-
Notifications
You must be signed in to change notification settings - Fork 0
/
acmain.py
167 lines (144 loc) · 7.67 KB
/
acmain.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
import model
import argparse
import torch
import dataset_generator
import torch.utils.data
from torch.autograd import Variable
import torchvision.transforms as transforms
import torchvision.utils as vutils
import os
import torch.nn.functional as F
import numpy as np
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--data-folder', required=True, help='path to folder with images')
parser.add_argument('--data-file', required=True, help='path to images description file')
parser.add_argument('--experiment', required=True, help='where to store samples and models')
parser.add_argument('--img-size', type=int, default=64, help='image size')
parser.add_argument('--batch-size', type=int, default=64, help='batch size')
parser.add_argument('--n-workers', type=int, default=2, help='number of workers to load dataset')
parser.add_argument('--d-iter', type=int, default=5, help='discriminator iters per generator iter')
parser.add_argument('--n-channels', type=int, default=3, help='number of channels')
parser.add_argument('--n-gen-features', type=int, default=64, help='number of generator features')
parser.add_argument('--n-disc-features', type=int, default=64, help='number of discriminator features')
parser.add_argument('--dim-z', type=int, default=100, help='dimensionality of a latent vector')
parser.add_argument('--n-iter', type=int, default=20, help='number of training epochs')
parser.add_argument('--cuda', action='store_true', help='enables cuda')
parser.add_argument('--net-gen', default='', help='path to generator')
parser.add_argument('--net-disc', default='', help='path to discriminator')
parser.add_argument('--gp-coef', type=float, default=10, help='GP coef')
parser.add_argument('--ac-coef', type=float, default=0.2, help='AC coef')
parser.add_argument('--classes', type=int, nargs='+', help='Classes labels to train')
parser.add_argument('--frequency', type=int, default=500, help='How often to save models and generate images')
args = parser.parse_args()
# Creating experiment folder
if not os.path.exists(args.experiment):
os.makedirs(args.experiment)
# Creating models
signs_subset = args.classes
print(signs_subset)
n_classes = len(args.classes)
if args.net_gen:
gen = torch.load(args.net_gen)
print('loaded generator')
else:
gen = model.BNGenerator(img_size=args.img_size, y_dim=n_classes, z_dim=args.dim_z,
n_channels=args.n_channels, n_features=args.n_gen_features)
gen.apply(model.weights_init)
if args.net_disc:
disc = torch.load(args.net_disc)
print('loaded discriminator')
else:
disc = model.ACDiscriminator(img_size=args.img_size, n_classes=n_classes, n_channels=args.n_channels,
n_features=args.n_disc_features)
disc.apply(model.weights_init)
if args.cuda:
gen.cuda()
disc.cuda()
# Creating optimizers
opt_gen = torch.optim.Adam(gen.parameters(), lr=1e-4, betas=(0, 0.9))
opt_disc = torch.optim.Adam(disc.parameters(), lr=1e-4, betas=(0, 0.9))
# Creating dataloader
dataset = dataset_generator.SignsDataset(args.data_file, args.data_folder, signs_subset=signs_subset,
transform=transforms.Compose([
transforms.Scale(args.img_size),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]))
dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size,
shuffle=True, num_workers=args.n_workers, drop_last=True)
# Creating auxiliary tensors
noise = torch.FloatTensor(args.batch_size, args.dim_z)
fixed_noise = torch.FloatTensor(n_classes * 8, args.dim_z).normal_(0, 1)
fixed_labels = np.repeat(np.arange(n_classes), 8)
fixed_labels = torch.LongTensor(fixed_labels)
eps = torch.FloatTensor(args.batch_size, 1)
ones = torch.FloatTensor(args.batch_size).fill_(1.0)
weights = dataset.get_weights()
one_hot = torch.FloatTensor(args.batch_size, 8)
if args.cuda:
noise = noise.cuda()
fixed_noise = fixed_noise.cuda()
fixed_labels = fixed_labels.cuda()
eps = eps.cuda()
ones = ones.cuda()
weights = weights.cuda()
one_hot = one_hot.cuda()
# Training loop
ac_coef = args.ac_coef
gp_coef = args.gp_coef
gen_iteration = 0
while gen_iteration < args.n_iter:
data = iter(dataloader)
i = 0
while i < len(dataloader):
j = 0
while j < args.d_iter and i < len(dataloader):
real_images, real_labels = next(data)
noise.normal_(0, 1)
if args.cuda:
real_images = real_images.cuda()
real_labels = real_labels.cuda()
current_weights = Variable(weights[real_labels])
fake_labels = real_labels.clone()
fake_images = gen(Variable(noise), Variable(fake_labels))
eps.uniform_()
alpha = eps.unsqueeze(2).unsqueeze(3).expand(real_images.size())
interpolate_images = alpha * real_images + (1 - alpha) * fake_images.data
opt_disc.zero_grad()
wass_fake, logits_fake = disc(fake_images)
wass_real, logits_real = disc(Variable(real_images))
wass_loss = (wass_fake - wass_real)
ce_loss = F.cross_entropy(logits_real, Variable(real_labels), weight=weights)
ones.resize_as_(wass_loss.data)
grad_norms = disc.grad_norm(interpolate_images, ones)
grad_loss = (grad_norms - 1) ** 2
disc_loss = (torch.mean(wass_loss.squeeze() * current_weights) + ce_loss +
gp_coef * torch.mean(grad_loss * current_weights))
opt_disc.zero_grad()
disc_loss.backward()
opt_disc.step()
j += 1
i += 1
noise.normal_(0, 1)
current_weights = Variable(weights[fake_labels])
fake_images = gen(Variable(noise), Variable(fake_labels))
opt_gen.zero_grad()
wass_fake, logits_fake = disc(fake_images)
ce_loss = F.cross_entropy(logits_fake, Variable(fake_labels), weight=weights)
gen_loss = -torch.mean(wass_fake.squeeze() * current_weights) + ac_coef * ce_loss
gen_loss.backward()
opt_gen.step()
gen_iteration += 1
print('[{}/{}] disc_loss = {} \t gen_loss = {}'.format(gen_iteration, args.n_iter,
disc_loss.cpu().data.numpy()[0],
gen_loss.cpu().data.numpy()[0]))
if gen_iteration % args.frequency == 0 or gen_iteration == 1:
disc_path = os.path.join(args.experiment, 'disc_{}.pth'.format(gen_iteration))
gen_path = os.path.join(args.experiment, 'gen_{}.pth'.format(gen_iteration))
torch.save(disc, disc_path)
torch.save(gen, gen_path)
fake_data = gen(Variable(fixed_noise), Variable(fixed_labels)).cpu().data
fake_data = fake_data.mul(0.5).add(0.5)
fake_data_path = os.path.join(args.experiment, 'fake_samples_{}.png'.format(gen_iteration))
vutils.save_image(fake_data, fake_data_path, nrow=n_classes)