forked from MingjieChen/DYGANVC
-
Notifications
You must be signed in to change notification settings - Fork 0
/
solver.py
executable file
·310 lines (240 loc) · 12.5 KB
/
solver.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
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
from generator import Generator
from discriminator import Discriminator
from speaker_encoder import SPEncoder
import torch
import torch.nn.functional as F
import os
from os.path import join, basename, exists
import time
import datetime
import numpy as np
from tqdm import tqdm
import numpy as np
import copy
class Solver(object):
def __init__(self, train_loader, config):
"""Initialize configurations."""
self.train_loader = train_loader
self.sampling_rate = config.sampling_rate
self.D_name = config.discriminator
self.SPE_name = config.spenc
self.G_name = config.generator
self.g_hidden_size = config.g_hidden_size
self.num_speakers = config.num_speakers
self.spk_emb_dim = config.spk_emb_dim
self.lambda_rec = config.lambda_rec
self.lambda_id = config.lambda_id
self.lambda_adv = config.lambda_adv
self.batch_size = config.batch_size
self.num_iters = config.num_iters
self.g_lr = config.g_lr
self.d_lr = config.d_lr
self.beta1 = config.beta1
self.beta2 = config.beta2
self.resume_iters = config.resume_iters
self.use_ema = config.use_ema
self.auto_resume = config.auto_resume
self.kernel = config.kernel
self.num_heads = config.num_heads
self.num_res_blocks = config.num_res_blocks
self.use_tensorboard = config.use_tensorboard
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.log_dir = config.log_dir
self.model_save_dir = config.model_save_dir
self.log_step = config.log_step
self.sample_step = config.sample_step
self.model_save_step = config.model_save_step
self.build_model()
if self.use_tensorboard:
self.build_tensorboard()
def build_model(self):
"""Create a generator and a discriminator."""
self.generator = eval(self.G_name)(num_speakers=self.num_speakers,
kernel = self.kernel,
num_heads = self.num_heads,
num_res_blocks = self.num_res_blocks,
spk_emb_dim = self.spk_emb_dim,
)
self.discriminator = eval(self.D_name)(num_speakers=self.num_speakers)
self.sp_enc = eval(self.SPE_name)(num_speakers = self.num_speakers, spk_emb_dim = self.spk_emb_dim)
self.sp_enc.to(self.device)
self.generator.to(self.device)
self.discriminator.to(self.device)
g_params = list(self.generator.parameters())
g_params += list(self.sp_enc.parameters())
d_params = list(self.discriminator.parameters())
self.g_optimizer = torch.optim.Adam(g_params, self.g_lr, [self.beta1, self.beta2])
self.d_optimizer = torch.optim.Adam(d_params, self.d_lr, [self.beta1, self.beta2])
# restore model
if not self.auto_resume:
if self.resume_iters and not self.resume_ft:
print("resuming step %d ..."% self.resume_iters, flush=True)
self.restore_model(self.resume_iters)
else:
ckpt_files = [ int(x.split('-')[0]) for x in os.listdir(self.model_save_dir)]
last_step = sorted(ckpt_files, reverse = True)[0]
print("auto resuming step %d ..."% last_step, flush=True)
self.restore_model(last_step)
self.resume_iters = last_step
if self.use_ema:
self.generator_ema = copy.deepcopy(self.generator)
self.sp_enc_ema = copy.deepcopy(self.sp_enc)
self.print_network(self.generator, 'Generator')
self.print_network(self.discriminator, 'Discriminator')
self.print_network(self.sp_enc, 'SpeakerEncoder')
if self.use_ema:
self.generator_ema.to(self.device)
self.sp_enc_ema.to(self.device)
def print_network(self, model, name):
"""Print out the network information."""
num_params = 0
for p in model.parameters():
num_params += p.numel()
print(model, flush=True)
print(name,flush=True)
print("The number of parameters: {}".format(num_params), flush=True)
def moving_average(self, model, model_test, beta = 0.999):
for param, param_test in zip(model.parameters(), model_test.parameters()):
param_test.data = torch.lerp(param.data, param_test.data, beta)
def restore_model(self, resume_iters, resume_ft = False):
"""Restore the trained generator and discriminator."""
print('Loading the trained models from step {}...'.format(resume_iters), flush=True)
g_path = os.path.join(self.model_save_dir, '{}-G.ckpt'.format(resume_iters))
d_path = os.path.join(self.model_save_dir, '{}-D.ckpt'.format(resume_iters))
sp_path = os.path.join(self.model_save_dir, '{}-sp.ckpt'.format(resume_iters))
g_opt_path = os.path.join(self.model_save_dir, '{}-g_opt.ckpt'.format(resume_iters))
d_opt_path = os.path.join(self.model_save_dir, '{}-d_opt.ckpt'.format(resume_iters))
self.generator.load_state_dict(torch.load(g_path, map_location=lambda storage, loc: storage))
self.discriminator.load_state_dict(torch.load(d_path, map_location=lambda storage, loc: storage))
self.sp_enc.load_state_dict(torch.load(sp_path, map_location=lambda storage, loc: storage))
print("loading optimizer",flush=True)
if exists(g_opt_path):
self.g_optimizer.load_state_dict(torch.load(g_opt_path, map_location = lambda storage, loc: storage))
if exists(d_opt_path):
self.d_optimizer.load_state_dict(torch.load(d_opt_path, map_location = lambda storage, loc: storage))
def build_tensorboard(self):
"""Build a tensorboard logger."""
from logger import Logger
self.logger = Logger(self.log_dir)
def update_lr(self, g_lr, d_lr):
"""Decay learning rates of the generator and discriminator."""
for param_group in self.g_optimizer.param_groups:
param_group['lr'] = g_lr
for param_group in self.d_optimizer.param_groups:
param_group['lr'] = d_lr
def reset_grad(self):
"""Reset the gradientgradient buffers."""
self.g_optimizer.zero_grad()
self.d_optimizer.zero_grad()
def label2onehot(self, labels, dim):
"""Convert label indices to one-hot vectors."""
batch_size = labels.size(0)
out = torch.zeros(batch_size, dim)
out[np.arange(batch_size), labels.long()] = 1
return out
def sample_spk_c(self, size):
spk_c = np.random.randint(0, self.num_speakers, size=size)
spk_c_cat = to_categorical(spk_c, self.num_speakers)
return torch.LongTensor(spk_c), torch.FloatTensor(spk_c_cat)
def classification_loss(self, logit, target):
"""Compute softmax cross entropy loss."""
return F.cross_entropy(logit, target)
def load_wav(self, wavfile, sr=16000):
wav, _ = librosa.load(wavfile, sr=sr, mono=True)
return wav_padding(wav, sr=16000, frame_period=5, multiple = 4)
def load_mel(self, melfile):
tmp_mel = np.load(melfile)
return tmp_mel
def train(self):
# Set data loader.
train_loader = self.train_loader
data_iter = iter(train_loader)
g_lr = self.g_lr
d_lr = self.d_lr
start_iters = 0
if self.resume_iters:
start_iters = self.resume_iters
print('Start training...', flush=True)
start_time = time.time()
for i in range(start_iters, self.num_iters):
try:
mc_src, spk_label_org, spk_c_org, mc_trg, spk_label_trg, spk_c_trg = next(data_iter)
except:
data_iter = iter(train_loader)
mc_src, spk_label_org, spk_c_org, mc_trg, spk_label_trg, spk_c_trg = next(data_iter)
mc_src.unsqueeze_(1)
mc_trg.unsqueeze_(1)
mc_src = mc_src.to(self.device)
mc_trg = mc_trg.to(self.device)
spk_label_org = spk_label_org.to(self.device)
spk_c_org = spk_c_org.to(self.device)
spk_label_trg = spk_label_trg.to(self.device)
spk_c_trg = spk_c_trg.to(self.device)
spk_c_trg = self.sp_enc(mc_trg, spk_label_trg)
spk_c_org = self.sp_enc(mc_src, spk_label_org)
d_out_src = self.discriminator(mc_src, spk_label_trg, spk_label_org)
d_loss_real = torch.mean( (1.0 - d_out_src)**2 )
mc_fake = self.generator(mc_src, spk_c_org, spk_c_trg)
d_out_fake = self.discriminator(mc_fake.detach(), spk_label_org, spk_label_trg)
d_loss_fake = torch.mean(d_out_fake ** 2)
# Backward and optimize.
d_loss = d_loss_real + d_loss_fake
self.reset_grad()
d_loss.backward()
self.d_optimizer.step()
# Logging.
loss = {}
loss['D/loss_real'] = d_loss_real.item()
loss['D/loss_fake'] = d_loss_fake.item()
loss['D/loss'] = d_loss.item()
spk_c_trg = self.sp_enc(mc_trg, spk_label_trg)
spk_c_org = self.sp_enc(mc_src, spk_label_org)
mc_fake = self.generator(mc_src, spk_c_org, spk_c_trg)
g_out_src = self.discriminator(mc_fake, spk_label_org, spk_label_trg)
g_loss_fake = torch.mean((1.0 - g_out_src)**2)
mc_reconst = self.generator(mc_fake, spk_c_trg, spk_c_org)
g_loss_rec = torch.mean(torch.abs(mc_src - mc_reconst))
mc_fake_id = self.generator(mc_src, spk_c_org, spk_c_org)
g_loss_id = torch.mean(torch.abs(mc_src - mc_fake_id))
# Backward and optimize.
g_loss = self.lambda_adv * g_loss_fake \
+ self.lambda_rec * g_loss_rec \
+ self.lambda_id * g_loss_id
self.reset_grad()
g_loss.backward()
self.g_optimizer.step()
# Logging.
loss['G/loss_fake'] = g_loss_fake.item()
loss['G/loss_rec'] = g_loss_rec.item()
loss['G/loss_id'] = g_loss_id.item()
if self.use_ema:
self.moving_average(self.generator, self.generator_ema)
self.moving_average(self.sp_enc, self.sp_enc_ema)
if (i+1) % self.log_step == 0:
et = time.time() - start_time
et = str(datetime.timedelta(seconds=et))[:-7]
log = "Elapsed [{}], Iteration [{}/{}]".format(et, i+1, self.num_iters)
for tag, value in loss.items():
log += ", {}: {:.4f}".format(tag, value)
print(log, flush=True)
if self.use_tensorboard:
for tag, value in loss.items():
self.logger.scalar_summary(tag, value, i+1)
if (i+1) % self.model_save_step == 0:
g_path = os.path.join(self.model_save_dir, '{}-G.ckpt'.format(i+1))
g_path_ema = os.path.join(self.model_save_dir, '{}-G.ckpt.ema'.format(i+1))
d_path = os.path.join(self.model_save_dir, '{}-D.ckpt'.format(i+1))
sp_path = os.path.join(self.model_save_dir, '{}-sp.ckpt'.format(i+1))
sp_path_ema = os.path.join(self.model_save_dir, '{}-sp.ckpt.ema'.format(i+1))
g_opt_path = os.path.join(self.model_save_dir, '{}-g_opt.ckpt'.format(i+1))
d_opt_path = os.path.join(self.model_save_dir, '{}-d_opt.ckpt'.format(i+1))
torch.save(self.generator.state_dict(), g_path)
if self.use_ema:
torch.save(self.generator_ema.state_dict(), g_path_ema)
torch.save(self.discriminator.state_dict(), d_path)
torch.save(self.sp_enc.state_dict(), sp_path)
if self.use_ema:
torch.save(self.sp_enc_ema.state_dict(), sp_path_ema)
torch.save(self.g_optimizer.state_dict(), g_opt_path)
torch.save(self.d_optimizer.state_dict(), d_opt_path)
print('Saved model checkpoints into {}...'.format(self.model_save_dir), flush=True)