forked from Z-yq/TensorflowASR
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest_asr.py
275 lines (245 loc) · 14 KB
/
test_asr.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
import logging
import os
import numpy as np
from asr.models.conformer_blocks import ConformerEncoder, StreamingConformerEncoder, CTCDecoder, Translator, tf
from utils.speech_featurizers import SpeechFeaturizer
from utils.text_featurizers import TextFeaturizer
from utils.user_config import UserConfig
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
class ASR():
def __init__(self, config):
self.running_config = config['running_config']
self.speech_config = config['speech_config']
self.model_config = config['model_config']
self.opt_config = config['optimizer_config']
self.phone_featurizer = TextFeaturizer(config['inp_config'])
self.text_featurizer = TextFeaturizer(config['tar_config'])
self.speech_featurizer = SpeechFeaturizer(self.speech_config)
self.chunk = self.speech_config['sample_rate'] * self.speech_config['streaming_bucket']
self.compile()
def compile(self):
if not self.speech_config['streaming']:
self.encoder = ConformerEncoder(dmodel=self.model_config['dmodel'],
reduction_factor=self.model_config['reduction_factor'],
num_blocks=self.model_config['num_blocks'],
head_size=self.model_config['head_size'],
num_heads=self.model_config['num_heads'],
kernel_size=self.model_config['kernel_size'],
fc_factor=self.model_config['fc_factor'],
dropout=self.model_config['dropout'],
add_wav_info=self.speech_config['add_wav_info'],
sample_rate=self.speech_config['sample_rate'],
n_mels=self.speech_config['num_feature_bins'],
mel_layer_type=self.speech_config['mel_layer_type'],
mel_layer_trainable=self.speech_config['mel_layer_trainable'],
stride_ms=self.speech_config['stride_ms'],
name="conformer_encoder", )
else:
assert 'Streaming' in self.model_config['name'],'am_data.yml set streaming=True,But model.yml is OfflineCTC'
self.encoder = StreamingConformerEncoder(dmodel=self.model_config['dmodel'],
reduction_factor=self.model_config['reduction_factor'],
num_blocks=self.model_config['num_blocks'],
head_size=self.model_config['head_size'],
num_heads=self.model_config['num_heads'],
kernel_size=self.model_config['kernel_size'],
fc_factor=self.model_config['fc_factor'],
dropout=self.model_config['dropout'],
add_wav_info=self.speech_config['add_wav_info'],
sample_rate=self.speech_config['sample_rate'],
n_mels=self.speech_config['num_feature_bins'],
mel_layer_type=self.speech_config['mel_layer_type'],
mel_layer_trainable=self.speech_config['mel_layer_trainable'],
stride_ms=self.speech_config['stride_ms'],
name="stream_conformer_encoder")
self.encoder.add_chunk_size(
chunk_size=int(self.speech_config['streaming_bucket'] * self.speech_config['sample_rate']),
mel_size=self.speech_config['num_feature_bins'],
hop_size=int(self.speech_config['stride_ms'] * self.speech_config['sample_rate'] // 1000) *
self.model_config['reduction_factor'])
self.encoder.set_inference_func()
self.ctc_model = CTCDecoder(num_classes=self.phone_featurizer.num_classes,
dmodel=self.model_config['dmodel'],
num_blocks=self.model_config['ctcdecoder_num_blocks'],
head_size=self.model_config['head_size'],
num_heads=self.model_config['num_heads'],
kernel_size=self.model_config['ctcdecoder_kernel_size'],
dropout=self.model_config['ctcdecoder_dropout'],
fc_factor=self.model_config['ctcdecoder_fc_factor'],
)
self.translator = Translator(inp_classes=self.phone_featurizer.num_classes,
tar_classes=self.text_featurizer.num_classes,
dmodel=self.model_config['dmodel'],
num_blocks=self.model_config['translator_num_blocks'],
head_size=self.model_config['head_size'],
num_heads=self.model_config['num_heads'],
kernel_size=self.model_config['translator_kernel_size'],
dropout=self.model_config['translator_dropout'],
fc_factor=self.model_config['translator_fc_factor'], )
self.encoder._build()
self.ctc_model._build()
self.translator._build()
self.load_checkpoint()
self.encoder.summary(line_length=100)
self.ctc_model.summary(line_length=100)
self.translator.summary(line_length=100)
def load_checkpoint(self, ):
"""Load checkpoint."""
self.checkpoint_dir = os.path.join(self.running_config["outdir"], "encoder-ckpt")
files = os.listdir(self.checkpoint_dir)
files.sort(key=lambda x: int(x.split('_')[-1].replace('.h5', '')))
self.encoder.load_weights(os.path.join(self.checkpoint_dir, files[-1]),by_name=True)
logging.info('encoder load at {}'.format(os.path.join(self.checkpoint_dir, files[-1])))
self.checkpoint_dir = os.path.join(self.running_config["outdir"], "ctc_decoder-ckpt")
files = os.listdir(self.checkpoint_dir)
files.sort(key=lambda x: int(x.split('_')[-1].replace('.h5', '')))
self.ctc_model.load_weights(os.path.join(self.checkpoint_dir, files[-1]))
logging.info('ctc_model load at {}'.format(os.path.join(self.checkpoint_dir, files[-1])))
self.checkpoint_dir = os.path.join(self.running_config["outdir"], "translator-ckpt")
files = os.listdir(self.checkpoint_dir)
files.sort(key=lambda x: int(x.split('_')[-1].replace('.h5', '')))
self.translator.load_weights(os.path.join(self.checkpoint_dir, files[-1]))
logging.info('translator load at {}'.format(os.path.join(self.checkpoint_dir, files[-1])))
def stream_stt(self, wav_path):
data = self.speech_featurizer.load_wav(wav_path)
enc_outputs = None
for i in range(9999):
s = i * self.chunk
e = s + self.chunk
if s >= len(data):
break
input_wav = data[int(s):int(e)]
input_wav = input_wav.reshape([1, -1, 1])
es = time.time()
enc_output = self.encoder.inference(input_wav)
ee = time.time()
enc_output = enc_output.numpy()
if enc_outputs is not None:
enc_outputs = np.hstack((enc_outputs, enc_output))
else:
enc_outputs = enc_output
# 这里为每chunk预测一次，也可以最后预测一次
ds = time.time()
ctc_output = self.ctc_model(enc_outputs, training=False)
de = time.time()
ctc_output = tf.nn.softmax(ctc_output, -1)
input_length = np.array([enc_outputs.shape[1]], 'int32')
ctc_decode = tf.keras.backend.ctc_decode(ctc_output, input_length)[0][0]
ctc_decode = tf.cast(tf.clip_by_value(ctc_decode, 0, self.phone_featurizer.num_classes), tf.int32)
ts = time.time()
ctc_result = []
for n in ctc_decode[0].numpy():
if n != 0:
ctc_result.append(n)
ctc_result += [0] * 10
translator_out = self.translator.inference(np.array([ctc_result], 'int32'), enc_outputs)
translator_out = tf.argmax(translator_out, -1)
te = time.time()
print('extract cost time:', ee - es, 'ctc decode time:', de - ds, 'translator cost time:', te - ts)
ctc_result = []
for n in ctc_decode[0].numpy():
if n != 0:
ctc_result.append(n)
txt_result = []
for n in translator_out[0].numpy():
if n != 0:
txt_result.append(n)
if n==self.text_featurizer.endid():
break
phone = self.phone_featurizer.iextract(ctc_result)
txt = self.text_featurizer.iextract(txt_result)
return ' '.join(phone), ''.join(txt)
def remove_blank(self, labels, blank=0):
new_labels = []
# 合并相同的标签
previous = None
for l in labels:
if l != previous:
new_labels.append(l)
previous = l
# 删除blank
new_labels = [l for l in new_labels if l != blank]
return new_labels
def greedy_decode(self, y, blank=1331):
# 按列取最大值，即每个时刻t上最大值对应的下标
raw_rs = np.argmax(y, axis=1)
# 移除blank,值为0的位置表示这个位置是blank
rs = self.remove_blank(raw_rs, blank)
return rs
def offline_stt(self, wav_path):
# am_result is token id
data = self.speech_featurizer.load_wav(wav_path)
input_wav = data.reshape([1, -1, 1])
es = time.time()
enc_outputs = self.encoder(input_wav, training=False)
ee = time.time()
ds = time.time()
ctc_output = self.ctc_model(enc_outputs, training=False)
de = time.time()
ctc_output = tf.nn.softmax(ctc_output, -1)
input_length = np.array([enc_outputs.shape[1]], 'int32')
ctc_decode = tf.keras.backend.ctc_decode(ctc_output, input_length)[0][0]
ctc_decode = tf.cast(tf.clip_by_value(ctc_decode, 0, self.phone_featurizer.num_classes), tf.int32)
ts = time.time()
translator_out = self.translator([ctc_decode, enc_outputs], training=False)
translator_out = tf.argmax(translator_out, -1)
te = time.time()
print('extract feature cost:', ee - es, 'ctc cost time:', de - ds, 'translator cost time:', te - ts)
ctc_result = []
for n in ctc_decode[0].numpy():
if n != 0:
ctc_result.append(n)
txt_result = []
for n in translator_out[0].numpy():
if n != 0:
txt_result.append(n)
if n==self.text_featurizer.endid():
break
phone = self.phone_featurizer.iextract(ctc_result)
txt = self.text_featurizer.iextract(txt_result)
return ' '.join(phone), ''.join(txt)
def stt(self, wav_path):
if self.speech_config['streaming']:
return self.stream_stt(wav_path)
else:
return self.offline_stt(wav_path)
def convert_to_onnx(self):
import tf2onnx
self.encoder.set_inference_func()
self.ctc_model.set_inference_func()
self.translator.set_inference_func()
tf2onnx.convert.from_function(self.encoder.inference,
input_signature=[ tf.TensorSpec([None, None,1], dtype=tf.float32)],opset=13,output_path='./encoder.onnx')
tf2onnx.convert.from_function(self.ctc_model.inference,
input_signature=[ tf.TensorSpec([None, None,self.ctc_model.dmodel], dtype=tf.float32),], opset=13,
output_path='./ctc_model.onnx')
tf2onnx.convert.from_function(self.translator.inference,
input_signature=[ tf.TensorSpec([None, None], dtype=tf.int32),
tf.TensorSpec([None, None, self.translator.dmodel], dtype=tf.float32),], opset=13,
output_path='./translator.onnx')
def convert_to_pb(self,export_path):
self.encoder.set_inference_func()
self.ctc_model.set_inference_func()
self.translator.set_inference_func()
encoder=os.path.join(export_path,'encoder')
ctc=os.path.join(export_path,'ctc_decoder')
translator=os.path.join(export_path,'translator')
concrete_func = self.encoder.inference.get_concrete_function()
tf.saved_model.save(self.encoder, encoder, signatures=concrete_func)
concrete_func = self.ctc_model.inference.get_concrete_function()
tf.saved_model.save(self.ctc_model, ctc, signatures=concrete_func)
concrete_func = self.translator.inference.get_concrete_function()
tf.saved_model.save(self.translator, translator, signatures=concrete_func)
if __name__ == '__main__':
import time
# USE CPU:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
# USE one GPU:
# os.environ['CUDA_VISIBLE_DEVICES'] = '0'
# limit cpu to 1 core:
# import tensorflow as tf
# tf.config.threading.set_inter_op_parallelism_threads(1)
# tf.config.threading.set_intra_op_parallelism_threads(1)
am_config = UserConfig(r'./asr/configs/am_data.yml', r'./asr/configs/conformerS.yml')
asr = ASR(am_config)
print(asr.stt('./asr/BAC009S0764W0121.wav'))
asr.convert_to_onnx()