forked from febman/cp-vton
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
191 lines (153 loc) · 7.48 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
#coding=utf-8
import torch
import torch.nn as nn
import torch.nn.functional as F
import argparse
import os
import time
from cp_dataset import CPDataset, CPDataLoader
from networks import GMM, UnetGenerator, VGGLoss, load_checkpoint, save_checkpoint
from tensorboardX import SummaryWriter
from visualization import board_add_image, board_add_images
def get_opt():
parser = argparse.ArgumentParser()
parser.add_argument("--name", default = "GMM")
parser.add_argument("--gpu_ids", default = "")
parser.add_argument('-j', '--workers', type=int, default=1)
parser.add_argument('-b', '--batch-size', type=int, default=4)
parser.add_argument("--dataroot", default = "data")
parser.add_argument("--datamode", default = "train")
parser.add_argument("--stage", default = "GMM")
parser.add_argument("--data_list", default = "train_pairs.txt")
parser.add_argument("--fine_width", type=int, default = 192)
parser.add_argument("--fine_height", type=int, default = 256)
parser.add_argument("--radius", type=int, default = 5)
parser.add_argument("--grid_size", type=int, default = 5)
parser.add_argument('--lr', type=float, default=0.0001, help='initial learning rate for adam')
parser.add_argument('--tensorboard_dir', type=str, default='tensorboard', help='save tensorboard infos')
parser.add_argument('--checkpoint_dir', type=str, default='checkpoints', help='save checkpoint infos')
parser.add_argument('--checkpoint', type=str, default='', help='model checkpoint for initialization')
parser.add_argument("--display_count", type=int, default = 20)
parser.add_argument("--save_count", type=int, default = 100)
parser.add_argument("--keep_step", type=int, default = 100000)
parser.add_argument("--decay_step", type=int, default = 100000)
parser.add_argument("--shuffle", action='store_true', help='shuffle input data')
opt = parser.parse_args()
return opt
def train_gmm(opt, train_loader, model, board):
model.cuda()
model.train()
# criterion
criterionL1 = nn.L1Loss()
# optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=opt.lr, betas=(0.5, 0.999))
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda = lambda step: 1.0 -
max(0, step - opt.keep_step) / float(opt.decay_step + 1))
for step in range(opt.keep_step + opt.decay_step):
iter_start_time = time.time()
inputs = train_loader.next_batch()
im = inputs['image'].cuda()
im_pose = inputs['pose_image'].cuda()
im_h = inputs['head'].cuda()
shape = inputs['shape'].cuda()
agnostic = inputs['agnostic'].cuda()
c = inputs['cloth'].cuda()
cm = inputs['cloth_mask'].cuda()
im_c = inputs['parse_cloth'].cuda()
im_g = inputs['grid_image'].cuda()
grid, theta = model(agnostic, c)
warped_cloth = F.grid_sample(c, grid, padding_mode='border')
warped_mask = F.grid_sample(cm, grid, padding_mode='zeros')
warped_grid = F.grid_sample(im_g, grid, padding_mode='zeros')
visuals = [ [im_h, shape, im_pose],
[c, warped_cloth, im_c],
[warped_grid, (warped_cloth+im)*0.5, im]]
loss = criterionL1(warped_cloth, im_c)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (step+1) % opt.display_count == 0:
board_add_images(board, 'combine', visuals, step+1)
board.add_scalar('metric', loss.item(), step+1)
t = time.time() - iter_start_time
print('step: %8d, time: %.3f, loss: %4f' % (step+1, t, loss.item()), flush=True)
if (step+1) % opt.save_count == 0:
save_checkpoint(model, os.path.join(opt.checkpoint_dir, opt.name, 'step_%06d.pth' % (step+1)))
def train_tom(opt, train_loader, model, board):
model.cuda()
model.train()
# criterion
criterionL1 = nn.L1Loss()
criterionVGG = VGGLoss()
criterionMask = nn.L1Loss()
# optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=opt.lr, betas=(0.5, 0.999))
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda = lambda step: 1.0 -
max(0, step - opt.keep_step) / float(opt.decay_step + 1))
for step in range(opt.keep_step + opt.decay_step):
iter_start_time = time.time()
inputs = train_loader.next_batch()
im = inputs['image'].cuda()
im_pose = inputs['pose_image']
im_h = inputs['head']
shape = inputs['shape']
agnostic = inputs['agnostic'].cuda()
c = inputs['cloth'].cuda()
cm = inputs['cloth_mask'].cuda()
outputs = model(torch.cat([agnostic, c],1))
p_rendered, m_composite = torch.split(outputs, 3,1)
p_rendered = F.tanh(p_rendered)
m_composite = F.sigmoid(m_composite)
p_tryon = c * m_composite+ p_rendered * (1 - m_composite)
visuals = [ [im_h, shape, im_pose],
[c, cm*2-1, m_composite*2-1],
[p_rendered, p_tryon, im]]
loss_l1 = criterionL1(p_tryon, im)
loss_vgg = criterionVGG(p_tryon, im)
loss_mask = criterionMask(m_composite, cm)
loss = loss_l1 + loss_vgg + loss_mask
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (step+1) % opt.display_count == 0:
board_add_images(board, 'combine', visuals, step+1)
board.add_scalar('metric', loss.item(), step+1)
board.add_scalar('L1', loss_l1.item(), step+1)
board.add_scalar('VGG', loss_vgg.item(), step+1)
board.add_scalar('MaskL1', loss_mask.item(), step+1)
t = time.time() - iter_start_time
print('step: %8d, time: %.3f, loss: %.4f, l1: %.4f, vgg: %.4f, mask: %.4f'
% (step+1, t, loss.item(), loss_l1.item(),
loss_vgg.item(), loss_mask.item()), flush=True)
if (step+1) % opt.save_count == 0:
save_checkpoint(model, os.path.join(opt.checkpoint_dir, opt.name, 'step_%06d.pth' % (step+1)))
def main():
opt = get_opt()
print(opt)
print("Start to train stage: %s, named: %s!" % (opt.stage, opt.name))
# create dataset
train_dataset = CPDataset(opt)
# create dataloader
train_loader = CPDataLoader(opt, train_dataset)
# visualization
if not os.path.exists(opt.tensorboard_dir):
os.makedirs(opt.tensorboard_dir)
board = SummaryWriter(log_dir = os.path.join(opt.tensorboard_dir, opt.name))
# create model & train & save the final checkpoint
if opt.stage == 'GMM':
model = GMM(opt)
if not opt.checkpoint =='' and os.path.exists(opt.checkpoint):
load_checkpoint(model, opt.checkpoint)
train_gmm(opt, train_loader, model, board)
save_checkpoint(model, os.path.join(opt.checkpoint_dir, opt.name, 'gmm_final.pth'))
elif opt.stage == 'TOM':
model = UnetGenerator(25, 4, 6, ngf=64, norm_layer=nn.InstanceNorm2d)
if not opt.checkpoint =='' and os.path.exists(opt.checkpoint):
load_checkpoint(model, opt.checkpoint)
train_tom(opt, train_loader, model, board)
save_checkpoint(model, os.path.join(opt.checkpoint_dir, opt.name, 'tom_final.pth'))
else:
raise NotImplementedError('Model [%s] is not implemented' % opt.stage)
print('Finished training %s, nameed: %s!' % (opt.stage, opt.name))
if __name__ == "__main__":
main()