-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathtrainer.py
211 lines (166 loc) · 7.81 KB
/
trainer.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
from networks import Generator, Discriminator
import torch
import torch.optim as optim
import torch.nn.functional as F
import tf_recorder as tensorboard
from tqdm import tqdm
from dataloader import Dataloader
from torch.autograd import grad
import amp_support as amp
import random
def requires_grad(model, flag=True):
for p in model.parameters():
p.requires_grad = flag
class DummyDataParallel(torch.nn.Module):
def __init__(self, module):
super(DummyDataParallel, self).__init__()
self.module = module
def forward(self, *args, **kwargs):
return self.module(*args, **kwargs)
def DataParallel(module):
if torch.cuda.device_count() > 1:
return torch.nn.DataParallel(module)
else:
return DummyDataParallel(module) # For a consistent model structure
def cuda(module):
if torch.cuda.device_count() > 0:
return module.cuda()
else:
print('Warning: cuda cannot be activated.')
return module
class Trainer:
def __init__(self, dataset_dir, log_dir, generator_channels, discriminator_channels, nz, style_depth, lrs, betas, eps,
phase_iter, batch_size, n_cpu, opt_level):
self.nz = nz
self.dataloader = Dataloader(dataset_dir, batch_size, phase_iter * 2, n_cpu)
self.generator = cuda(DataParallel(Generator(generator_channels, nz, style_depth)))
self.discriminator = cuda(DataParallel(Discriminator(discriminator_channels)))
self.tb = tensorboard.tf_recorder('StyleGAN', log_dir)
self.phase_iter = phase_iter
self.lrs = lrs
self.betas = betas
self.opt_level = opt_level
def generator_trainloop(self, batch_size, alpha):
requires_grad(self.generator, True)
requires_grad(self.discriminator, False)
# mixing regularization
if random.random() < 0.9:
z = [torch.randn(batch_size, self.nz).cuda(),
torch.randn(batch_size, self.nz).cuda()]
else:
z = torch.randn(batch_size, self.nz).cuda()
fake = self.generator(z, alpha=alpha)
d_fake = self.discriminator(fake, alpha=alpha)
loss = F.softplus(-d_fake).mean()
self.optimizer_g.zero_grad()
with amp.scale_loss(loss, self.optimizer_g, loss_id=0) as scaled_loss:
scaled_loss.backward()
self.optimizer_g.step()
return loss.item()
def discriminator_trainloop(self, real, alpha):
requires_grad(self.generator, False)
requires_grad(self.discriminator, True)
real.requires_grad = True
self.optimizer_d.zero_grad()
d_real = self.discriminator(real, alpha=alpha)
loss_real = F.softplus(-d_real).mean()
with amp.scale_loss(loss_real, self.optimizer_d, loss_id=1) as scaled_loss_real:
scaled_loss_real.backward(retain_graph=True)
grad_real = grad(
outputs=d_real.sum(), inputs=real, create_graph=True
)[0]
grad_penalty = (
grad_real.view(grad_real.size(0), -1).norm(2, dim=1) ** 2
).mean()
grad_penalty = 10 / 2 * grad_penalty
with amp.scale_loss(grad_penalty, self.optimizer_d, loss_id=1) as scaled_grad_penalty:
scaled_grad_penalty.backward()
if random.random() < 0.9:
z = [torch.randn(real.size(0), self.nz).cuda(),
torch.randn(real.size(0), self.nz).cuda()]
else:
z = torch.randn(real.size(0), self.nz).cuda()
fake = self.generator(z, alpha=alpha)
d_fake = self.discriminator(fake, alpha=alpha)
loss_fake = F.softplus(d_fake).mean()
with amp.scale_loss(loss_fake, self.optimizer_d) as scaled_loss_fake:
scaled_loss_fake.backward()
loss = scaled_loss_real + scaled_loss_fake + scaled_grad_penalty
self.optimizer_d.step()
return loss.item(), (d_real.mean().item(), d_fake.mean().item())
def run(self, log_iter, checkpoint):
global_iter = 0
test_z = torch.randn(4, self.nz).cuda()
if checkpoint:
self.load_checkpoint(checkpoint)
else:
self.grow()
while True:
print('train {}X{} images...'.format(self.dataloader.img_size, self.dataloader.img_size))
for iter, ((data, _), n_trained_samples) in enumerate(tqdm(self.dataloader), 1):
real = data.cuda()
alpha = min(1, n_trained_samples / self.phase_iter) if self.dataloader.img_size > 8 else 1
loss_d, (real_score, fake_score) = self.discriminator_trainloop(real, alpha)
loss_g = self.generator_trainloop(real.size(0), alpha)
if global_iter % log_iter == 0:
self.log(loss_d, loss_g, real_score, fake_score, test_z, alpha)
# save 3 times during training
if iter % (len(self.dataloader) // 4 + 1) == 0:
self.save_checkpoint(n_trained_samples)
global_iter += 1
self.tb.iter(data.size(0))
self.save_checkpoint()
self.grow()
def log(self, loss_d, loss_g, real_score, fake_score, test_z, alpha):
with torch.no_grad():
fake = self.generator.module(test_z, alpha=alpha)
fake = (fake + 1) * 0.5
fake = torch.clamp(fake, min=0.0, max=1.0)
self.tb.add_scalar('loss_d', loss_d)
self.tb.add_scalar('loss_g', loss_g)
self.tb.add_scalar('real_score', real_score)
self.tb.add_scalar('fake_score', fake_score)
self.tb.add_images('fake', fake)
def grow(self):
self.discriminator = cuda(DataParallel(self.discriminator.module.grow()))
self.generator = cuda(DataParallel(self.generator.module.grow()))
self.dataloader.grow()
self.tb.renew('{}x{}'.format(self.dataloader.img_size, self.dataloader.img_size))
self.lr = self.lrs.get(str(self.dataloader.img_size), 0.001)
self.style_lr = self.lr * 0.01
self.optimizer_d = optim.Adam(params=self.discriminator.parameters(), lr=self.lr, betas=self.betas)
self.optimizer_g = optim.Adam([
{'params': self.generator.module.model.parameters(), 'lr':self.lr},
{'params': self.generator.module.style_mapper.parameters(), 'lr': self.style_lr},
],
betas=self.betas
)
[self.generator, self.discriminator], [self.optimizer_g, self.optimizer_d] = amp.initialize(
[self.generator, self.discriminator],
[self.optimizer_g, self.optimizer_d],
opt_level=self.opt_level,
num_losses=2,
)
def save_checkpoint(self, tick='last'):
torch.save({
'generator': self.generator.state_dict(),
'discriminator': self.discriminator.state_dict(),
'generator_optimizer': self.optimizer_g.state_dict(),
'discriminator_optimizer': self.optimizer_d.state_dict(),
'img_size': self.dataloader.img_size,
'tick': tick,
}, 'checkpoints/{}x{}_{}.pth'.format(self.dataloader.img_size, self.dataloader.img_size, tick))
def load_checkpoint(self, filename):
checkpoint = torch.load(filename)
print('load {}x{} checkpoint'.format(checkpoint['img_size'], checkpoint['img_size']))
while self.dataloader.img_size < checkpoint['img_size']:
self.grow()
self.generator.load_state_dict(checkpoint['generator'])
self.discriminator.load_state_dict(checkpoint['discriminator'])
self.optimizer_g.load_state_dict(checkpoint['generator_optimizer'])
self.optimizer_d.load_state_dict(checkpoint['discriminator_optimizer'])
if checkpoint['tick'] == 'last':
self.grow()
else:
self.dataloader.set_checkpoint(checkpoint['tick'])
self.tb.iter(checkpoint['tick'])