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

Merge branch 'PaddlePaddle:develop' into develop

Browse Source
pull/2615/head
HuangLiangJie 3 years ago committed by GitHub
parent a20ca46eac 1d6f8ea8d5
commit 26a6fb5c75
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 128 additions and 142 deletions
Show Stats Download Patch File Download Diff File

13
examples/wenetspeech/asr1/local/quant.sh
Unescape View File

@ -1,7 +1,8 @@
#!/bin/bash #!/bin/bash
# ./local/quant.sh conf/chunk_conformer_u2pp.yaml conf/tuning/chunk_decode.yaml exp/chunk_conformer_u2pp/checkpoints/avg_10 data/wav.aishell.test.scp
if [ $# != 4 ];then if [ $# != 4 ];then
echo "usage: ${0} config_path decode_config_path ckpt_path_prefix audio_file" echo "usage: ${0} config_path decode_config_path ckpt_path_prefix audio_scp"
exit -1 exit -1
fi fi
@ -11,16 +12,15 @@ echo "using $ngpu gpus..."
config_path=$1 config_path=$1
decode_config_path=$2 decode_config_path=$2
ckpt_prefix=$3 ckpt_prefix=$3
audio_file=$4 audio_scp=$4
mkdir -p data mkdir -p data
wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
exit 1 exit 1
fi fi
if [ ! -f ${audio_file} ]; then if [ ! -f ${audio_scp} ]; then
echo "Plase input the right audio_file path" echo "Plase input the right audio_scp path"
exit 1 exit 1
fi fi
@ -49,7 +49,8 @@ for type in attention_rescoring; do
--checkpoint_path ${ckpt_prefix} \ --checkpoint_path ${ckpt_prefix} \
--opts decode.decoding_method ${type} \ --opts decode.decoding_method ${type} \
--opts decode.decode_batch_size ${batch_size} \ --opts decode.decode_batch_size ${batch_size} \
--audio_file ${audio_file} --num_utts 200 \
--audio_scp ${audio_scp}
if [ $? -ne 0 ]; then if [ $? -ne 0 ]; then
echo "Failed in evaluation!" echo "Failed in evaluation!"

4
examples/wenetspeech/asr1/run.sh
Unescape View File

@ -54,3 +54,7 @@ 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} ${decode_conf_path} exp/${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
if [ ${stage} -le 8 ] && [ ${stop_stage} -ge 8 ]; then
# export quant model, plesae see local/quant.sh
fi

252
paddlespeech/s2t/exps/u2/bin/quant.py
Unescape View File

@ -11,13 +11,9 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# 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 U2 model.""" """Quantzation U2 model."""
import os
import sys
from pathlib import Path
import paddle import paddle
import soundfile from kaldiio import ReadHelper
from paddleslim import PTQ from paddleslim import PTQ
from yacs.config import CfgNode from yacs.config import CfgNode
@ -34,7 +30,7 @@ class U2Infer():
def __init__(self, config, args): def __init__(self, config, args):
self.args = args self.args = args
self.config = config self.config = config
self.audio_file = args.audio_file self.audio_scp = args.audio_scp
self.preprocess_conf = config.preprocess_config self.preprocess_conf = config.preprocess_config
self.preprocess_args = {"train": False} self.preprocess_args = {"train": False}
@ -63,133 +59,112 @@ class U2Infer():
self.model.set_state_dict(model_dict) self.model.set_state_dict(model_dict)
def run(self): def run(self):
check(args.audio_file) cnt = 0
with ReadHelper(f"scp:{self.audio_scp}") as reader:
with paddle.no_grad(): for key, (rate, audio) in reader:
# read assert rate == 16000
audio, sample_rate = soundfile.read( cnt += 1
self.audio_file, dtype="int16", always_2d=True) if cnt > args.num_utts:
audio = audio[:, 0] break
logger.info(f"audio shape: {audio.shape}")
with paddle.no_grad():
# fbank logger.info(f"audio shape: {audio.shape}")
feat = self.preprocessing(audio, **self.preprocess_args)
logger.info(f"feat shape: {feat.shape}") # fbank
feat = self.preprocessing(audio, **self.preprocess_args)
ilen = paddle.to_tensor(feat.shape[0]) logger.info(f"feat shape: {feat.shape}")
xs = paddle.to_tensor(feat, dtype='float32').unsqueeze(0)
decode_config = self.config.decode ilen = paddle.to_tensor(feat.shape[0])
logger.info(f"decode cfg: {decode_config}") xs = paddle.to_tensor(feat, dtype='float32').unsqueeze(0)
reverse_weight = getattr(decode_config, 'reverse_weight', 0.0) decode_config = self.config.decode
result_transcripts = self.model.decode( logger.info(f"decode cfg: {decode_config}")
xs, result_transcripts = self.model.decode(
ilen, xs,
text_feature=self.text_feature, ilen,
decoding_method=decode_config.decoding_method, text_feature=self.text_feature,
beam_size=decode_config.beam_size, decoding_method=decode_config.decoding_method,
ctc_weight=decode_config.ctc_weight, beam_size=decode_config.beam_size,
decoding_chunk_size=decode_config.decoding_chunk_size, ctc_weight=decode_config.ctc_weight,
num_decoding_left_chunks=decode_config.num_decoding_left_chunks, decoding_chunk_size=decode_config.decoding_chunk_size,
simulate_streaming=decode_config.simulate_streaming, num_decoding_left_chunks=decode_config.
reverse_weight=reverse_weight) num_decoding_left_chunks,
rsl = result_transcripts[0][0] simulate_streaming=decode_config.simulate_streaming,
utt = Path(self.audio_file).name reverse_weight=decode_config.reverse_weight)
logger.info(f"hyp: {utt} {rsl}") rsl = result_transcripts[0][0]
# print(self.model) utt = key
# print(self.model.forward_encoder_chunk) logger.info(f"hyp: {utt} {rsl}")
# print(self.model)
logger.info("-------------start quant ----------------------") # print(self.model.forward_encoder_chunk)
batch_size = 1
feat_dim = 80 logger.info("-------------start quant ----------------------")
model_size = 512 batch_size = 1
num_left_chunks = -1 feat_dim = 80
reverse_weight = 0.3 model_size = 512
logger.info( num_left_chunks = -1
f"U2 Export Model Params: batch_size {batch_size}, feat_dim {feat_dim}, model_size {model_size}, num_left_chunks {num_left_chunks}, reverse_weight {reverse_weight}" reverse_weight = 0.3
) logger.info(
f"U2 Export Model Params: batch_size {batch_size}, feat_dim {feat_dim}, model_size {model_size}, num_left_chunks {num_left_chunks}, reverse_weight {reverse_weight}"
# ######################## self.model.forward_encoder_chunk ############ )
# input_spec = [
# # (T,), int16 # ######################## self.model.forward_encoder_chunk ############
# paddle.static.InputSpec(shape=[None], dtype='int16'), # input_spec = [
# ] # # (T,), int16
# self.model.forward_feature = paddle.jit.to_static( # paddle.static.InputSpec(shape=[None], dtype='int16'),
# self.model.forward_feature, input_spec=input_spec) # ]
# self.model.forward_feature = paddle.jit.to_static(
######################### self.model.forward_encoder_chunk ############ # self.model.forward_feature, input_spec=input_spec)
input_spec = [
# xs, (B, T, D) ######################### self.model.forward_encoder_chunk ############
paddle.static.InputSpec( input_spec = [
shape=[batch_size, None, feat_dim], dtype='float32'), # xs, (B, T, D)
# offset, int, but need be tensor paddle.static.InputSpec(
paddle.static.InputSpec(shape=[1], dtype='int32'), shape=[batch_size, None, feat_dim], dtype='float32'),
# required_cache_size, int # offset, int, but need be tensor
num_left_chunks, paddle.static.InputSpec(shape=[1], dtype='int32'),
# att_cache # required_cache_size, int
paddle.static.InputSpec( num_left_chunks,
shape=[None, None, None, None], dtype='float32'), # att_cache
# cnn_cache paddle.static.InputSpec(
paddle.static.InputSpec( shape=[None, None, None, None], dtype='float32'),
shape=[None, None, None, None], dtype='float32') # cnn_cache
] paddle.static.InputSpec(
self.model.forward_encoder_chunk = paddle.jit.to_static( shape=[None, None, None, None], dtype='float32')
self.model.forward_encoder_chunk, input_spec=input_spec) ]
self.model.forward_encoder_chunk = paddle.jit.to_static(
######################### self.model.ctc_activation ######################## self.model.forward_encoder_chunk, input_spec=input_spec)
input_spec = [
# encoder_out, (B,T,D) ######################### self.model.ctc_activation ########################
paddle.static.InputSpec( input_spec = [
shape=[batch_size, None, model_size], dtype='float32') # encoder_out, (B,T,D)
] paddle.static.InputSpec(
self.model.ctc_activation = paddle.jit.to_static( shape=[batch_size, None, model_size], dtype='float32')
self.model.ctc_activation, input_spec=input_spec) ]
self.model.ctc_activation = paddle.jit.to_static(
######################### self.model.forward_attention_decoder ######################## self.model.ctc_activation, input_spec=input_spec)
input_spec = [
# hyps, (B, U) ######################### self.model.forward_attention_decoder ########################
paddle.static.InputSpec(shape=[None, None], dtype='int64'), input_spec = [
# hyps_lens, (B,) # hyps, (B, U)
paddle.static.InputSpec(shape=[None], dtype='int64'), paddle.static.InputSpec(shape=[None, None], dtype='int64'),
# encoder_out, (B,T,D) # hyps_lens, (B,)
paddle.static.InputSpec( paddle.static.InputSpec(shape=[None], dtype='int64'),
shape=[batch_size, None, model_size], dtype='float32'), # encoder_out, (B,T,D)
reverse_weight paddle.static.InputSpec(
] shape=[batch_size, None, model_size], dtype='float32'),
self.model.forward_attention_decoder = paddle.jit.to_static( reverse_weight
self.model.forward_attention_decoder, input_spec=input_spec) ]
################################################################################ self.model.forward_attention_decoder = paddle.jit.to_static(
self.model.forward_attention_decoder, input_spec=input_spec)
# jit save ################################################################################
logger.info(f"export save: {self.args.export_path}")
config = { # jit save
'is_static': True, logger.info(f"export save: {self.args.export_path}")
'combine_params': True, self.ptq.ptq._convert(self.model)
'skip_forward': True paddle.jit.save(
} self.model,
self.ptq.save_quantized_model(self.model, self.args.export_path) self.args.export_path,
# paddle.jit.save( combine_params=True,
# self.model, skip_forward=True)
# self.args.export_path,
# combine_params=True,
# skip_forward=True)
def check(audio_file):
if not os.path.isfile(audio_file):
print("Please input the right audio file path")
sys.exit(-1)
logger.info("checking the audio file format......")
try:
sig, sample_rate = soundfile.read(audio_file)
except Exception as e:
logger.error(str(e))
logger.error(
"can not open the wav file, please check the audio file format")
sys.exit(-1)
logger.info("The sample rate is %d" % sample_rate)
assert (sample_rate == 16000)
logger.info("The audio file format is right")
def main(config, args): def main(config, args):
@ -202,11 +177,16 @@ if __name__ == "__main__":
parser.add_argument( parser.add_argument(
"--result_file", type=str, help="path of save the asr result") "--result_file", type=str, help="path of save the asr result")
parser.add_argument( parser.add_argument(
"--audio_file", type=str, help="path of the input audio file") "--audio_scp", type=str, help="path of the input audio file")
parser.add_argument(
"--num_utts",
type=int,
default=200,
help="num utts for quant calibrition.")
parser.add_argument( parser.add_argument(
"--export_path", "--export_path",
type=str, type=str,
default='export', default='export.jit.quant',
help="path of the input audio file") help="path of the input audio file")
args = parser.parse_args() args = parser.parse_args()

1
tests/unit/cli/test_cli.sh
Unescape View File

@ -16,6 +16,7 @@ paddlespeech asr --model conformer_aishell --input ./zh.wav
paddlespeech asr --model conformer_online_aishell --input ./zh.wav paddlespeech asr --model conformer_online_aishell --input ./zh.wav
paddlespeech asr --model conformer_online_wenetspeech --input ./zh.wav paddlespeech asr --model conformer_online_wenetspeech --input ./zh.wav
paddlespeech asr --model conformer_online_multicn --input ./zh.wav paddlespeech asr --model conformer_online_multicn --input ./zh.wav
paddlespeech asr --model conformer_u2pp_online_wenetspeech --lang zh --input zh.wav
paddlespeech asr --model transformer_librispeech --lang en --input ./en.wav paddlespeech asr --model transformer_librispeech --lang en --input ./en.wav
paddlespeech asr --model deepspeech2offline_aishell --input ./zh.wav paddlespeech asr --model deepspeech2offline_aishell --input ./zh.wav
paddlespeech asr --model deepspeech2online_wenetspeech --input ./zh.wav paddlespeech asr --model deepspeech2online_wenetspeech --input ./zh.wav

Write Preview
Loading…
Cancel
Save