-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathtrain.py
166 lines (143 loc) · 6.24 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
# -*- coding: utf-8 -*-
# @Date : 2019-07-25
# @Author : Xinyu Gong (xy_gong@tamu.edu)
# @Link : None
# @Version : 0.0
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import cfg
import models
import datasets
from functions import train, validate, LinearLrDecay, load_params, copy_params
from utils.utils import set_log_dir, save_checkpoint, create_logger
from utils.inception_score import _init_inception
from utils.fid_score import create_inception_graph, check_or_download_inception
import torch
import os
import numpy as np
import torch.nn as nn
from tensorboardX import SummaryWriter
from tqdm import tqdm
from copy import deepcopy
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
def main():
args = cfg.parse_args()
torch.cuda.manual_seed(args.random_seed)
# set tf env
_init_inception()
inception_path = check_or_download_inception(None)
create_inception_graph(inception_path)
# import network
gen_net = eval('models.'+args.model+'.Generator')(args=args).cuda()
dis_net = eval('models.'+args.model+'.Discriminator')(args=args).cuda()
# weight init
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv2d') != -1:
if args.init_type == 'normal':
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif args.init_type == 'orth':
nn.init.orthogonal_(m.weight.data)
elif args.init_type == 'xavier_uniform':
nn.init.xavier_uniform(m.weight.data, 1.)
else:
raise NotImplementedError('{} unknown inital type'.format(args.init_type))
elif classname.find('BatchNorm2d') != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0.0)
gen_net.apply(weights_init)
dis_net.apply(weights_init)
# set optimizer
gen_optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, gen_net.parameters()),
args.g_lr, (args.beta1, args.beta2))
dis_optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, dis_net.parameters()),
args.d_lr, (args.beta1, args.beta2))
gen_scheduler = LinearLrDecay(gen_optimizer, args.g_lr, 0.0, 0, args.max_iter * args.n_critic)
dis_scheduler = LinearLrDecay(dis_optimizer, args.d_lr, 0.0, 0, args.max_iter * args.n_critic)
# set up data_loader
dataset = datasets.ImageDataset(args)
train_loader = dataset.train
# fid stat
if args.dataset.lower() == 'cifar10':
fid_stat = 'fid_stat/fid_stats_cifar10_train.npz'
elif args.dataset.lower() == 'stl10':
fid_stat = 'fid_stat/stl10_train_unlabeled_fid_stats_48.npz'
else:
raise NotImplementedError(f'no fid stat for {args.dataset.lower()}')
assert os.path.exists(fid_stat)
# epoch number for dis_net
args.max_epoch = args.max_epoch * args.n_critic
if args.max_iter:
args.max_epoch = np.ceil(args.max_iter * args.n_critic / len(train_loader))
# initial
fixed_z = torch.cuda.FloatTensor(np.random.normal(0, 1, (25, args.latent_dim)))
gen_avg_param = copy_params(gen_net)
start_epoch = 0
best_fid = 1e4
# set writer
if args.load_path:
print(f'=> resuming from {args.load_path}')
assert os.path.exists(args.load_path)
checkpoint_file = os.path.join(args.load_path, 'Model', 'checkpoint.pth')
assert os.path.exists(checkpoint_file)
checkpoint = torch.load(checkpoint_file)
start_epoch = checkpoint['epoch']
best_fid = checkpoint['best_fid']
gen_net.load_state_dict(checkpoint['gen_state_dict'])
dis_net.load_state_dict(checkpoint['dis_state_dict'])
gen_optimizer.load_state_dict(checkpoint['gen_optimizer'])
dis_optimizer.load_state_dict(checkpoint['dis_optimizer'])
avg_gen_net = deepcopy(gen_net)
avg_gen_net.load_state_dict(checkpoint['avg_gen_state_dict'])
gen_avg_param = copy_params(avg_gen_net)
del avg_gen_net
args.path_helper = checkpoint['path_helper']
logger = create_logger(args.path_helper['log_path'])
logger.info(f'=> loaded checkpoint {checkpoint_file} (epoch {start_epoch})')
else:
# create new log dir
assert args.exp_name
args.path_helper = set_log_dir('logs', args.exp_name)
logger = create_logger(args.path_helper['log_path'])
logger.info(args)
writer_dict = {
'writer': SummaryWriter(args.path_helper['log_path']),
'train_global_steps': start_epoch * len(train_loader),
'valid_global_steps': start_epoch // args.val_freq,
}
# train loop
lr_schedulers = (gen_scheduler, dis_scheduler) if args.lr_decay else None
for epoch in tqdm(range(int(start_epoch), int(args.max_epoch)), desc='total progress'):
train(args, gen_net, dis_net, gen_optimizer, dis_optimizer, gen_avg_param, train_loader, epoch, writer_dict,
lr_schedulers)
if epoch and epoch % args.val_freq == 0 or epoch == int(args.max_epoch)-1:
backup_param = copy_params(gen_net)
load_params(gen_net, gen_avg_param)
inception_score, fid_score = validate(args, fixed_z, fid_stat, gen_net, writer_dict)
logger.info(f'Inception score: {inception_score}, FID score: {fid_score} || @ epoch {epoch}.')
load_params(gen_net, backup_param)
if fid_score < best_fid:
best_fid = fid_score
is_best = True
else:
is_best = False
else:
is_best = False
avg_gen_net = deepcopy(gen_net)
load_params(avg_gen_net, gen_avg_param)
save_checkpoint({
'epoch': epoch + 1,
'model': args.model,
'gen_state_dict': gen_net.state_dict(),
'dis_state_dict': dis_net.state_dict(),
'avg_gen_state_dict': avg_gen_net.state_dict(),
'gen_optimizer': gen_optimizer.state_dict(),
'dis_optimizer': dis_optimizer.state_dict(),
'best_fid': best_fid,
'path_helper': args.path_helper
}, is_best, args.path_helper['ckpt_path'])
del avg_gen_net
if __name__ == '__main__':
main()