-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
229 lines (205 loc) · 8.16 KB
/
train.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
import jittor as jt
import numpy as np
from model import StyledGenerator, Discriminator
from dataset import SymbolDataset
from tqdm import tqdm
import argparse
import math
import random
jt.flags.use_cuda = True
jt.flags.log_silent = True
def requires_grad(model, flag=True):
for p in model.parameters():
p.requires_grad = flag
def accumulate(model1, model2, decay=0.999):
par1 = dict(model1.named_parameters())
par2 = dict(model2.named_parameters())
for k in par1.keys():
par1[k].update(par1[k] * decay + (1 - decay) * par2[k].detach())
def adjust_lr(optimizer, lr):
for group in optimizer.param_groups:
mult = group.get('mult', 1)
group['lr'] = lr * mult
if __name__ == '__main__':
path="/home/user/Desktop/stylegan/data/symbol"
ckpt=None
code_size=512
batch_size={4: 512, 8: 256, 16: 128, 32: 64, 64: 32, 128: 16}
lr = 1e-3
init_size=8
max_size=128
use_loss='wgan-gp'
from_rgb_act=True
mixing = True
sched=False
generator=StyledGenerator(code_dim=code_size)
discriminator=Discriminator(from_rgb_activate=from_rgb_act)
g_running = StyledGenerator(code_size)
g_optimizer = jt.optim.Adam(generator.generator.parameters(), lr=lr, betas=(0.0, 0.99))
g_optimizer.add_param_group({
'params': generator.style.parameters(),
'lr': lr * 0.01,
'mult': 0.01,
}
)
d_optimizer = jt.optim.Adam(discriminator.parameters(), lr=lr, betas=(0.0, 0.99))
accumulate(g_running, generator, 0)
if ckpt is not None:
ckpt = jt.load(ckpt)
generator.load_state_dict(ckpt['generator'])
discriminator.load_state_dict(ckpt['discriminator'])
g_running.load_state_dict(ckpt['g_running'])
print('from checkpoint .......')
if sched:
lr={128: 0.0015, 256: 0.002, 512: 0.003, 1024: 0.003}
batch_size={4: 512, 8: 256, 16: 128, 32: 64, 64: 32, 128: 16}
else:
lr={}
batch_size={}
gen_sample={512: (8, 4), 1024: (4, 2)}
batch_default=32
phase = 200_000
max_iter = 300_000
## Actual Training
step = int(math.log2(init_size) - 2)
resolution=int(4 * 2**step)
image_loader = SymbolDataset(path,resolution).set_attrs(
batch_size=batch_size.get(resolution, batch_default),
shuffle=True
)
data_loader = iter(image_loader)
adjust_lr(g_optimizer, lr.get(resolution, 0.001))
adjust_lr(d_optimizer, lr.get(resolution, 0.001))
pbar = tqdm(range(max_iter))
requires_grad(generator, False)
requires_grad(discriminator, True)
disc_loss_val = 0
gen_loss_val = 0
grad_loss_val = 0
alpha = 0
used_sample = 0
max_step = int(math.log2(max_size) - 2)
final_progress = False
for i in pbar:
alpha = min(1, 1 / phase * (used_sample + 1))
if (resolution == init_size and ckpt is None) or final_progress:
alpha = 1
if used_sample > phase * 2:
used_sample = 0
step += 1
if step > max_step:
step = max_step
final_progress = True
ckpt_step = step + 1
else:
alpha = 0
ckpt_step = step
resolution = 4 * 2 ** step
image_loader = SymbolDataset(path,resolution).set_attrs(
batch_size=batch_size.get(resolution, batch_default),
shuffle=True
)
data_loader = iter(image_loader)
jt.save(
{
'generator': generator.state_dict(),
'discriminator': discriminator.state_dict(),
'g_optimizer': g_optimizer.state_dict(),
'd_optimizer': d_optimizer.state_dict(),
'g_running': g_running.state_dict(),
},
f'checkpoint/train_step-{ckpt_step}.model',
)
adjust_lr(g_optimizer, lr.get(resolution, 0.001))
adjust_lr(d_optimizer, lr.get(resolution, 0.001))
try:
real_image = next(data_loader)
except (OSError, StopIteration):
data_loader = iter(image_loader)
real_image = next(data_loader)
used_sample += real_image.shape[0]
b_size = real_image.size(0)
if mixing and random.random() < 0.9:
gen_in11, gen_in12, gen_in21, gen_in22 = jt.randn(4, b_size, code_size).chunk(4, 0)
gen_in1 = [gen_in11.squeeze(0), gen_in12.squeeze(0)]
gen_in2 = [gen_in21.squeeze(0), gen_in22.squeeze(0)]
else:
gen_in1, gen_in2 = jt.randn(2, b_size, code_size).chunk(2, 0)
gen_in1 = gen_in1.squeeze(0)
gen_in2 = gen_in2.squeeze(0)
fake_image = generator(gen_in1, step=step, alpha=alpha)
fake_predict = discriminator(fake_image, step=step, alpha=alpha)
if use_loss == 'wgan-gp':
real_predict = discriminator(real_image, step=step, alpha=alpha)
real_predict = -real_predict.mean() + 0.001 * (real_predict ** 2).mean()
fake_predict = fake_predict.mean()
eps = jt.randn(b_size, 1, 1, 1)
x_hat = eps * real_image.data + (1 - eps) * fake_image.data
x_hat.requires_grad = True
hat_predict = discriminator(x_hat, step=step, alpha=alpha)
grad_x_hat = jt.grad(hat_predict.sum(),x_hat)
grad_penalty = (
(grad_x_hat.reshape(grad_x_hat.size(0), -1).norm(2, dim=1) - 1) ** 2
).mean()
grad_penalty = 10 * grad_penalty
if i%10 == 0:
grad_loss_val = grad_penalty.item()
if i % 10 == 0:
disc_loss_val = (real_predict + fake_predict).item()
loss_D = real_predict+grad_penalty+fake_predict
d_optimizer.step(loss_D)
elif use_loss == 'r1':
real_image.requires_grad = True
real_scores = discriminator(real_image, step=step, alpha=alpha)
real_predict = jt.nn.softplus(-real_scores).mean()
grad_real = jt.grad(real_scores.sum(), real_image)
grad_penalty = (
grad_real.reshape(grad_real.size(0), -1).norm(2, dim=1) ** 2
).mean()
grad_penalty = 10 / 2 * grad_penalty
if i % 10 == 0:
grad_loss_val = grad_penalty.item()
fake_predict = jt.nn.softplus(fake_predict).mean()
if i % 10 == 0:
disc_loss_val = (real_predict + fake_predict).item()
loss_D = real_predict + grad_penalty + fake_predict
d_optimizer.step(loss_D)
# optimize generator
requires_grad(generator, True)
requires_grad(discriminator, False)
fake_image = generator(gen_in2, step=step, alpha=alpha)
predict = discriminator(fake_image, step=step, alpha=alpha)
if use_loss == 'wgan-gp':
loss_G = -predict.mean()
elif use_loss == 'r1':
loss_G = jt.nn.softplus(-predict).mean()
if i % 10 == 0:
gen_loss_val = loss_G.item()
g_optimizer.step(loss_G)
accumulate(g_running, generator)
requires_grad(generator, False)
requires_grad(discriminator, True)
if (i + 1) % 100 == 0:
images = []
gen_i, gen_j = gen_sample.get(resolution, (10, 5))
with jt.no_grad():
for _ in range(gen_i):
images.append(
g_running(
jt.randn(gen_j, code_size), step=step, alpha=alpha
).data
)
jt.save_image(
jt.concat(images, 0),
f'sample/{str(i + 1).zfill(6)}.png',
nrow=gen_i,
normalize=True,
range=(-1, 1),
)
if (i + 1) % 100 == 0:
jt.save({'g_running':g_running.state_dict(),'step':step}, f'checkpoint/{str(i + 1).zfill(6)}.model')
state_msg = (
f'Size: {4 * 2 ** step}; G: {gen_loss_val:.3f}; D: {disc_loss_val:.3f};'
f' Grad: {grad_loss_val:.3f}; Alpha: {alpha:.5f}'
)
pbar.set_description(state_msg)