This repository has been archived by the owner on Dec 13, 2021. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 40
/
train.py
216 lines (167 loc) · 7.69 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 functools
import imlib as im
import numpy as np
import pylib as py
import tensorboardX
import torch
import torchlib
import torchprob as gan
import tqdm
import data
import module
# ==============================================================================
# = param =
# ==============================================================================
# command line
py.arg('--dataset', default='fashion_mnist', choices=['cifar10', 'fashion_mnist', 'mnist', 'celeba', 'anime', 'custom'])
py.arg('--batch_size', type=int, default=64)
py.arg('--epochs', type=int, default=25)
py.arg('--lr', type=float, default=0.0002)
py.arg('--beta_1', type=float, default=0.5)
py.arg('--n_d', type=int, default=1) # # d updates per g update
py.arg('--z_dim', type=int, default=128)
py.arg('--adversarial_loss_mode', default='gan', choices=['gan', 'hinge_v1', 'hinge_v2', 'lsgan', 'wgan'])
py.arg('--gradient_penalty_mode', default='none', choices=['none', '1-gp', '0-gp', 'lp'])
py.arg('--gradient_penalty_sample_mode', default='line', choices=['line', 'real', 'fake', 'dragan'])
py.arg('--gradient_penalty_weight', type=float, default=10.0)
py.arg('--experiment_name', default='none')
py.arg('--gradient_penalty_d_norm', default='layer_norm', choices=['instance_norm', 'layer_norm']) # !!!
args = py.args()
# output_dir
if args.experiment_name == 'none':
args.experiment_name = '%s_%s' % (args.dataset, args.adversarial_loss_mode)
if args.gradient_penalty_mode != 'none':
args.experiment_name += '_%s_%s' % (args.gradient_penalty_mode, args.gradient_penalty_sample_mode)
output_dir = py.join('output', args.experiment_name)
py.mkdir(output_dir)
# save settings
py.args_to_yaml(py.join(output_dir, 'settings.yml'), args)
# others
use_gpu = torch.cuda.is_available()
device = torch.device("cuda" if use_gpu else "cpu")
# ==============================================================================
# = data =
# ==============================================================================
# setup dataset
if args.dataset in ['cifar10', 'fashion_mnist', 'mnist']: # 32x32
data_loader, shape = data.make_32x32_dataset(args.dataset, args.batch_size, pin_memory=use_gpu)
n_G_upsamplings = n_D_downsamplings = 3
elif args.dataset == 'celeba': # 64x64
img_paths = py.glob('data/img_align_celeba', '*.jpg')
data_loader, shape = data.make_celeba_dataset(img_paths, args.batch_size, pin_memory=use_gpu)
n_G_upsamplings = n_D_downsamplings = 4
elif args.dataset == 'anime': # 64x64
img_paths = py.glob('data/faces', '*.jpg')
data_loader, shape = data.make_anime_dataset(img_paths, args.batch_size, pin_memory=use_gpu)
n_G_upsamplings = n_D_downsamplings = 4
elif args.dataset == 'custom':
# ======================================
# = custom =
# ======================================
img_paths = ... # image paths of custom dataset
data_loader = data.make_custom_dataset(img_paths, args.batch_size, pin_memory=use_gpu)
n_G_upsamplings = n_D_downsamplings = ... # 3 for 32x32 and 4 for 64x64
# ======================================
# = custom =
# ======================================
# ==============================================================================
# = model =
# ==============================================================================
# setup the normalization function for discriminator
if args.gradient_penalty_mode == 'none':
d_norm = 'batch_norm'
else: # cannot use batch normalization with gradient penalty
d_norm = args.gradient_penalty_d_norm
# networks
G = module.ConvGenerator(args.z_dim, shape[-1], n_upsamplings=n_G_upsamplings).to(device)
D = module.ConvDiscriminator(shape[-1], n_downsamplings=n_D_downsamplings, norm=d_norm).to(device)
print(G)
print(D)
# adversarial_loss_functions
d_loss_fn, g_loss_fn = gan.get_adversarial_losses_fn(args.adversarial_loss_mode)
# optimizer
G_optimizer = torch.optim.Adam(G.parameters(), lr=args.lr, betas=(args.beta_1, 0.999))
D_optimizer = torch.optim.Adam(D.parameters(), lr=args.lr, betas=(args.beta_1, 0.999))
# ==============================================================================
# = train step =
# ==============================================================================
def train_G():
G.train()
D.train()
z = torch.randn(args.batch_size, args.z_dim, 1, 1).to(device)
x_fake = G(z)
x_fake_d_logit = D(x_fake)
G_loss = g_loss_fn(x_fake_d_logit)
G.zero_grad()
G_loss.backward()
G_optimizer.step()
return {'g_loss': G_loss}
def train_D(x_real):
G.train()
D.train()
z = torch.randn(args.batch_size, args.z_dim, 1, 1).to(device)
x_fake = G(z).detach()
x_real_d_logit = D(x_real)
x_fake_d_logit = D(x_fake)
x_real_d_loss, x_fake_d_loss = d_loss_fn(x_real_d_logit, x_fake_d_logit)
gp = gan.gradient_penalty(functools.partial(D), x_real, x_fake, gp_mode=args.gradient_penalty_mode, sample_mode=args.gradient_penalty_sample_mode)
D_loss = (x_real_d_loss + x_fake_d_loss) + gp * args.gradient_penalty_weight
D.zero_grad()
D_loss.backward()
D_optimizer.step()
return {'d_loss': x_real_d_loss + x_fake_d_loss, 'gp': gp}
@torch.no_grad()
def sample(z):
G.eval()
return G(z)
# ==============================================================================
# = run =
# ==============================================================================
# load checkpoint if exists
ckpt_dir = py.join(output_dir, 'checkpoints')
py.mkdir(ckpt_dir)
try:
ckpt = torchlib.load_checkpoint(ckpt_dir)
ep, it_d, it_g = ckpt['ep'], ckpt['it_d'], ckpt['it_g']
D.load_state_dict(ckpt['D'])
G.load_state_dict(ckpt['G'])
D_optimizer.load_state_dict(ckpt['D_optimizer'])
G_optimizer.load_state_dict(ckpt['G_optimizer'])
except:
ep, it_d, it_g = 0, 0, 0
# sample
sample_dir = py.join(output_dir, 'samples_training')
py.mkdir(sample_dir)
# main loop
writer = tensorboardX.SummaryWriter(py.join(output_dir, 'summaries'))
z = torch.randn(100, args.z_dim, 1, 1).to(device) # a fixed noise for sampling
for ep_ in tqdm.trange(args.epochs, desc='Epoch Loop'):
if ep_ < ep:
continue
ep += 1
# train for an epoch
for x_real in tqdm.tqdm(data_loader, desc='Inner Epoch Loop'):
x_real = x_real.to(device)
D_loss_dict = train_D(x_real)
it_d += 1
for k, v in D_loss_dict.items():
writer.add_scalar('D/%s' % k, v.data.cpu().numpy(), global_step=it_d)
if it_d % args.n_d == 0:
G_loss_dict = train_G()
it_g += 1
for k, v in G_loss_dict.items():
writer.add_scalar('G/%s' % k, v.data.cpu().numpy(), global_step=it_g)
# sample
if it_g % 100 == 0:
x_fake = sample(z)
x_fake = np.transpose(x_fake.data.cpu().numpy(), (0, 2, 3, 1))
img = im.immerge(x_fake, n_rows=10).squeeze()
im.imwrite(img, py.join(sample_dir, 'iter-%09d.jpg' % it_g))
# save checkpoint
torchlib.save_checkpoint({'ep': ep, 'it_d': it_d, 'it_g': it_g,
'D': D.state_dict(),
'G': G.state_dict(),
'D_optimizer': D_optimizer.state_dict(),
'G_optimizer': G_optimizer.state_dict()},
py.join(ckpt_dir, 'Epoch_(%d).ckpt' % ep),
max_keep=1)