-
Notifications
You must be signed in to change notification settings - Fork 12
/
train.py
216 lines (164 loc) · 7.79 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
import torch
from torch import nn, autograd, optim
from torch.nn import functional as F
from stylegan2.stylegan2_pytorch.model import Generator, MLP
from stylegan2.stylegan2_pytorch.model import Discriminator
from torch.utils.data import DataLoader
from criteria.lpips.lpips import LPIPS
from torchvision import utils, transforms
from stylegan2.stylegan2_pytorch.train import g_path_regularize, make_noise, mixing_noise, d_r1_loss
from configs import transforms_config
from datasets.images_dataset import ImagesDataset
from stylegan2.stylegan2_pytorch.non_leaking import augment, AdaptiveAugment
from torch.utils.tensorboard import SummaryWriter
import numpy as np
writer = SummaryWriter(log_dir='outcome/log8')
def adv_loss(y_fake, y_real):
# adv_loss = torch.min(0, -1 + y_real) + torch.min(0, -1 - y_fake)
# return adv_loss
real_loss = F.softplus(-y_real)
fake_loss = F.softplus(y_fake)
return real_loss.mean() + fake_loss.mean()
def g_nonsaturating_loss(fake_pred):
loss = F.softplus(-fake_pred).mean()
return loss
def requires_grad(model, flag=True):
for p in model.parameters():
p.requires_grad = flag
def configure_datasets():
transforms_dict = transforms_config.EncodeTransforms.get_transforms()
train_dataset = ImagesDataset(source_root='/home/ai/project/dataset/agile_train',
source_transform=transforms_dict['transform_gt_train']
)
test_dataset = ImagesDataset(source_root='/home/ai/project/dataset/portrait2_test',
source_transform=transforms_dict['transform_gt_train']
)
print(f"Number of training samples: {len(train_dataset)}")
print(f"Number of test samples: {len(test_dataset)}")
return train_dataset, test_dataset
def train():
G_path = "/home/ai/project/agile3/pretrained_models/stylegan2-ffhq-config-f.pt"
device0 = 'cuda:0'
device1 = 'cuda:1'
batch_size = 2
output_size = 1024
learning_rate = 0.0002
w1 = 1
w2 = 0.4
# load generator0 and genetator t
G0 = Generator(output_size, 512, 8).to(device0)
Gt = Generator(output_size, 512, 8).to(device1)
checkpoint = torch.load(G_path)
G0.load_state_dict(checkpoint["g_ema"], strict=False)
Gt.load_state_dict(checkpoint["g_ema"], strict=False)
mlp = MLP(1024, 512, 8).to(device0)
mlp.load_state_dict(checkpoint["g_ema"], strict=False)
# load discriminator
D = Discriminator(size=1024).to(device1)
D.load_state_dict(checkpoint["d"], strict=True)
train_dataset, test_dataset = configure_datasets()
train_loader = DataLoader(train_dataset,
batch_size,
shuffle=True,
num_workers=4,
drop_last=True,
)
lpips = LPIPS(net_type='vgg').to(device0).eval() # DISCARD 9 LAYERS
loss_dict = {}
num_epochs = 1000
Gt_optim = optim.Adam(Gt.parameters(), lr=learning_rate)
D_optim = optim.Adam(D.parameters(), lr=learning_rate)
number = 0
for epoch in range(num_epochs):
for batch_idx, batch in enumerate(train_loader):
number += 1
real_img = batch
real_img = real_img.to(device0)
real_img = F.interpolate(real_img, size=[1024, 1024])
# Update discriminator
requires_grad(Gt, False)
requires_grad(G0, False)
requires_grad(D, True)
# noise = mixing_noise(batch_size, latent_dim=512, prob=0.9, device=device0)
noise = torch.randn(batch_size, 18, 512).to(device0).unbind(0)
z = mlp(noise)
z = torch.stack(z, dim=0).to(device1)
fake_img_t = Gt(z, input_is_latent=True, noise=None, return_latents=False)
fake_img_t = fake_img_t.detach()
#fake_img_t256 = F.interpolate(fake_img_t, size = [256, 256])
y_fake = D(fake_img_t)
y_real = D(real_img.to(device1))
loss_adv = adv_loss(y_fake, y_real)
D_loss = loss_adv
loss_dict["d"] = D_loss.to(device0)
D.zero_grad()
D_loss.backward()
D_optim.step()
real_img.requires_grad = True
y_real = D(real_img.to(device1))
r1_loss = d_r1_loss(y_real, real_img)
D.zero_grad()
r1_loss.backward()
D_optim.step()
loss_dict["r1"] = r1_loss.to(device0)
# Update generator
requires_grad(Gt, True)
requires_grad(G0, False)
requires_grad(D, False)
z0 = torch.randn(batch_size, 18, 512, device=device1)
z = mlp(z0.to(device0))
z = torch.stack(z, dim=0)
fake_img_t = Gt(z.to(device1), input_is_latent=True, noise=None, return_latents=False)
#fake_img_t256 = F.interpolate(fake_img_t, size=[256, 256])
y_fake = D(fake_img_t)
nonsaturating_loss = g_nonsaturating_loss(y_fake).to(device0)
loss_dict["nonsa_loss"] = nonsaturating_loss
fake_img_0 = G0(z, input_is_latent=True, noise=None).detach()
lpips_loss = lpips(fake_img_t.to(device0), fake_img_0).to(device0)
loss_dict["lpip_loss"] = lpips_loss
if number % 20 == 0:
image = torch.cat((fake_img_0.to(device0), fake_img_t.to(device0)), 0)
utils.save_image(
image,
"outcome/images8/" + str(number) + "output.jpg",
nrow=2,
normalize=True,
range=(-1, 1),
)
Gt_loss = w1 * lpips_loss + w2 * nonsaturating_loss
loss_dict["g"] = Gt_loss
Gt.zero_grad()
Gt_loss.backward()
Gt_optim.step()
z0 = torch.randn(batch_size, 18, 512, device=device0)
z = mlp(z0)
z = torch.stack(z, dim=0).to(device1)
fake_img_t, latents = Gt(z, input_is_latent=True, noise=None, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(fake_img_t, latents, mean_path_length=0)
Gt.zero_grad()
path_loss = 2 * path_loss
loss_dict["path"] = path_loss.to(device0)
path_loss.backward()
Gt_optim.step()
if number % 20 == 0:
print("step: ", number)
print("d loss: ", loss_dict["d"].item())
print("g loss: ", loss_dict["g"].item())
print("nonsa loss: ", loss_dict["nonsa_loss"].item())
print("lpip loss: ", loss_dict["lpip_loss"].item())
print("r1 loss: ", loss_dict["r1"].item())
print("path loss: ", loss_dict["path"].item())
values = loss_dict.values()
print("loss: ", loss_dict["g"].item() + loss_dict["d"].item())
print()
writer.add_scalar('Loss/train/total', loss_dict["g"].item() + loss_dict["d"].item(), number)
writer.add_scalar('Loss/train/total', loss_dict["nonsa_loss"].item(), number)
writer.add_scalar('Loss/train/total', loss_dict["lpip_loss"].item(), number)
writer.add_scalar('Loss/train/total', loss_dict["r1"].item(), number)
writer.add_scalar('Loss/train/total', loss_dict["path"].item(), number)
writer.add_scalar('Loss/train/d', loss_dict["d"].item(), number)
writer.add_scalar('Loss/train/g2', loss_dict["g"].item(), number)
if ((epoch + 1) * batch_idx) % 50 == 0:
state = {'Gt': Gt.state_dict()}
torch.save(state, '/home/ai/project/agile_stage2/outcome/model_save8/Gt' + str(number) +'.pth')
train()