-
Notifications
You must be signed in to change notification settings - Fork 4
/
07_train_vc_gan.py
202 lines (147 loc) · 6.26 KB
/
07_train_vc_gan.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
# coding: utf-8
"""Train VC model.
"""
from datetime import datetime
import glob
import os
import numpy as np
import torch
import torch.nn as nn
from torch.nn import functional as F
import torch.optim as optim
from torch.optim.lr_scheduler import StepLR
from tqdm import tqdm
import config
from data import SpeechGenDataset
from model import SpeechRecogModel, SpeechGenModel, DiscrimintorModel
device = "cuda:{}".format(config.GPU_ID) if torch.cuda.is_available() else "cpu"
def train(train_loader, recog_model, gen_model, dis_model, criterion_G, optimizer_G, optimizer_D, scheduler_G, scheduler_D):
running_loss_G = 0.0
running_loss_D = 0.0
running_loss_MSE = 0.0
for x, y in tqdm(train_loader, total=len(train_loader), desc='train'):
x, y = x.to(device), y.to(device)
# Train Discriminator
gen_model.eval()
dis_model.train()
optimizer_D.zero_grad()
output_real = dis_model(y)
x_tmp = recog_model(x)
x_tmp = F.softmax(x_tmp, dim=1)
x_tmp = gen_model(x_tmp)
output_fake = dis_model(x_tmp)
loss_D = - torch.mean(output_real) + torch.mean(output_fake)
loss_D.backward()
optimizer_D.step()
running_loss_D += loss_D.item()
# Clip weights of Discriminator
for p in dis_model.parameters():
p.data.clamp_(-0.01, 0.01)
# Train Generator
gen_model.train()
dis_model.eval()
optimizer_G.zero_grad()
x_tmp = recog_model(x)
x_tmp = F.softmax(x_tmp, dim=1)
x_tmp = gen_model(x_tmp)
output_fake = dis_model(x_tmp)
loss_G = - torch.mean(output_fake) * 5
loss_MSE = criterion_G(x_tmp, y)
loss_G_all = loss_G + loss_MSE
loss_G_all.backward()
optimizer_G.step()
running_loss_G += loss_G.item()
running_loss_MSE += loss_MSE.item()
train_loss_D = running_loss_D / len(train_loader)
train_loss_G = running_loss_G / len(train_loader)
train_loss_MSE = running_loss_MSE / len(train_loader)
# scheduler.step()
return train_loss_D, train_loss_G, train_loss_MSE
def valid(valid_loader, recog_model, gen_model, dis_model, criterion_G):
gen_model.eval()
dis_model.eval()
running_loss_G = 0.0
running_loss_D = 0.0
running_loss_MSE = 0.0
with torch.no_grad():
for x, y in tqdm(valid_loader, total=len(valid_loader), desc='valid'):
x, y = x.to(device), y.to(device)
output_real = dis_model(y)
x_tmp = recog_model(x)
x_tmp = F.softmax(x_tmp, dim=1)
x_tmp = gen_model(x_tmp)
output_fake = dis_model(x_tmp)
loss_D = - torch.mean(output_real) + torch.mean(output_fake)
running_loss_D += loss_D.item()
x_tmp = recog_model(x)
x_tmp = F.softmax(x_tmp, dim=1)
x_tmp = gen_model(x_tmp)
output_fake = dis_model(x_tmp)
loss_G = - torch.mean(output_fake) * 5
loss_MSE = criterion_G(x_tmp, y)
loss_G_all = loss_G + loss_MSE
running_loss_G += loss_G.item()
running_loss_MSE += loss_MSE.item()
valid_loss_G = running_loss_G / len(valid_loader)
valid_loss_D = running_loss_D / len(valid_loader)
valid_loss_MSE = running_loss_MSE / len(valid_loader)
return valid_loss_D, valid_loss_G, valid_loss_MSE
def main():
ct = datetime.now()
log_dir = os.path.join(config.LOG_ROOT, "VC", ct.strftime("%Y%m%d_%H%M%S"))
if not os.path.exists(log_dir):
os.makedirs(log_dir)
else:
raise FileExistsError("Directory {} already exists".format(log_dir))
print("Output dir: {}".format(log_dir))
train_loader = torch.utils.data.DataLoader(
SpeechGenDataset(phase = "train", data = "train", spk = config.VC_SPEAKER),
batch_size = config.VC_BATCH_SIZE,
num_workers = config.NUM_WORKERS,
shuffle = True
)
valid_loader = torch.utils.data.DataLoader(
SpeechGenDataset(phase = "train", data = "valid", spk = config.VC_SPEAKER),
batch_size = config.VC_BATCH_SIZE,
num_workers = config.NUM_WORKERS,
shuffle = True
)
recog_model_path = glob.glob(os.path.join(config.ASR_INFERENCE_ROOT, "*best.pth"))[0]
recog_model = SpeechRecogModel().to(device)
recog_model.load_state_dict(torch.load(recog_model_path, map_location='cuda:0'))
recog_model.eval()
gen_model = SpeechGenModel().to(device)
dis_model = DiscrimintorModel().to(device)
print(recog_model)
print(gen_model)
print(dis_model)
optimizer_G = optim.Adam(
gen_model.parameters(),
lr = config.VC_LR
)
optimizer_D = optim.Adam(
dis_model.parameters(),
lr = config.VC_LR
)
scheduler_G = StepLR(optimizer_G, step_size=100, gamma=0.1)
scheduler_D = StepLR(optimizer_D, step_size=100, gamma=0.1)
criterion_G = nn.MSELoss()
best_loss = None
best_model = None
for epoch in range(1, config.VC_N_EPOCHS + 1):
train_loss_D, train_loss_G, train_loss_MSE = train(train_loader, recog_model, gen_model, dis_model, criterion_G, optimizer_G, optimizer_D, scheduler_G, scheduler_D)
valid_loss_D, valid_loss_G, valid_loss_MSE = valid(valid_loader, recog_model, gen_model, dis_model, criterion_G)
print("epoch [{}/{}]\ntrain_loss_D: {:.3f}, train_loss_G: {:.3f}, train_loss_MSE: {:.3f},\nvalid_loss_D: {:.3f}, valid_loss_G: {:.3f}, valid_loss_MSE: {:.3f}"\
.format(epoch, config.VC_N_EPOCHS, train_loss_D, train_loss_G, train_loss_MSE, valid_loss_D, valid_loss_G, valid_loss_MSE))
# if epoch == 1 or (valid_loss < best_loss):
# if best_loss is not None:
# print(' => valid_loss improved from {:.3f} to {:.3f}!'.format(best_loss, valid_loss))
# os.remove(best_model)
# best_loss = valid_loss
# best_model = os.path.join(log_dir, 'epoch{:03d}_{:.3f}_best.pth'.format(epoch, valid_loss))
# torch.save(gen_model.state_dict(), best_model)
if epoch % config.SAVE_MODEL_FREQ == 0:
current_model = os.path.join(log_dir, 'epoch{:03d}_{:.3f}.pth'.format(epoch, valid_loss_G))
torch.save(gen_model.state_dict(), current_model)
if __name__ == "__main__":
main()