forked from Aixile/chainer-cyclegan
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
195 lines (155 loc) · 8.28 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
#!/usr/bin/env python
import argparse
from chainer import training
from chainer.training import extensions
import common.datasets as datasets
from common.evaluation.visualization import *
from common.models.discriminators import *
from common.models.transformers import *
from common.utils import *
from updater import *
import matplotlib
matplotlib.use('Agg')
def main():
parser = argparse.ArgumentParser(
description='Train CycleGAN')
parser.add_argument('--batchsize', '-b', type=int, default=1)
parser.add_argument('--max_iter', '-m', type=int, default=200000)
parser.add_argument('--gpu', '-g', type=int, default=0,
help='GPU ID (negative value indicates CPU)')
parser.add_argument('--out', '-o', default='result',
help='Directory to output the result')
parser.add_argument('--eval_interval', type=int, default=400,
help='Interval of evaluating generator')
parser.add_argument('--snapshot_interval', type=int, default=4000,
help='Interval of model snapshot')
parser.add_argument("--learning_rate_g", type=float, default=0.0002,
help="Learning rate for generator")
parser.add_argument("--learning_rate_d", type=float, default=0.0002,
help="Learning rate for discriminator")
parser.add_argument("--data_train_x", default='', help='path of train data of x')
parser.add_argument("--data_train_y", default='', help='path of train data of y')
parser.add_argument("--data_test_x", type=str, help='path of test data of x')
parser.add_argument("--data_test_y", type=str, help='path of test data of y')
parser.add_argument("--resume", type = str, help='trainer snapshot to be resumed')
parser.add_argument("--load_gen_f_model", type=str, help='load generator model')
parser.add_argument("--load_gen_g_model", type=str, help='load generator model')
parser.add_argument("--load_dis_x_model", type=str, help='load discriminator model')
parser.add_argument("--load_dis_y_model", type=str, help='load discriminator model')
parser.add_argument("--resize_to", type=int, default=286, help='resize the image to')
parser.add_argument("--crop_to", type=int, default=256, help='crop the resized image to')
parser.add_argument("--lambda1", type=float, default=10.0, help='lambda for reconstruction loss')
parser.add_argument("--lambda2", type=float, default=1.0, help='lambda for adversarial loss')
parser.add_argument("--lambda_idt", type=float, default=0.5, help='lambda for identity mapping loss')
parser.add_argument("--bufsize", type=int, default=50, help='size of buffer')
# parser.add_argument("--cfmap_loss", type=int, choices = [0,1,2], help='use of cfmap loss 0: penalize gen, 1: penalize dis, 2: penalize both')
# parser.add_argument("--lambda_cfmap", type=float, default=1.0,
# help='lambda for cfmap loss')
parser.add_argument("--learning_rate_anneal", type=float, default=0.000002, help='anneal the learning rate')
parser.add_argument("--learning_rate_anneal_interval", type=int, default=1000, help='interval of learning rate anneal')
parser.add_argument("--learning_rate_anneal_trigger", type=int, default=100000, help='trigger of learning rate anneal')
parser.add_argument("--norm", type=str, default='instance', choices = ['instance','bn','None'], help='normalization method')
parser.add_argument("--reflect", type=int, choices = [0,1,2],default=2, help='reflect padding setting 0: no use, 1: at the beginning, 2: each time')
# parser.add_argument("--norm_noaffine", action='store_true')
# parser.add_argument("--norm_gnorm", action='store_true')
args = parser.parse_args()
print(args)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
if args.norm == 'None': args.norm = None
gen_g = ResNetImageTransformer(norm=args.norm, reflect=args.reflect)
gen_f = ResNetImageTransformer(norm=args.norm, reflect=args.reflect)
dis_x = DCGANDiscriminator(norm=args.norm)
dis_y = DCGANDiscriminator(norm=args.norm)
if args.load_gen_g_model:
serializers.load_npz(args.load_gen_g_model, gen_g)
print("Generator G(X->Y) model loaded")
if args.load_gen_f_model:
serializers.load_npz(args.load_gen_f_model, gen_f)
print("Generator F(Y->X) model loaded")
if args.load_dis_x_model:
serializers.load_npz(args.load_dis_x_model, dis_x)
print("Discriminator X model loaded")
if args.load_dis_y_model:
serializers.load_npz(args.load_dis_y_model, dis_y)
print("Discriminator Y model loaded")
if args.gpu >= 0:
gen_g.to_gpu()
gen_f.to_gpu()
dis_x.to_gpu()
dis_y.to_gpu()
print("use gpu {}".format(args.gpu))
opt_g=make_adam(gen_g, lr=args.learning_rate_g, beta1=0.5)
opt_f=make_adam(gen_f, lr=args.learning_rate_g, beta1=0.5)
opt_x=make_adam(dis_x, lr=args.learning_rate_d, beta1=0.5)
opt_y=make_adam(dis_y, lr=args.learning_rate_d, beta1=0.5)
train_dataset = datasets.image_pairs_train(args.data_train_x, args.data_train_y,
resize_to=args.resize_to, crop_to=args.crop_to)
train_iter = chainer.iterators.MultiprocessIterator(
train_dataset, args.batchsize, n_processes=4)
if args.data_test_x:
test_dataset = datasets.image_pairs_train(args.data_test_x, args.data_test_y,
resize_to=args.crop_to, crop_to=args.crop_to)
# Set up a trainer
updater = Updater(
models=(gen_g, gen_f, dis_x, dis_y),
iterator={
'main': train_iter,
},
optimizer={
'gen_g': opt_g,
'gen_f': opt_f,
'dis_x': opt_x,
'dis_y': opt_y
},
device=args.gpu,
params={
'lambda1': args.lambda1,
'lambda2': args.lambda2,
'lambda_idt': args.lambda_idt,
'image_size' : args.crop_to,
'buffer_size' : args.bufsize,
'learning_rate_anneal' : args.learning_rate_anneal,
'learning_rate_anneal_trigger' : args.learning_rate_anneal_trigger,
'learning_rate_anneal_interval' : args.learning_rate_anneal_interval,
# 'cfmap_loss' : args.cfmap_loss,
# 'lambda_cfmap' : args.lambda_cfmap,
})
trainer = training.Trainer(updater, (args.max_iter, 'iteration'), out=args.out)
model_save_interval = (args.snapshot_interval, 'iteration')
trainer.extend(extensions.snapshot_object(
gen_g, 'gen_g{.updater.iteration}.npz'), trigger=model_save_interval)
trainer.extend(extensions.snapshot_object(
gen_f, 'gen_f{.updater.iteration}.npz'), trigger=model_save_interval)
trainer.extend(extensions.snapshot_object(
dis_x, 'dis_x{.updater.iteration}.npz'), trigger=model_save_interval)
trainer.extend(extensions.snapshot_object(
dis_y, 'dis_y{.updater.iteration}.npz'), trigger=model_save_interval)
trainer.extend(extensions.snapshot(), trigger=model_save_interval)
log_keys = ['epoch', 'iteration', 'gen_g/loss_rec', 'gen_f/loss_rec', 'gen_g/loss_adv',
'gen_f/loss_adv', 'gen_g/loss_idt', 'gen_f/loss_idt', 'dis_x/loss', 'dis_y/loss']
# if args.cfmap_loss != None:
# log_keys += ['loss_cfmap_X', 'loss_cfmap_Y']
trainer.extend(extensions.LogReport(keys=log_keys, trigger=(20, 'iteration')))
trainer.extend(extensions.PrintReport(log_keys), trigger=(20, 'iteration'))
trainer.extend(extensions.ProgressBar(update_interval=50))
if extensions.PlotReport.available():
trainer.extend(
extensions.PlotReport(
['epoch', 'gen_g/loss_rec', 'gen_f/loss_rec', 'gen_g/loss_adv',
'gen_f/loss_adv', 'gen_g/loss_idt', 'gen_f/loss_idt', 'dis_x/loss', 'dis_y/loss'], 'iteration',
trigger=(100, 'iteration'), file_name='loss.png'))
if args.data_test_x:
eval_dataset = test_dataset
else:
eval_dataset = train_dataset
eval_interval = (args.eval_interval, 'iteration')
trainer.extend(
visualization(gen_g, gen_f, eval_dataset, os.path.join(args.out, 'preview'), 1),
trigger=eval_interval
)
if args.resume:
serializers.load_npz(args.resume, trainer)
trainer.run()
if __name__ == '__main__':
main()