/
main.py
222 lines (168 loc) · 8.32 KB
/
main.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
import torch
from torch import nn
from torch import optim
from torch.utils.data import Dataset, DataLoader
from torch.nn import functional as F
from torchvision import datasets, transforms
from torch.autograd import Variable
from torchvision.utils import save_image
from torchvision.utils import make_grid
import itertools
import numpy as np
import argparse
import matplotlib.pyplot as plt
import matplotlib.cm as cm
from PIL import Image
import os
import cv2
from operator import itemgetter
import matplotlib.pyplot as plt
from PIL import Image
from data import *
from model import *
from utils import *
from PIL import Image
import torch.nn.functional as F
from torch.autograd import Variable
import torch.autograd as autograd
# file revised
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# setting args
parser = argparse.ArgumentParser(description = "StarGAN")
parser.add_argument('--batch_size', type = int, default =64, help = "batch_size")
parser.add_argument('--epoch', type = int, default = 150, help = "epoch")
parser.add_argument('--learning_rate_g', type = float, default = 0.0001, help = "learning_rate")
parser.add_argument('--learning_rate_d', type = float, default = 0.0001, help = "learning_rate")
parser.add_argument("--b1", type=float, default=0.5, help="adam: decay of first order momentum of gradient")
parser.add_argument("--b2", type=float, default=0.999, help="adam: decay of first order momentum of gradient")
parser.add_argument("--recon_lambda", type= float, default=10.0, help="recon lambda")
parser.add_argument("--cls_lambda", type= float, default=1.0, help="cls lambda")
parser.add_argument("--gp_lambda", type= float, default=10.0, help="gp lambda")
parser.add_argument("--n_residual_blocks", type=int, default=6, help="number of residual blocks")
parser.add_argument("--n_attribute", type=int, default=5, help="number of attributes")
parser.add_argument("--n_domain", type=int, default=5, help="number of domains")
parser.add_argument("--out_nc", type=int, default=64, help="number of output channels")
args = parser.parse_args()
os.makedirs('result', exist_ok = True)
transforms_ = transforms.Compose([
transforms.CenterCrop(178),
transforms.Resize(128),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
#transforms_ = [
# transforms.CenterCrop(178),
# transforms.Resize(128),
# transforms.ToTensor(),
# transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
#]
#
def weights_init_normal(m):
classname = m.__class__.__name__
if classname.find("Conv") != -1:
torch.nn.init.normal_(m.weight.data, 0.0, 0.02)
elif classname.find("BatchNorm2d") != -1:
torch.nn.init.normal_(m.weight.data, 1.0, 0.02)
torch.nn.init.constant_(m.bias.data, 0.0)
train_data = CelebADataset('dataset/celeba', transforms_ = transforms_, mode = "train")
## test_data = CelebDataset('dataset/celeba', train = False, transform = transforms_)
train_loader = DataLoader(dataset = train_data, batch_size = args.batch_size, shuffle = True)
## test_loader = DataLoader(dataset = test_data, batch_size = args.batch_size, shuffle = True)
print("data loaded")
generator = Generator(args.n_residual_blocks, args.out_nc, args.n_attribute).to(device)
discriminator = Discriminator(args.out_nc, args.n_domain).to(device)
generator.apply(weights_init_normal)
discriminator.apply(weights_init_normal)
optimizer_g = optim.Adam(generator.parameters(), lr = args.learning_rate_g, betas = (args.b1, args.b2))
optimizer_d = optim.Adam(discriminator.parameters(), lr = args.learning_rate_d, betas = (args.b1, args.b2))
criterion_gan = nn.BCELoss()
criterion_recon = nn.L1Loss()
""" Below two functions from https://github.com/eriklindernoren/PyTorch-GAN/blob/master/implementations/stargan/stargan.py """
def criterion_cls(input, target):
return F.binary_cross_entropy_with_logits(input, target, size_average=False) / input.size(0)
def compute_gradient_penalty(D, real_samples, fake_samples):
"""Calculates the gradient penalty loss for WGAN GP"""
# Random weight term for interpolation between real and fake samples
Tensor = torch.cuda.FloatTensor
alpha = Tensor(np.random.random((real_samples.size(0), 1, 1, 1)))
# Get random interpolation between real and fake samples
interpolates = (alpha * real_samples + ((1 - alpha) * fake_samples)).requires_grad_(True)
d_interpolates, _ = D(interpolates)
fake = Variable(Tensor(np.ones(d_interpolates.shape)), requires_grad=False)
# Get gradient w.r.t. interpolates
gradients = autograd.grad(
outputs=d_interpolates,
inputs=interpolates,
grad_outputs=fake,
create_graph=True,
retain_graph=True,
only_inputs=True,
)[0]
gradients = gradients.view(gradients.size(0), -1)
gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean()
return gradient_penalty
generator.train()
discriminator.train()
for epoch in range(args.epoch):
for i, (image, attr) in enumerate(train_loader):
image = Variable(image.to(device))
image = image.float()
origin_attr = Variable(attr.to(device))
target_attr = []
for index in range(args.batch_size):
target_attr.append([np.random.randint(2) for index in range(origin_attr.shape[1])])
target_attr = torch.tensor(target_attr).to(device)
target_attr = target_attr.float()
if image.shape[0] < args.batch_size: break
# update generator
fake = generator(image, target_attr)
# update discriminator
optimizer_d.zero_grad()
adv_real, cls_real = discriminator(image)
adv_fake, cls_fake = discriminator(fake.detach())
gradient_penalty = compute_gradient_penalty(discriminator, image, fake)
#loss_d_adv_fake = criterion_gan(adv_fake.detach(), fake(adv_fake))
#loss_d_adv_real = criterion_gan(adv_real, valid(adv_real))
#loss_d_adv = 0.5 * (loss_d_adv_fake + loss_d_adv_real)
loss_d_adv = -torch.mean(adv_real) + torch.mean(adv_fake) + args.gp_lambda * gradient_penalty
loss_d_cls = criterion_cls(cls_real, origin_attr)
loss_d = loss_d_adv + args.cls_lambda * loss_d_cls
loss_d.backward()
optimizer_d.step()
optimizer_g.zero_grad()
if i % 5 == 0:
generated_image = generator(image, target_attr)
recon = generator(generated_image, origin_attr)
adv_fake, cls_fake = discriminator(generated_image)
loss_g_recon = criterion_recon(recon, image)
loss_g_cls = criterion_cls(cls_fake, target_attr)
# generator_loss
loss_g_adv = -torch.mean(adv_fake)
# loss_g_adv = criterion_gan(adv_fake, valid(adv_fake))
# criterion_gan(adv_fake, valid(adv_fake))
loss_g = loss_g_adv + args.recon_lambda * loss_g_recon + args.cls_lambda * loss_g_cls
loss_g.backward()
optimizer_g.step()
total_loss = loss_d + loss_g
if i % 20 == 0:
print('Epoch [{}/{}], Step [{}/{}], Loss:{:.4f}, G-Loss: {:.4f}, G-ADV-Loss:{:.4f}, G-CLS-Loss:{:.4f}, G-Recon-Loss:{:.4f}, D-Loss: {:.4f}, D-ADV-Loss:{:4f}, D-CLS-Loss:{:4f} '.format(epoch+1, args.epoch, i, int(len(train_loader)), total_loss.item(), loss_g.item(), loss_g_adv.item(), loss_g_cls.item(), loss_g_recon.item(), loss_d.item(), loss_d_adv.item(), loss_d_cls.item()))
with torch.no_grad():
original = image[0]
orig_attr = attr[0]
fake1 = generator(original.unsqueeze(0), torch.FloatTensor([(orig_attr[0] + 1) % 2,0,0,0,0]).unsqueeze(0).to(device))
fake2 = generator(original.unsqueeze(0), torch.FloatTensor([0,(orig_attr[1] + 1) % 2,0,0,0]).unsqueeze(0).to(device))
fake3 = generator(original.unsqueeze(0), torch.FloatTensor([0,0, (orig_attr[2] + 1) % 2,0,0]).unsqueeze(0).to(device))
fake4 = generator(original.unsqueeze(0), torch.FloatTensor([0,0,0,(orig_attr[3] + 1) % 2,0]).unsqueeze(0).to(device))
fake5 = generator(original.unsqueeze(0), torch.FloatTensor([0,0,0,0,(orig_attr[4] + 1) % 2]).unsqueeze(0).to(device))
result = torch.cat((original, fake1.squeeze(0), fake2.squeeze(0), fake3.squeeze(0), fake4.squeeze(0), fake5.squeeze(0)), 1)
save_image(result, 'result/%d_%d.png' % (epoch, i), normalize = True)
#fake = make_grid(fake[0], nrow = 1, normalize = True)
#recon = make_grid(recon[0], nrow = 1, normalize = True)
#image_grid = torch.cat((original, fake, recon), 1)
#save_image(image_grid,
#plot_list = []
#plot_list.append(image[0])
#plot_list.append(fake[0])
#plot_list.append(recon[0])
#plot_list = torch.stack(plot_list)
#save_image(result, 'result/%d_%d | ' % (epoch, i) + str(origin_attr[0])[8:26] + '|' + str(target_attr[0])[8:26] + '.png', normalize = False)