-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_model.py
201 lines (174 loc) · 7.52 KB
/
train_model.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
from resemblyzer import preprocess_wav, VoiceEncoder
import os,librosa
import numpy as np
import tensorflow as tf
from tensorflow.keras import Model
from tensorflow.keras.layers import Multiply, ZeroPadding2D, concatenate, Conv2D, Input, Dense, Reshape, BatchNormalization, Activation, LSTM, Lambda, Bidirectional
from tensorflow.keras.utils import Sequence
from tensorflow.keras.callbacks import Callback
from tensorflow.keras.callbacks import ModelCheckpoint
from Audio import Audio,HyperParams
import soundfile as sf
import voice_demo
def get_model():
hyper_params = HyperParams() # model parameters
T_dim = int(hyper_params.length/hyper_params.hop_length)
emb_dim = hyper_params.embedder_emb_dim
num_freq = hyper_params.num_freq
lstm_dim = hyper_params.model_lstm_dim
fc1_dim = hyper_params.model_fc1_dim
fc2_dim = hyper_params.model_fc2_dim # num_freq
dvec_inp = Input(shape=(emb_dim),name="dvec")
input_spec = Input(shape=(T_dim,num_freq),name="input_spec")
x = Reshape((T_dim,num_freq,1))(input_spec)
#cnn1
x = ZeroPadding2D(((0,0), (3,3)))(x)
x = Conv2D(filters=64, kernel_size=[1,7], dilation_rate=[1, 1])(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
#cnn2
x = ZeroPadding2D(((3,3), (0,0)))(x)
x = Conv2D(filters=64, kernel_size=[7,1], dilation_rate=[1, 1])(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
#cnn3
x = ZeroPadding2D(((2,2), (2,2)))(x)
x = Conv2D(filters=64, kernel_size=[5,5], dilation_rate=[1, 1])(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
#cnn4
x = ZeroPadding2D(((4,4), (2,2)))(x)
x = Conv2D(filters=64, kernel_size=[5,5], dilation_rate=[2, 1])(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
#cnn5
x = ZeroPadding2D(((8,8), (2,2)))(x)
x = Conv2D(filters=64, kernel_size=[5,5], dilation_rate=[4, 1])(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
#cnn6
x = ZeroPadding2D(((16,16), (2,2)))(x)
x = Conv2D(filters=64, kernel_size=[5,5], dilation_rate=[8, 1])(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
#cnn7
x = ZeroPadding2D(((32,32), (2,2)))(x)
x = Conv2D(filters=64, kernel_size=[5,5], dilation_rate=[16, 1])(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
#cnn8
x = Conv2D(filters=8, kernel_size=[1,1], dilation_rate=[1, 1])(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Reshape((x.shape[1],x.shape[2]*x.shape[3]))(x) #else use -1 as last arg
#x = tf.reshape(x, [x.shape[0],x.shape[1],-1])
dvec = Lambda(lambda a : tf.expand_dims(a,1))(dvec_inp)
dvec = Lambda(lambda a : tf.repeat(a,repeats =x.shape[1],axis =1))(dvec)
#dvec= tf.expand_dims(dvec_inp,1)
#dvec= tf.repeat(dvec,repeats =x.shape[1],axis =1)
x = concatenate([x,dvec],-1)
#x= tf.concat([x,dvec],-1)
#lstm
x = Bidirectional(LSTM(lstm_dim,return_sequences=True))(x)
#fc1
x = Dense(fc1_dim,activation ="relu")(x)
#fc2
mask = Dense(fc2_dim,activation ="sigmoid",name="mask")(x) #soft mask
#element-wise
output = Multiply()([input_spec,mask])
model = Model(inputs=[input_spec,dvec_inp], outputs=output)
model.summary()
return model
def generate_data(data_path, audio, embed_average):
ids=np.random.randint(1,21,size=3)
embeddings = []
wavs = []
spec = []
# encoder = VoiceEncoder()
length = 80000
for i in [0,1,2]:
path = data_path + "ID" + str(ids[i]) + "/"
files = os.listdir(path)
np.random.shuffle(files)
for file in files:
wav,_ = librosa.load(path+file, sr=16000)
wav = (0.8 * np.random.rand() + 0.2) * wav
if len(wav) >= length:
break
wavs.append(wav[0:length - 1])
# embeddings.append(encoder.embed_utterance(wavs[i]))
embeddings.append(embed_average[ids[i]-1])
wav_spec,_ = audio.wave2spec(wavs[i])
spec.append(wav_spec)
mixed_wav = wavs[0]
mixed_wav += wavs[1]
mixed_wav += wavs[2]
mixed_spec,mixed_phase = audio.wave2spec(mixed_wav)
return embeddings, spec, mixed_spec, mixed_phase
def create_dataset(data_path, audio, train_num=2):
"""生成数据集"""
print("\nStart Creating Dataset\n")
_,embed_average = voice_demo.read_voices("./train/")
input_spec = []
input_embeddings = []
output_spec = []
for n in range(0,train_num,1):
embeddings, spec, mixed_spec, mixed_phase = generate_data(data_path, audio, embed_average)
# for i in range(0,3,1):
input_embeddings.append(embeddings[0])
input_spec.append(mixed_spec)
output_spec.append(spec[0])
input_spec = tf.convert_to_tensor(input_spec)
input_embeddings = tf.convert_to_tensor(input_embeddings)
output_spec = tf.convert_to_tensor(output_spec)
print("\nCreate Dataset Done\n")
return input_spec,input_embeddings,output_spec
# class Dataset(Sequence):
# def __init__(self, train_num, data_path):
# self.train_num = train_num
# self.datapath = data_path
# self.input_spec,self.input_embedding,self.output_spec = create_dataset(data_path)
# def __len__(self):
# return self.datapath
# def __getitem__(self, idx):
# return ({'input_spec':self.input_spec, 'dvec': self.input_embedding}, self.output_spec)
def train_model(epochs, size, weights_path):
hyper_params = HyperParams()
audio = Audio(hyper_params)
encoder = VoiceEncoder()
# 保存训练好的weight的路径
if not os.path.isdir(weights_path):
os.mkdir(weights_path)
model_checkpoint_callback = ModelCheckpoint(os.path.join(weights_path,'weights_epoch{epoch:04d}.h5'),save_weights_only=True,save_freq='epoch')
input_spec, input_embedding, output_spec = create_dataset('./train/', audio, size)
model = get_model()
model.compile(optimizer='adam', loss='mse')
model.load_weights(os.path.join(weights_path,'weights_epoch0001.h5'))
history = model.fit({'input_spec':input_spec, 'dvec':input_embedding}, output_spec,
epochs=epochs, shuffle=True,
callbacks=[model_checkpoint_callback])
return model
if __name__ == "__main__":
with tf.device('/gpu:0'):
# 训练参数
epochs = 2 # 估计loss在较小的epoch时就已经稳定了，为了节省时间就少训练几个epoch
size = 400 # 在这里改训练数据量（先从500开始，之后1000 2000 5000 10000 20000）
weights_path = './model/'
train_model(epochs, size, weights_path)
# model = get_model()
# model.compile(optimizer='adam', loss='mse')
# model.load_weights(os.path.join(weights_path,'weights_epoch%04d.h5'%epochs))
# hyper_params = HyperParams()
# audio = Audio(hyper_params)
# encoder = VoiceEncoder()
# wav,_ = librosa.load('./test_offline/task3/combine001.mp4', sr=16000)
# input_spec,mixed_phase = audio.wave2spec(wav)
# input_spec = tf.convert_to_tensor([input_spec])
# wav_7,_ = librosa.load('./test_offline/task3_gt/001_left.wav', sr=16000)
# # wav_7 = preprocess_wav('./test_offline/task3_gt/001_left.wav', 16000)
# embed = tf.convert_to_tensor([encoder.embed_utterance(wav_7)])
# print(input_spec.shape)
# print(embed.shape)
# out_spec = model.predict(x={'input_spec':input_spec, 'dvec':embed},verbose=1)
# out_wav = audio.spec2wave(out_spec[0],mixed_phase)
# sf.write('./predict7.wav', out_wav, 16000)