-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathbuild_wavenet.py
206 lines (163 loc) · 7.74 KB
/
build_wavenet.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
import tensorflow as tf
import keras
import argparse
from keras_wavenet.utils.audio_generator_utils import WavGenerator
import numpy as np
from keras.optimizers import Adam
from keras_wavenet.weightnorm import AdamWithWeightnorm
import keras.backend as K
import pickle
from keras.layers import (Lambda,Reshape,Conv1D,Add,Activation,
Concatenate,Dense,RepeatVector,GRU,Bidirectional,
Multiply,MaxPool1D,Flatten,GlobalMaxPool1D,TimeDistributed,
Dropout,ZeroPadding1D,GaussianDropout)
from keras_wavenet.models.wavenet import build_wavenet_decoder,build_wavenet_encoder
from keras_wavenet.models.audio_outputs import get_output_processor
import sys
import os
import json
from inspect import signature,Parameter
fs = os.path.sep
def get_default_args(func):
sig = signature(func)
return {k: v.default
for k, v in sig.parameters.items()
if v.default is not Parameter.empty }
def build_model(input_shape,dec_width=512,dec_skip_width=256,
enc_width=512,
num_en_layers=8,num_en_stages=2,enc_pool_size=None,
num_dec_layers=30,num_dec_stages=10,
latent_size=8,filter_len=3,epsilon_std=1.0,
final_conditioning=True,
final_activation='softmax',
output_channels=257,
preprocess_func_str = "lambda x : x/128.",
stochastic_encoding=False,
output_processor='sparse_categorical',
output_processor_kwargs = None,
):
num_steps,input_channels = input_shape
input_layer = keras.layers.Input(shape = input_shape)
preprocess_func = eval(preprocess_func_str)
scaled_input = Lambda(preprocess_func,output_shape=input_shape,
name='preprocessor')(input_layer)
#Encoder
encoding = build_wavenet_encoder(scaled_input,width=enc_width,filter_len=filter_len,
num_layers=num_en_layers,num_stages=num_en_stages
)
encoding = Activation('relu')(encoding)
if enc_pool_size is None:
en = GlobalMaxPool1D()(encoding)
en = Reshape((1,enc_width))(en)
else:
assert input_shape[0] % enc_pool_size == 0
en = MaxPool1D(pool_size=enc_pool_size)(encoding)
#Latent Sampling
z_mean = Conv1D(latent_size,kernel_size=1,name='z_mean')(en)
if stochastic_encoding:
z_log_var = Conv1D(latent_size,kernel_size=1,
kernel_initializer=keras.initializers.constant(-0.1)
)(en)
z_log_var = Lambda(lambda x: K.minimum(x,6.0),name='z_log_var',)(z_log_var)
_,enc_timesteps,_ = z_mean._keras_shape
def sampling(args):
z_mean, z_log_var = args
epsilon = K.random_normal(shape=(K.shape(z_mean)[0], enc_timesteps, latent_size),
mean=0., stddev=epsilon_std)
return z_mean + K.exp(z_log_var/2) * epsilon
z = Lambda(sampling, output_shape=(enc_timesteps,latent_size,),
name='decoder_input')([z_mean, z_log_var])
else:
z = Lambda(lambda x: x,name='decoder_input')(z_mean)
#Decoder
decoder_out = build_wavenet_decoder(scaled_input,z,
width=dec_width,skip_width=dec_skip_width,
out_width=output_channels,
num_layers=num_dec_layers,num_stages=num_dec_stages,
final_conditioning=final_conditioning,
final_activation=final_activation
)
model = keras.models.Model(input_layer, decoder_out)
#build loss
px_loss_func = get_output_processor(output_processor,
output_channels,
output_processor_kwargs).loss
def vae_loss(y_true, model_output):
px_loss = px_loss_func(y_true,model_output)
if stochastic_encoding:
kl_loss = - 0.5 * K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=(1,2))
return K.mean(px_loss + kl_loss )
return px_loss
return model,vae_loss
parser = argparse.ArgumentParser()
parser.add_argument('--train_folder', dest='train_folder',
action='store', required=True,
help='train folder to load')
parser.add_argument('--valid_folder', dest='valid_folder',
action='store', required=True,
help='validation folder to load')
parser.add_argument('--save_path', dest='save_path',
action='store', required=True,
help='file name to save the model as')
parser.add_argument('--config_json', dest='config_json',
action='store', default=None,
help='Path to the config json')
args = parser.parse_args()
'''Defaults
'''
generator_dict = get_default_args(WavGenerator)
model_dict = get_default_args(build_model)
train_dict = {'batch_size':8,'patience':10,'epochs':250}
if args.config_json is not None:
config_json = json.load(open(args.config_json,'r'))
generator_dict.update(config_json['generator_dict'])
model_dict.update(config_json['model_dict'])
train_dict.update(config_json['train_dict'])
all_dict = {'generator_dict':generator_dict,'model_dict':model_dict}
json.dump(all_dict,open(args.save_path+'_options.json','w'),indent=4)
train_generator = WavGenerator(**generator_dict
)
test_generator = WavGenerator(**generator_dict,
)
test_generator.random_transforms = False
pickle.dump(test_generator,open(args.save_path+'_generator.pkl','wb'))
print('loading data')
train_gen = train_generator.flow_from_directory(args.train_folder,
shuffle=True,
follow_links=True,
batch_size=train_dict['batch_size'])
valid_gen = test_generator.flow_from_directory(args.valid_folder,
shuffle=True,
follow_links=True,
batch_size=train_dict['batch_size'])
test_x,test_y,filenames = train_gen.next(return_filenames=True)
#sys.exit()
train_gen.reset()
input_shape = np.shape(test_x)[1:]
num_train_samples = train_gen.samples
num_train_steps_per_epoch = np.ceil(num_train_samples/train_dict['batch_size'])
num_valid_samples = valid_gen.samples
num_valid_steps_per_epoch = np.ceil(num_valid_samples/train_dict['batch_size'])
#Model set up
print('model setup')
model,vae_loss = build_model(input_shape,
**model_dict
)
model.compile(optimizer=AdamWithWeightnorm(),
loss=vae_loss)
#sys.exit()
early_stop=keras.callbacks.EarlyStopping(monitor='val_loss',
patience=train_dict['patience'],
verbose=0, mode='auto')
csv_log = keras.callbacks.CSVLogger(args.save_path+'.csv')
model_checkpoint = keras.callbacks.ModelCheckpoint(args.save_path+'.hdf5',
save_best_only=True)
lr_plateu = keras.callbacks.ReduceLROnPlateau(factor=0.2,patience=7)
model.fit_generator(train_gen,
steps_per_epoch=num_train_steps_per_epoch,
validation_data=valid_gen,
validation_steps=num_valid_steps_per_epoch,
epochs=train_dict['epochs'],
verbose=1,
callbacks=[early_stop,model_checkpoint,csv_log,lr_plateu],
)