ckpt_prefix}.${type}.rsl \ --result_file ${ckpt_prefix}.${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"

5
examples/tiny/asr1/run.sh
Unescape View File

@ -6,6 +6,7 @@ gpus=0
stage=0 stage=0
stop_stage=50 stop_stage=50
conf_path=conf/transformer.yaml conf_path=conf/transformer.yaml
decode_conf_path=conf/tuning/decode.yaml
avg_num=1 avg_num=1
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1; source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
@ -31,12 +32,12 @@ fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
# ctc alignment of test data # ctc alignment of test data
CUDA_VISIBLE_DEVICES=${gpus} ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=${gpus} ./local/align.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then

77
examples/wenetspeech/asr1/conf/conformer.yaml
Unescape View File

@ -1,5 +1,8 @@
# network architecture ############################################
model: # Network Architecture #
############################################
cmvn_file:
cmvn_file_type: "json"
# encoder related # encoder related
encoder: conformer encoder: conformer
encoder_conf: encoder_conf:
@ -37,43 +40,41 @@ model:
length_normalized_loss: false length_normalized_loss: false
# https://yaml.org/type/float.html # https://yaml.org/type/float.html
data: ###########################################
# Data #
###########################################
train_manifest: data/manifest.train train_manifest: data/manifest.train
dev_manifest: data/manifest.dev dev_manifest: data/manifest.dev
test_manifest: data/manifest.test test_manifest: data/manifest.test
min_input_len: 0.1 # second
max_input_len: 12.0 # second
min_output_len: 1.0
max_output_len: 400.0
min_output_input_ratio: 0.05
max_output_input_ratio: 10.0
collator: ###########################################
# Dataloader #
###########################################
vocab_filepath: data/lang_char/vocab.txt vocab_filepath: data/lang_char/vocab.txt
unit_type: 'char' unit_type: 'char'
preprocess_config: conf/preprocess.yaml
spm_model_prefix: '' spm_model_prefix: ''
augmentation_config: conf/preprocess.yaml
batch_size: 64
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80 feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0 stride_ms: 10.0
window_ms: 25.0 window_ms: 25.0
use_dB_normalization: True sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs
target_dB: -20 batch_size: 64
random_seed: 0 maxlen_in: 512 # if input length > maxlen-in, batchsize is automatically reduced
keep_transcription_text: False maxlen_out: 150 # if output length > maxlen-out, batchsize is automatically reduced
sortagrad: True minibatches: 0 # for debug
shuffle_method: batch_shuffle batch_count: auto
num_workers: 2 batch_bins: 0
batch_frames_in: 0
batch_frames_out: 0
batch_frames_inout: 0
num_workers: 0
subsampling_factor: 1
num_encs: 1
training: ###########################################
# Training #
###########################################
n_epoch: 240 n_epoch: 240
accum_grad: 16 accum_grad: 16
global_grad_clip: 5.0 global_grad_clip: 5.0
@ -84,28 +85,8 @@ training:
optim: adam optim: adam
optim_conf: optim_conf:
lr: 0.001 lr: 0.001
weight_decay: 1e-6 weight_decay: 1.0e-6
scheduler: warmuplr scheduler: warmuplr
scheduler_conf: scheduler_conf:
warmup_steps: 5000 warmup_steps: 5000
lr_decay: 1.0 lr_decay: 1.0
decoding:
batch_size: 128
error_rate_type: cer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 10
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

11
examples/wenetspeech/asr1/conf/tuning/decode.yaml
Unescape View File

@ -0,0 +1,11 @@
decode_batch_size: 128
error_rate_type: cer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
beam_size: 10
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.
# >0: for decoding, use fixed chunk size as set.
# 0: used for training, it's prohibited here.
num_decoding_left_chunks: -1 # number of left chunks for decoding. Defaults to -1.
simulate_streaming: False # simulate streaming inference. Defaults to False.

17
examples/wenetspeech/asr1/local/test.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 2 ];then if [ $# != 3 ];then
echo "usage: ${0} config_path ckpt_path_prefix" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
exit -1 exit -1
fi fi
@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
ckpt_prefix=$3
chunk_mode=false chunk_mode=false
if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
@ -36,10 +37,11 @@ for type in attention ctc_greedy_search; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.rsl \ --result_file ${output_dir}/${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"
@ -55,10 +57,11 @@ for type in ctc_prefix_beam_search attention_rescoring; do
python3 -u ${BIN_DIR}/test.py \ python3 -u ${BIN_DIR}/test.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.rsl \ --result_file ${output_dir}/${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} --opts decode.decode_batch_size ${batch_size}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"

14
examples/wenetspeech/asr1/local/test_wav.sh
Unescape View File

@ -1,7 +1,7 @@
#!/bin/bash #!/bin/bash
if [ $# != 3 ];then if [ $# != 4 ];then
echo "usage: ${0} config_path ckpt_path_prefix audio_file" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix audio_file"
exit -1 exit -1
fi fi
@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
echo "using $ngpu gpus..." echo "using $ngpu gpus..."
config_path=$1 config_path=$1
ckpt_prefix=$2 decode_config_path=$2
audio_file=$3 ckpt_prefix=$3
audio_file=$4
mkdir -p data mkdir -p data
wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/ wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/
@ -43,10 +44,11 @@ for type in attention_rescoring; do
python3 -u ${BIN_DIR}/test_wav.py \ python3 -u ${BIN_DIR}/test_wav.py \
--ngpu ${ngpu} \ --ngpu ${ngpu} \
--config ${config_path} \ --config ${config_path} \
--decode_cfg ${decode_config_path} \
--result_file ${output_dir}/${type}.rsl \ --result_file ${output_dir}/${type}.rsl \
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decoding.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decoding.batch_size ${batch_size} \ --opts decode.decode_batch_size ${batch_size} \
--audio_file ${audio_file} --audio_file ${audio_file}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then

8
examples/wenetspeech/asr1/run.sh
Unescape View File

@ -7,7 +7,7 @@ gpus=0,1,2,3,4,5,6,7
stage=0 stage=0
stop_stage=100 stop_stage=100
conf_path=conf/conformer.yaml conf_path=conf/conformer.yaml
decode_conf_path=conf/tuning/decode.yaml
average_checkpoint=true average_checkpoint=true
avg_num=10 avg_num=10
@ -36,12 +36,12 @@ fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# test ckpt avg_n # test ckpt avg_n
CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
# ctc alignment of test data # ctc alignment of test data
CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
fi fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
@ -51,5 +51,5 @@ fi
if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
# test a single .wav file # test a single .wav file
CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1 CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1
fi fi

32
paddlespeech/s2t/exps/deepspeech2/bin/deploy/runtime.py
Unescape View File

@ -80,13 +80,13 @@ def inference(config, args):
def start_server(config, args): def start_server(config, args):
"""Start the ASR server""" """Start the ASR server"""
config.defrost() config.defrost()
config.data.manifest = config.data.test_manifest config.manifest = config.test_manifest
dataset = ManifestDataset.from_config(config) dataset = ManifestDataset.from_config(config)
config.collator.augmentation_config = "" config.augmentation_config = ""
config.collator.keep_transcription_text = True config.keep_transcription_text = True
config.collator.batch_size = 1 config.batch_size = 1
config.collator.num_workers = 0 config.num_workers = 0
collate_fn = SpeechCollator.from_config(config) collate_fn = SpeechCollator.from_config(config)
test_loader = DataLoader(dataset, collate_fn=collate_fn, num_workers=0) test_loader = DataLoader(dataset, collate_fn=collate_fn, num_workers=0)
@ -105,14 +105,14 @@ def start_server(config, args):
paddle.to_tensor(audio), paddle.to_tensor(audio),
paddle.to_tensor(audio_len), paddle.to_tensor(audio_len),
vocab_list=test_loader.collate_fn.vocab_list, vocab_list=test_loader.collate_fn.vocab_list,
decoding_method=config.decoding.decoding_method, decoding_method=config.decode.decoding_method,
lang_model_path=config.decoding.lang_model_path, lang_model_path=config.decode.lang_model_path,
beam_alpha=config.decoding.alpha, beam_alpha=config.decode.alpha,
beam_beta=config.decoding.beta, beam_beta=config.decode.beta,
beam_size=config.decoding.beam_size, beam_size=config.decode.beam_size,
cutoff_prob=config.decoding.cutoff_prob, cutoff_prob=config.decode.cutoff_prob,
cutoff_top_n=config.decoding.cutoff_top_n, cutoff_top_n=config.decode.cutoff_top_n,
num_processes=config.decoding.num_proc_bsearch) num_processes=config.decode.num_proc_bsearch)
return result_transcript[0] return result_transcript[0]
# warming up with utterrances sampled from Librispeech # warming up with utterrances sampled from Librispeech
@ -179,12 +179,16 @@ if __name__ == "__main__":
config = get_cfg_defaults() config = get_cfg_defaults()
if args.config: if args.config:
config.merge_from_file(args.config) config.merge_from_file(args.config)
if args.decode_cfg:
decode_confs = CfgNode(new_allowed=True)
decode_confs.merge_from_file(args.decode_cfg)
config.decode = decode_confs
if args.opts: if args.opts:
config.merge_from_list(args.opts) config.merge_from_list(args.opts)
config.freeze() config.freeze()
print(config) print(config)
args.warmup_manifest = config.data.test_manifest args.warmup_manifest = config.test_manifest
print_arguments(args, globals()) print_arguments(args, globals())
if args.dump_config: if args.dump_config:

32
paddlespeech/s2t/exps/deepspeech2/bin/deploy/server.py
Unescape View File

@ -33,13 +33,13 @@ from paddlespeech.s2t.utils.utility import print_arguments
def start_server(config, args): def start_server(config, args):
"""Start the ASR server""" """Start the ASR server"""
config.defrost() config.defrost()
config.data.manifest = config.data.test_manifest config.manifest = config.test_manifest
dataset = ManifestDataset.from_config(config) dataset = ManifestDataset.from_config(config)
config.collator.augmentation_config = "" config.augmentation_config = ""
config.collator.keep_transcription_text = True config.keep_transcription_text = True
config.collator.batch_size = 1 config.batch_size = 1
config.collator.num_workers = 0 config.num_workers = 0
collate_fn = SpeechCollator.from_config(config) collate_fn = SpeechCollator.from_config(config)
test_loader = DataLoader(dataset, collate_fn=collate_fn, num_workers=0) test_loader = DataLoader(dataset, collate_fn=collate_fn, num_workers=0)
@ -62,14 +62,14 @@ def start_server(config, args):
paddle.to_tensor(audio), paddle.to_tensor(audio),
paddle.to_tensor(audio_len), paddle.to_tensor(audio_len),
vocab_list=test_loader.collate_fn.vocab_list, vocab_list=test_loader.collate_fn.vocab_list,
decoding_method=config.decoding.decoding_method, decoding_method=config.decode.decoding_method,
lang_model_path=config.decoding.lang_model_path, lang_model_path=config.decode.lang_model_path,
beam_alpha=config.decoding.alpha, beam_alpha=config.decode.alpha,
beam_beta=config.decoding.beta, beam_beta=config.decode.beta,
beam_size=config.decoding.beam_size, beam_size=config.decode.beam_size,
cutoff_prob=config.decoding.cutoff_prob, cutoff_prob=config.decode.cutoff_prob,
cutoff_top_n=config.decoding.cutoff_top_n, cutoff_top_n=config.decode.cutoff_top_n,
num_processes=config.decoding.num_proc_bsearch) num_processes=config.decode.num_proc_bsearch)
return result_transcript[0] return result_transcript[0]
# warming up with utterrances sampled from Librispeech # warming up with utterrances sampled from Librispeech
@ -114,12 +114,16 @@ if __name__ == "__main__":
config = get_cfg_defaults() config = get_cfg_defaults()
if args.config: if args.config:
config.merge_from_file(args.config) config.merge_from_file(args.config)
if args.decode_cfg:
decode_confs = CfgNode(new_allowed=True)
decode_confs.merge_from_file(args.decode_cfg)
config.decode = decode_confs
if args.opts: if args.opts:
config.merge_from_list(args.opts) config.merge_from_list(args.opts)
config.freeze() config.freeze()
print(config) print(config)
args.warmup_manifest = config.data.test_manifest args.warmup_manifest = config.test_manifest
print_arguments(args, globals()) print_arguments(args, globals())
if args.dump_config: if args.dump_config:

6
paddlespeech/s2t/exps/deepspeech2/bin/test.py
Unescape View File

@ -12,6 +12,8 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
"""Evaluation for DeepSpeech2 model.""" """Evaluation for DeepSpeech2 model."""
from yacs.config import CfgNode
from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults
from paddlespeech.s2t.exps.deepspeech2.model import DeepSpeech2Tester as Tester from paddlespeech.s2t.exps.deepspeech2.model import DeepSpeech2Tester as Tester
from paddlespeech.s2t.training.cli import default_argument_parser from paddlespeech.s2t.training.cli import default_argument_parser
@ -44,6 +46,10 @@ if __name__ == "__main__":
config = get_cfg_defaults(args.model_type) config = get_cfg_defaults(args.model_type)
if args.config: if args.config:
config.merge_from_file(args.config) config.merge_from_file(args.config)
if args.decode_cfg:
decode_confs = CfgNode(new_allowed=True)
decode_confs.merge_from_file(args.decode_cfg)
config.decode = decode_confs
if args.opts: if args.opts:
config.merge_from_list(args.opts) config.merge_from_list(args.opts)
config.freeze() config.freeze()

6
paddlespeech/s2t/exps/deepspeech2/bin/test_export.py
Unescape View File

@ -12,6 +12,8 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
"""Evaluation for DeepSpeech2 model.""" """Evaluation for DeepSpeech2 model."""
from yacs.config import CfgNode
from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults
from paddlespeech.s2t.exps.deepspeech2.model import DeepSpeech2ExportTester as ExportTester from paddlespeech.s2t.exps.deepspeech2.model import DeepSpeech2ExportTester as ExportTester
from paddlespeech.s2t.training.cli import default_argument_parser from paddlespeech.s2t.training.cli import default_argument_parser
@ -49,6 +51,10 @@ if __name__ == "__main__":
config = get_cfg_defaults(args.model_type) config = get_cfg_defaults(args.model_type)
if args.config: if args.config:
config.merge_from_file(args.config) config.merge_from_file(args.config)
if args.decode_cfg:
decode_confs = CfgNode(new_allowed=True)
decode_confs.merge_from_file(args.decode_cfg)
config.decode = decode_confs
if args.opts: if args.opts:
config.merge_from_list(args.opts) config.merge_from_list(args.opts)
config.freeze() config.freeze()

21
paddlespeech/s2t/exps/deepspeech2/bin/test_wav.py
Unescape View File

@ -18,6 +18,7 @@ from pathlib import Path
import paddle import paddle
import soundfile import soundfile
from yacs.config import CfgNode
from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults
from paddlespeech.s2t.frontend.featurizer.text_featurizer import TextFeaturizer from paddlespeech.s2t.frontend.featurizer.text_featurizer import TextFeaturizer
@ -41,7 +42,7 @@ class DeepSpeech2Tester_hub():
self.audio_file = args.audio_file self.audio_file = args.audio_file
self.collate_fn_test = SpeechCollator.from_config(config) self.collate_fn_test = SpeechCollator.from_config(config)
self._text_featurizer = TextFeaturizer( self._text_featurizer = TextFeaturizer(
unit_type=config.collator.unit_type, vocab=None) unit_type=config.unit_type, vocab=None)
def compute_result_transcripts(self, audio, audio_len, vocab_list, cfg): def compute_result_transcripts(self, audio, audio_len, vocab_list, cfg):
result_transcripts = self.model.decode( result_transcripts = self.model.decode(
@ -74,7 +75,7 @@ class DeepSpeech2Tester_hub():
audio = paddle.unsqueeze(audio, axis=0) audio = paddle.unsqueeze(audio, axis=0)
vocab_list = collate_fn_test.vocab_list vocab_list = collate_fn_test.vocab_list
result_transcripts = self.compute_result_transcripts( result_transcripts = self.compute_result_transcripts(
audio, audio_len, vocab_list, cfg.decoding) audio, audio_len, vocab_list, cfg.decode)
logger.info("result_transcripts: " + result_transcripts[0]) logger.info("result_transcripts: " + result_transcripts[0])
def run_test(self): def run_test(self):
@ -110,13 +111,13 @@ class DeepSpeech2Tester_hub():
def setup_model(self): def setup_model(self):
config = self.config.clone() config = self.config.clone()
with UpdateConfig(config): with UpdateConfig(config):
config.model.input_dim = self.collate_fn_test.feature_size config.input_dim = self.collate_fn_test.feature_size
config.model.output_dim = self.collate_fn_test.vocab_size config.output_dim = self.collate_fn_test.vocab_size
if self.args.model_type == 'offline': if self.args.model_type == 'offline':
model = DeepSpeech2Model.from_config(config.model) model = DeepSpeech2Model.from_config(config)
elif self.args.model_type == 'online': elif self.args.model_type == 'online':
model = DeepSpeech2ModelOnline.from_config(config.model) model = DeepSpeech2ModelOnline.from_config(config)
else: else:
raise Exception("wrong model type") raise Exception("wrong model type")
@ -134,8 +135,8 @@ class DeepSpeech2Tester_hub():
self.checkpoint_dir = checkpoint_dir self.checkpoint_dir = checkpoint_dir
self.checkpoint = Checkpoint( self.checkpoint = Checkpoint(
kbest_n=self.config.training.checkpoint.kbest_n, kbest_n=self.config.checkpoint.kbest_n,
latest_n=self.config.training.checkpoint.latest_n) latest_n=self.config.checkpoint.latest_n)
def resume(self): def resume(self):
"""Resume from the checkpoint at checkpoints in the output """Resume from the checkpoint at checkpoints in the output
@ -190,6 +191,10 @@ if __name__ == "__main__":
config = get_cfg_defaults(args.model_type) config = get_cfg_defaults(args.model_type)
if args.config: if args.config:
config.merge_from_file(args.config) config.merge_from_file(args.config)
if args.decode_cfg:
decode_confs = CfgNode(new_allowed=True)
decode_confs.merge_from_file(args.decode_cfg)
config.decode = decode_confs
if args.opts: if args.opts:
config.merge_from_list(args.opts) config.merge_from_list(args.opts)
config.freeze() config.freeze()

11
paddlespeech/s2t/exps/deepspeech2/config.py
Unescape View File

@ -23,17 +23,6 @@ from paddlespeech.s2t.models.ds2_online import DeepSpeech2ModelOnline
def get_cfg_defaults(model_type='offline'): def get_cfg_defaults(model_type='offline'):
_C = CfgNode() _C = CfgNode()
_C.data = ManifestDataset.params()
_C.collator = SpeechCollator.params()
_C.training = DeepSpeech2Trainer.params()
_C.decoding = DeepSpeech2Tester.params()
if model_type == 'offline':
_C.model = DeepSpeech2Model.params()
else:
_C.model = DeepSpeech2ModelOnline.params()
"""Get a yacs CfgNode object with default values for my_project."""
# Return a clone so that the defaults will not be altered
# This is for the "local variable" use pattern
config = _C.clone() config = _C.clone()
config.set_new_allowed(True) config.set_new_allowed(True)
return config return config

132
paddlespeech/s2t/exps/deepspeech2/model.py
Unescape View File

@ -69,8 +69,8 @@ class DeepSpeech2Trainer(Trainer):
super().__init__(config, args) super().__init__(config, args)
def train_batch(self, batch_index, batch_data, msg): def train_batch(self, batch_index, batch_data, msg):
batch_size = self.config.collator.batch_size batch_size = self.config.batch_size
accum_grad = self.config.training.accum_grad accum_grad = self.config.accum_grad
start = time.time() start = time.time()
@ -133,7 +133,7 @@ class DeepSpeech2Trainer(Trainer):
total_loss += float(loss) * num_utts total_loss += float(loss) * num_utts
valid_losses['val_loss'].append(float(loss)) valid_losses['val_loss'].append(float(loss))
if (i + 1) % self.config.training.log_interval == 0: if (i + 1) % self.config.log_interval == 0:
valid_dump = {k: np.mean(v) for k, v in valid_losses.items()} valid_dump = {k: np.mean(v) for k, v in valid_losses.items()}
valid_dump['val_history_loss'] = total_loss / num_seen_utts valid_dump['val_history_loss'] = total_loss / num_seen_utts
@ -154,16 +154,16 @@ class DeepSpeech2Trainer(Trainer):
config = self.config.clone() config = self.config.clone()
with UpdateConfig(config): with UpdateConfig(config):
if self.train: if self.train:
config.model.input_dim = self.train_loader.collate_fn.feature_size config.input_dim = self.train_loader.collate_fn.feature_size
config.model.output_dim = self.train_loader.collate_fn.vocab_size config.output_dim = self.train_loader.collate_fn.vocab_size
else: else:
config.model.input_dim = self.test_loader.collate_fn.feature_size config.input_dim = self.test_loader.collate_fn.feature_size
config.model.output_dim = self.test_loader.collate_fn.vocab_size config.output_dim = self.test_loader.collate_fn.vocab_size
if self.args.model_type == 'offline': if self.args.model_type == 'offline':
model = DeepSpeech2Model.from_config(config.model) model = DeepSpeech2Model.from_config(config)
elif self.args.model_type == 'online': elif self.args.model_type == 'online':
model = DeepSpeech2ModelOnline.from_config(config.model) model = DeepSpeech2ModelOnline.from_config(config)
else: else:
raise Exception("wrong model type") raise Exception("wrong model type")
if self.parallel: if self.parallel:
@ -177,17 +177,13 @@ class DeepSpeech2Trainer(Trainer):
if not self.train: if not self.train:
return return
grad_clip = ClipGradByGlobalNormWithLog( grad_clip = ClipGradByGlobalNormWithLog(config.global_grad_clip)
config.training.global_grad_clip)
lr_scheduler = paddle.optimizer.lr.ExponentialDecay( lr_scheduler = paddle.optimizer.lr.ExponentialDecay(
learning_rate=config.training.lr, learning_rate=config.lr, gamma=config.lr_decay, verbose=True)
gamma=config.training.lr_decay,
verbose=True)
optimizer = paddle.optimizer.Adam( optimizer = paddle.optimizer.Adam(
learning_rate=lr_scheduler, learning_rate=lr_scheduler,
parameters=model.parameters(), parameters=model.parameters(),
weight_decay=paddle.regularizer.L2Decay( weight_decay=paddle.regularizer.L2Decay(config.weight_decay),
config.training.weight_decay),
grad_clip=grad_clip) grad_clip=grad_clip)
self.optimizer = optimizer self.optimizer = optimizer
self.lr_scheduler = lr_scheduler self.lr_scheduler = lr_scheduler
@ -198,66 +194,67 @@ class DeepSpeech2Trainer(Trainer):
config.defrost() config.defrost()
if self.train: if self.train:
# train # train
config.data.manifest = config.data.train_manifest config.manifest = config.train_manifest
train_dataset = ManifestDataset.from_config(config) train_dataset = ManifestDataset.from_config(config)
if self.parallel: if self.parallel:
batch_sampler = SortagradDistributedBatchSampler( batch_sampler = SortagradDistributedBatchSampler(
train_dataset, train_dataset,
batch_size=config.collator.batch_size, batch_size=config.batch_size,
num_replicas=None, num_replicas=None,
rank=None, rank=None,
shuffle=True, shuffle=True,
drop_last=True, drop_last=True,
sortagrad=config.collator.sortagrad, sortagrad=config.sortagrad,
shuffle_method=config.collator.shuffle_method) shuffle_method=config.shuffle_method)
else: else:
batch_sampler = SortagradBatchSampler( batch_sampler = SortagradBatchSampler(
train_dataset, train_dataset,
shuffle=True, shuffle=True,
batch_size=config.collator.batch_size, batch_size=config.batch_size,
drop_last=True, drop_last=True,
sortagrad=config.collator.sortagrad, sortagrad=config.sortagrad,
shuffle_method=config.collator.shuffle_method) shuffle_method=config.shuffle_method)
config.collator.keep_transcription_text = False config.keep_transcription_text = False
collate_fn_train = SpeechCollator.from_config(config) collate_fn_train = SpeechCollator.from_config(config)
self.train_loader = DataLoader( self.train_loader = DataLoader(
train_dataset, train_dataset,
batch_sampler=batch_sampler, batch_sampler=batch_sampler,
collate_fn=collate_fn_train, collate_fn=collate_fn_train,
num_workers=config.collator.num_workers) num_workers=config.num_workers)
# dev # dev
config.data.manifest = config.data.dev_manifest config.manifest = config.dev_manifest
dev_dataset = ManifestDataset.from_config(config) dev_dataset = ManifestDataset.from_config(config)
config.collator.augmentation_config = "" config.augmentation_config = ""
config.collator.keep_transcription_text = False config.keep_transcription_text = False
collate_fn_dev = SpeechCollator.from_config(config) collate_fn_dev = SpeechCollator.from_config(config)
self.valid_loader = DataLoader( self.valid_loader = DataLoader(
dev_dataset, dev_dataset,
batch_size=int(config.collator.batch_size), batch_size=int(config.batch_size),
shuffle=False, shuffle=False,
drop_last=False, drop_last=False,
collate_fn=collate_fn_dev, collate_fn=collate_fn_dev,
num_workers=config.collator.num_workers) num_workers=config.num_workers)
logger.info("Setup train/valid Dataloader!") logger.info("Setup train/valid Dataloader!")
else: else:
# test # test
config.data.manifest = config.data.test_manifest config.manifest = config.test_manifest
test_dataset = ManifestDataset.from_config(config) test_dataset = ManifestDataset.from_config(config)
config.collator.augmentation_config = "" config.augmentation_config = ""
config.collator.keep_transcription_text = True config.keep_transcription_text = True
collate_fn_test = SpeechCollator.from_config(config) collate_fn_test = SpeechCollator.from_config(config)
decode_batch_size = config.get('decode', dict()).get(
'decode_batch_size', 1)
self.test_loader = DataLoader( self.test_loader = DataLoader(
test_dataset, test_dataset,
batch_size=config.decoding.batch_size, batch_size=decode_batch_size,
shuffle=False, shuffle=False,
drop_last=False, drop_last=False,
collate_fn=collate_fn_test, collate_fn=collate_fn_test,
num_workers=config.collator.num_workers) num_workers=config.num_workers)
logger.info("Setup test Dataloader!") logger.info("Setup test Dataloader!")
@ -286,7 +283,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
def __init__(self, config, args): def __init__(self, config, args):
super().__init__(config, args) super().__init__(config, args)
self._text_featurizer = TextFeaturizer( self._text_featurizer = TextFeaturizer(
unit_type=config.collator.unit_type, vocab=None) unit_type=config.unit_type, vocab=None)
def ordid2token(self, texts, texts_len): def ordid2token(self, texts, texts_len):
""" ord() id to chr() chr """ """ ord() id to chr() chr """
@ -304,17 +301,17 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
texts, texts,
texts_len, texts_len,
fout=None): fout=None):
cfg = self.config.decoding decode_cfg = self.config.decode
errors_sum, len_refs, num_ins = 0.0, 0, 0 errors_sum, len_refs, num_ins = 0.0, 0, 0
errors_func = error_rate.char_errors if cfg.error_rate_type == 'cer' else error_rate.word_errors errors_func = error_rate.char_errors if decode_cfg.error_rate_type == 'cer' else error_rate.word_errors
error_rate_func = error_rate.cer if cfg.error_rate_type == 'cer' else error_rate.wer error_rate_func = error_rate.cer if decode_cfg.error_rate_type == 'cer' else error_rate.wer
vocab_list = self.test_loader.collate_fn.vocab_list vocab_list = self.test_loader.collate_fn.vocab_list
target_transcripts = self.ordid2token(texts, texts_len) target_transcripts = self.ordid2token(texts, texts_len)
result_transcripts = self.compute_result_transcripts(audio, audio_len, result_transcripts = self.compute_result_transcripts(
vocab_list, cfg) audio, audio_len, vocab_list, decode_cfg)
for utt, target, result in zip(utts, target_transcripts, for utt, target, result in zip(utts, target_transcripts,
result_transcripts): result_transcripts):
@ -327,29 +324,31 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
logger.info(f"Utt: {utt}") logger.info(f"Utt: {utt}")
logger.info(f"Ref: {target}") logger.info(f"Ref: {target}")
logger.info(f"Hyp: {result}") logger.info(f"Hyp: {result}")
logger.info("Current error rate [%s] = %f" % logger.info(
(cfg.error_rate_type, error_rate_func(target, result))) "Current error rate [%s] = %f" %
(decode_cfg.error_rate_type, error_rate_func(target, result)))
return dict( return dict(
errors_sum=errors_sum, errors_sum=errors_sum,
len_refs=len_refs, len_refs=len_refs,
num_ins=num_ins, num_ins=num_ins,
error_rate=errors_sum / len_refs, error_rate=errors_sum / len_refs,
error_rate_type=cfg.error_rate_type) error_rate_type=decode_cfg.error_rate_type)
def compute_result_transcripts(self, audio, audio_len, vocab_list, cfg): def compute_result_transcripts(self, audio, audio_len, vocab_list,
decode_cfg):
result_transcripts = self.model.decode( result_transcripts = self.model.decode(
audio, audio,
audio_len, audio_len,
vocab_list, vocab_list,
decoding_method=cfg.decoding_method, decoding_method=decode_cfg.decoding_method,
lang_model_path=cfg.lang_model_path, lang_model_path=decode_cfg.lang_model_path,
beam_alpha=cfg.alpha, beam_alpha=decode_cfg.alpha,
beam_beta=cfg.beta, beam_beta=decode_cfg.beta,
beam_size=cfg.beam_size, beam_size=decode_cfg.beam_size,
cutoff_prob=cfg.cutoff_prob, cutoff_prob=decode_cfg.cutoff_prob,
cutoff_top_n=cfg.cutoff_top_n, cutoff_top_n=decode_cfg.cutoff_top_n,
num_processes=cfg.num_proc_bsearch) num_processes=decode_cfg.num_proc_bsearch)
return result_transcripts return result_transcripts
@ -358,7 +357,6 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
def test(self): def test(self):
logger.info(f"Test Total Examples: {len(self.test_loader.dataset)}") logger.info(f"Test Total Examples: {len(self.test_loader.dataset)}")
self.model.eval() self.model.eval()
cfg = self.config
error_rate_type = None error_rate_type = None
errors_sum, len_refs, num_ins = 0.0, 0, 0 errors_sum, len_refs, num_ins = 0.0, 0, 0
with jsonlines.open(self.args.result_file, 'w') as fout: with jsonlines.open(self.args.result_file, 'w') as fout:
@ -412,11 +410,10 @@ class DeepSpeech2ExportTester(DeepSpeech2Tester):
if self.args.enable_auto_log is True: if self.args.enable_auto_log is True:
from paddlespeech.s2t.utils.log import Autolog from paddlespeech.s2t.utils.log import Autolog
self.autolog = Autolog( self.autolog = Autolog(
batch_size=self.config.decoding.batch_size, batch_size=self.config.decode.decode_batch_size,
model_name="deepspeech2", model_name="deepspeech2",
model_precision="fp32").getlog() model_precision="fp32").getlog()
self.model.eval() self.model.eval()
cfg = self.config
error_rate_type = None error_rate_type = None
errors_sum, len_refs, num_ins = 0.0, 0, 0 errors_sum, len_refs, num_ins = 0.0, 0, 0
with jsonlines.open(self.args.result_file, 'w') as fout: with jsonlines.open(self.args.result_file, 'w') as fout:
@ -441,7 +438,8 @@ class DeepSpeech2ExportTester(DeepSpeech2Tester):
if self.args.enable_auto_log is True: if self.args.enable_auto_log is True:
self.autolog.report() self.autolog.report()
def compute_result_transcripts(self, audio, audio_len, vocab_list, cfg): def compute_result_transcripts(self, audio, audio_len, vocab_list,
decode_cfg):
if self.args.model_type == "online": if self.args.model_type == "online":
output_probs, output_lens = self.static_forward_online(audio, output_probs, output_lens = self.static_forward_online(audio,
audio_len) audio_len)
@ -454,13 +452,15 @@ class DeepSpeech2ExportTester(DeepSpeech2Tester):
self.predictor.clear_intermediate_tensor() self.predictor.clear_intermediate_tensor()
self.predictor.try_shrink_memory() self.predictor.try_shrink_memory()
self.model.decoder.init_decode(cfg.alpha, cfg.beta, cfg.lang_model_path, self.model.decoder.init_decode(decode_cfg.alpha, decode_cfg.beta,
vocab_list, cfg.decoding_method) decode_cfg.lang_model_path, vocab_list,
decode_cfg.decoding_method)
result_transcripts = self.model.decoder.decode_probs( result_transcripts = self.model.decoder.decode_probs(
output_probs, output_lens, vocab_list, cfg.decoding_method, output_probs, output_lens, vocab_list, decode_cfg.decoding_method,
cfg.lang_model_path, cfg.alpha, cfg.beta, cfg.beam_size, decode_cfg.lang_model_path, decode_cfg.alpha, decode_cfg.beta,
cfg.cutoff_prob, cfg.cutoff_top_n, cfg.num_proc_bsearch) decode_cfg.beam_size, decode_cfg.cutoff_prob,
decode_cfg.cutoff_top_n, decode_cfg.num_proc_bsearch)
#replace the <space> with ' ' #replace the <space> with ' '
result_transcripts = [ result_transcripts = [
self._text_featurizer.detokenize(sentence) self._text_featurizer.detokenize(sentence)
@ -531,12 +531,10 @@ class DeepSpeech2ExportTester(DeepSpeech2Tester):
num_chunk = int(num_chunk) num_chunk = int(num_chunk)
chunk_state_h_box = np.zeros( chunk_state_h_box = np.zeros(
(self.config.model.num_rnn_layers, 1, (self.config.num_rnn_layers, 1, self.config.rnn_layer_size),
self.config.model.rnn_layer_size),
dtype=x.dtype) dtype=x.dtype)
chunk_state_c_box = np.zeros( chunk_state_c_box = np.zeros(
(self.config.model.num_rnn_layers, 1, (self.config.num_rnn_layers, 1, self.config.rnn_layer_size),
self.config.model.rnn_layer_size),
dtype=x.dtype) dtype=x.dtype)
input_names = self.predictor.get_input_names() input_names = self.predictor.get_input_names()

4
paddlespeech/s2t/exps/u2/bin/alignment.py
Unescape View File

@ -43,9 +43,9 @@ if __name__ == "__main__":
config = get_cfg_defaults() config = get_cfg_defaults()
if args.config: if args.config:
config.merge_from_file(args.config) config.merge_from_file(args.config)
if args.decode_config: if args.decode_cfg:
decode_confs = CfgNode(new_allowed=True) decode_confs = CfgNode(new_allowed=True)
decode_confs.merge_from_file(args.decode_config) decode_confs.merge_from_file(args.decode_cfg)
config.decode = decode_confs config.decode = decode_confs
if args.opts: if args.opts:
config.merge_from_list(args.opts) config.merge_from_list(args.opts)

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save