-
Notifications
You must be signed in to change notification settings - Fork 17
/
Copy pathtrain.py
207 lines (166 loc) · 8.14 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
import os
from os.path import join
import argparse
from math import log10
import torch
import torch.nn as nn
import torch.optim as optim
import torch.backends.cudnn as cudnn
from data import DataLoaderHelper
from torch.utils.data import DataLoader
from torch.autograd import Variable
from model import G, D, weights_init
from util import load_image, save_image
from skimage.measure import compare_ssim as ssim
parser = argparse.ArgumentParser(description='DeepRendering-implemention')
parser.add_argument('--dataset', required=True, help='output from unity')
parser.add_argument('--train_batch_size', type=int, default=1, help='batch size for training')
parser.add_argument('--test_batch_size', type=int, default=1, help='batch size for testing')
parser.add_argument('--n_epoch', type=int, default=200, help='number of iterations')
parser.add_argument('--n_channel_input', type=int, default=3, help='number of input channels')
parser.add_argument('--n_channel_output', type=int, default=3, help='number of output channels')
parser.add_argument('--n_generator_filters', type=int, default=64, help='number of initial generator filters')
parser.add_argument('--n_discriminator_filters', type=int, default=64, help='number of initial discriminator filters')
parser.add_argument('--lr', type=float, default=0.0002, help='learning rate')
parser.add_argument('--beta1', type=float, default=0.5, help='beta1')
parser.add_argument('--cuda', action='store_true', help='cuda')
parser.add_argument('--resume_G', help='resume G')
parser.add_argument('--resume_D', help='resume D')
parser.add_argument('--workers', type=int, default=4, help='number of threads for data loader')
parser.add_argument('--seed', type=int, default=123, help='random seed')
parser.add_argument('--lamda', type=int, default=100, help='L1 regularization factor')
opt = parser.parse_args()
cudnn.benchmark = True
torch.cuda.manual_seed(opt.seed)
print('=> Loading datasets')
root_dir = "dataset/"
train_dir = join(root_dir + opt.dataset, "train")
test_dir = join(root_dir + opt.dataset, "val")
train_set = DataLoaderHelper(train_dir)
val_set = DataLoaderHelper(test_dir)
batch_size = opt.train_batch_size
n_epoch = opt.n_epoch
train_data = DataLoader(dataset=train_set, num_workers=opt.workers, batch_size=opt.train_batch_size, shuffle=True)
val_data = DataLoader(dataset=val_set, num_workers=opt.workers, batch_size=opt.test_batch_size, shuffle=False)
print('=> Building model')
netG = G(opt.n_channel_input*4, opt.n_channel_output, opt.n_generator_filters)
netG.apply(weights_init)
netD = D(opt.n_channel_input*4, opt.n_channel_output, opt.n_discriminator_filters)
netD.apply(weights_init)
criterion = nn.BCELoss()
criterion_l1 = nn.L1Loss()
albedo = torch.FloatTensor(opt.train_batch_size, opt.n_channel_input, 256, 256)
direct = torch.FloatTensor(opt.train_batch_size, opt.n_channel_input, 256, 256)
normal = torch.FloatTensor(opt.train_batch_size, opt.n_channel_input, 256, 256)
depth = torch.FloatTensor(opt.train_batch_size, opt.n_channel_input, 256, 256)
gt = torch.FloatTensor(opt.train_batch_size, opt.n_channel_output, 256, 256)
label = torch.FloatTensor(opt.train_batch_size)
real_label = 1
fake_label = 0
netD = netD.cuda()
netG = netG.cuda()
criterion = criterion.cuda()
criterion_l1 = criterion_l1.cuda()
albedo = albedo.cuda()
direct = direct.cuda()
normal = normal.cuda()
depth = depth.cuda()
gt = gt.cuda()
label = label.cuda()
albedo = Variable(albedo)
direct = Variable(direct)
normal = Variable(normal)
depth = Variable(depth)
gt = Variable(gt)
label = Variable(label)
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr, betas=(opt.beta1, 0.999))
optimizerG = optim.Adam(netG.parameters(), lr=opt.lr, betas=(opt.beta1, 0.999))
lastEpoch = 0
if opt.resume_G:
if os.path.isfile(opt.resume_G):
print("=> loading generator checkpoint '{}'".format(opt.resume_G))
checkpoint = torch.load(opt.resume_G)
lastEpoch = checkpoint['epoch']
n_epoch = n_epoch - lastEpoch
netG.load_state_dict(checkpoint['state_dict_G'])
optimizerG.load_state_dict(checkpoint['optimizer_G'])
print("=> loaded generator checkpoint '{}' (epoch {})".format(opt.resume_G, checkpoint['epoch']))
else:
print("=> no checkpoint found")
if opt.resume_D:
if os.path.isfile(opt.resume_D):
print("=> loading discriminator checkpoint '{}'".format(opt.resume_D))
checkpoint = torch.load(opt.resume_D)
netD.load_state_dict(checkpoint['state_dict_D'])
optimizerD.load_state_dict(checkpoint['optimizer_D'])
print("=> loaded discriminator checkpoint '{}'".format(opt.resume_D))
def train(epoch):
for (i, images) in enumerate(train_data):
netD.zero_grad()
(albedo_cpu, direct_cpu, normal_cpu, depth_cpu, gt_cpu) = (images[0], images[1], images[2], images[3], images[4])
albedo.data.resize_(albedo_cpu.size()).copy_(albedo_cpu)
direct.data.resize_(direct_cpu.size()).copy_(direct_cpu)
normal.data.resize_(normal_cpu.size()).copy_(normal_cpu)
depth.data.resize_(depth_cpu.size()).copy_(depth_cpu)
gt.data.resize_(gt_cpu.size()).copy_(gt_cpu)
output = netD(torch.cat((albedo, direct, normal, depth, gt), 1))
label.data.resize_(output.size()).fill_(real_label)
err_d_real = criterion(output, label)
err_d_real.backward()
d_x_y = output.data.mean()
fake_B = netG(torch.cat((albedo, direct, normal, depth), 1))
output = netD(torch.cat((albedo, direct, normal, depth, fake_B.detach()), 1))
label.data.resize_(output.size()).fill_(fake_label)
err_d_fake = criterion(output, label)
err_d_fake.backward()
d_x_gx = output.data.mean()
err_d = (err_d_real + err_d_fake) * 0.5
optimizerD.step()
netG.zero_grad()
output = netD(torch.cat((albedo, direct, normal, depth, fake_B), 1))
label.data.resize_(output.size()).fill_(real_label)
err_g = criterion(output, label) + opt.lamda \
* criterion_l1(fake_B, gt)
err_g.backward()
d_x_gx_2 = output.data.mean()
optimizerG.step()
print ('=> Epoch[{}]({}/{}): Loss_D: {:.4f} Loss_G: {:.4f} D(x): {:.4f} D(G(z)): {:.4f}/{:.4f}'.format(
epoch,
i,
len(train_data),
err_d.data[0],
err_g.data[0],
d_x_y,
d_x_gx,
d_x_gx_2,
))
def save_checkpoint(epoch):
if not os.path.exists("checkpoint"):
os.mkdir("checkpoint")
if not os.path.exists(os.path.join("checkpoint", opt.dataset)):
os.mkdir(os.path.join("checkpoint", opt.dataset))
net_g_model_out_path = "checkpoint/{}/netG_model_epoch_{}.pth".format(opt.dataset, epoch)
net_d_model_out_path = "checkpoint/{}/netD_model_epoch_{}.pth".format(opt.dataset, epoch)
torch.save({'epoch':epoch+1, 'state_dict_G': netG.state_dict(), 'optimizer_G':optimizerG.state_dict()}, net_g_model_out_path)
torch.save({'state_dict_D': netD.state_dict(), 'optimizer_D':optimizerD.state_dict()}, net_d_model_out_path)
print("Checkpoint saved to {}".format("checkpoint" + opt.dataset))
if not os.path.exists("validation"):
os.mkdir("validation")
if not os.path.exists(os.path.join("validation", opt.dataset)):
os.mkdir(os.path.join("validation", opt.dataset))
for index, images in enumerate(val_data):
(albedo_cpu, direct_cpu, normal_cpu, depth_cpu, gt_cpu) = (images[0], images[1], images[2], images[3], images[4])
albedo.data.resize_(albedo_cpu.size()).copy_(albedo_cpu)
direct.data.resize_(direct_cpu.size()).copy_(direct_cpu)
normal.data.resize_(normal_cpu.size()).copy_(normal_cpu)
depth.data.resize_(depth_cpu.size()).copy_(depth_cpu)
out = netG(torch.cat((albedo, direct, normal, depth), 1))
out = out.cpu()
out_img = out.data[0]
save_image(out_img,"validation/{}/{}_Fake.png".format(opt.dataset, index))
save_image(gt_cpu[0],"validation/{}/{}_Real.png".format(opt.dataset, index))
save_image(direct_cpu[0],"validation/{}/{}_Direct.png".format(opt.dataset, index))
for epoch in range(n_epoch):
train(epoch+lastEpoch)
if epoch % 1 == 0:
save_checkpoint(epoch+lastEpoch)