-
Notifications
You must be signed in to change notification settings - Fork 7
/
train_person_image.py
131 lines (123 loc) · 5.85 KB
/
train_person_image.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
from models.common.Discriminator import Discriminator
from models.common.VGG19 import Vgg19
from models.person.AdaATModule import AdaATModule
from utils import get_scheduler, update_learning_rate
from config.config import PersonTrainingOptions
from utils import GANLoss
from dataset.dataset_person import Data
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
import random
import numpy as np
import os
import torch.nn.functional as F
if __name__ == "__main__":
'''
training code of person image generation
'''
# load config
opt = PersonTrainingOptions().parse_args()
# set seed
random.seed(opt.seed)
np.random.seed(opt.seed)
torch.cuda.manual_seed(opt.seed)
# load training data in memory
train_data = Data(opt.train_data,opt.train_img_dir)
training_data_loader = DataLoader(dataset=train_data, batch_size=opt.batch_size, shuffle=True, drop_last=True)
train_data_length = len(training_data_loader)
# init network
net_g = AdaATModule(opt.img_channel, opt.keypoint_num).cuda()
net_d = Discriminator(opt.img_channel + opt.keypoint_num, opt.D_block_expansion, opt.D_num_blocks,
opt.D_max_features).cuda()
net_vgg = Vgg19().cuda()
# parallel
net_g = nn.DataParallel(net_g)
net_d = nn.DataParallel(net_d)
net_vgg = nn.DataParallel(net_vgg)
# set optimizer
optimizer_g = optim.Adam(net_g.parameters(), lr=opt.lr_g)
optimizer_d = optim.Adam(net_d.parameters(), lr=opt.lr_d)
# resume
if opt.resume != 'None':
print('loading checkpoint {}'.format(opt.resume_path))
checkpoint = torch.load(opt.resume_path)
opt.start_epoch = checkpoint['epoch']
net_g_static = checkpoint['state_dict']['net_g']
net_g.load_state_dict(net_g_static)
net_d.load_state_dict(checkpoint['state_dict']['net_d'])
optimizer_g.load_state_dict(checkpoint['optimizer']['net_g'])
optimizer_d.load_state_dict(checkpoint['optimizer']['net_d'])
# set criterion
criterionGAN = GANLoss().cuda()
criterionL1 = nn.L1Loss().cuda()
# set scheduler
net_g_scheduler = get_scheduler(optimizer_g, opt.non_decay, opt.decay)
net_d_scheduler = get_scheduler(optimizer_d, opt.non_decay, opt.decay)
# start train
for epoch in range(opt.start_epoch, opt.non_decay + opt.decay + 1):
net_g.train()
for iteration, data in enumerate(training_data_loader):
# read data
source_tensor, source_fitting_lm,target_tensor,target_fitting_lm = data
source_tensor = source_tensor.float().cuda()
source_fitting_lm = source_fitting_lm.float().cuda()
target_tensor = target_tensor.float().cuda()
target_fitting_lm = target_fitting_lm.float().cuda()
# network forward
fake_out, target_heatmap = net_g(source_tensor,source_fitting_lm,target_fitting_lm)
# down sample output image and real image
fake_out_half = F.avg_pool2d(fake_out, 3, 2, 1, count_include_pad=False)
target_tensor_half = F.interpolate(target_tensor, scale_factor=0.5, mode='bilinear')
# (1) Update D network
optimizer_d.zero_grad()
# compute fake loss
condition_fake_d = torch.cat([fake_out, target_heatmap], 1)
_,pred_fake_d = net_d(condition_fake_d)
loss_d_fake = criterionGAN(pred_fake_d, False)
# compute real loss
condition_real_d = torch.cat([target_tensor, target_heatmap], 1)
_,pred_real_d = net_d(condition_real_d)
loss_d_real = criterionGAN(pred_real_d, True)
# Combine D loss
loss_dI = (loss_d_fake + loss_d_real) * 0.5
loss_dI.backward(retain_graph=True)
optimizer_d.step()
# (2) Update G network
_, pred_fake_dI = net_d(condition_fake_d)
optimizer_g.zero_grad()
# compute perception loss
perception_real = net_vgg(target_tensor)
perception_fake = net_vgg(fake_out)
perception_real_half = net_vgg(target_tensor_half)
perception_fake_half = net_vgg(fake_out_half)
loss_g_perception = 0
for i in range(len(perception_real)):
loss_g_perception += criterionL1(perception_fake[i], perception_real[i])
loss_g_perception += criterionL1(perception_fake_half[i], perception_real_half[i])
loss_g_perception = (loss_g_perception / len(perception_real) ) * opt.lamb_perception
# gan dI loss
loss_g_dI = criterionGAN(pred_fake_dI, True)
# combine perception loss and gan loss
loss_g = loss_g_perception + loss_g_dI
loss_g.backward()
optimizer_g.step()
print(
"===> Epoch[{}]({}/{}): Loss_DI: {:.4f} Loss_GI: {:.4f} Loss_perception: {:.4f} lr_g = {:.7f} lr_d = {:.7f}".format(
epoch, iteration, len(training_data_loader), float(loss_dI), float(loss_g_dI),
float(loss_g_perception), optimizer_g.param_groups[0]['lr'], optimizer_d.param_groups[0]['lr']))
update_learning_rate(net_g_scheduler, optimizer_g)
update_learning_rate(net_d_scheduler, optimizer_d)
# checkpoint
if epoch % opt.checkpoint == 0:
if not os.path.exists(opt.result_path):
os.mkdir(opt.result_path)
model_out_path = os.path.join(opt.result_path, 'person_epoch_{}.pth'.format(epoch))
states = {
'epoch': epoch + 1,
'state_dict': {'net_g': net_g.state_dict(), 'net_d': net_d.state_dict()},
'optimizer': {'net_g': optimizer_g.state_dict(), 'net_d': optimizer_d.state_dict()}
}
torch.save(states, model_out_path)
print("Checkpoint saved to {}".format(epoch))