-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain.py
177 lines (153 loc) · 8.01 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
from torch.utils.data.dataloader import DataLoader
import lib.image.warping as warping
import torch
import torch.nn.functional as F
import numpy as np
import cv2
import matplotlib.pyplot as plt
from torch.utils.tensorboard import SummaryWriter
import kornia
import os
import time
from lib.image.warping import *
from config.config import parse_config
from lib.model.localtrans import LocalTrans
from lib.data.dataset_homo import HomoDataset
import matplotlib
matplotlib.use('Agg')
if __name__ == '__main__':
args = parse_config()
os.makedirs('results/tb_log/%s' % args.name, exist_ok=True)
os.makedirs('results/checkpoints/%s' % args.name, exist_ok=True)
writer = SummaryWriter('results/tb_log/%s' % args.name, flush_secs=10)
dataset = HomoDataset(args.dataroot, bias=args.random_bias,
downsample=args.downsample, random_color=args.random_color,
random_noise=args.random_noise, random_identity=False)
dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True, num_workers=4)
net1 = LocalTrans()
net2 = LocalTrans()
net3 = LocalTrans()
nets = [net1, net2, net3]
cuda = torch.device('cuda:%d' % args.gpu_id)
torch.cuda.set_device(args.gpu_id)
if args.resume or args.resume_dir is not None:
id = 0
if args.resume_dir is not None:
resume_dir = args.resume_dir
else:
resume_dir = os.path.join('results/checkpoints/%s' % args.name)
for net in nets:
if os.path.isfile(os.path.join(resume_dir, 'net%d_latest.pt' % id)):
print('load from %s' % (os.path.join(resume_dir, 'net%d_latest.pt' % id)))
net.load_state_dict(torch.load(os.path.join(resume_dir, 'net%d_latest.pt' % id)))
id += 1
for net in nets:
net.to(cuda)
lr = args.lr
optimizer1 = torch.optim.Adam(nets[0].parameters(), lr)
optimizer2 = torch.optim.Adam(nets[1].parameters(), lr)
optimizer3 = torch.optim.Adam(nets[2].parameters(), lr)
optimizers = [optimizer1, optimizer2, optimizer3]
epoch_idx = 0
train_idx = 0
if args.resume or args.resume_dir is not None:
id = 0
if args.resume_dir is not None:
resume_dir = args.resume_dir
else:
resume_dir = os.path.join('results/checkpoints/%s' % args.name)
for optim in optimizers:
if os.path.isfile(os.path.join(resume_dir, 'optim%d_latest.pt' % id)):
print('load from %s' % (os.path.join(resume_dir, 'optim%d_latest.pt' % id)))
optim.load_state_dict(torch.load(os.path.join(resume_dir, 'optim%d_latest.pt' % id)))
id += 1
if os.path.isfile(os.path.join(resume_dir, 'state.pt')):
state = torch.load(os.path.join(resume_dir, 'state.pt'))
epoch_idx = state['epoch_idx']
train_idx = state['train_idx']
EPOCH = args.epoch
training = True
for epoch in range(epoch_idx, EPOCH):
np.random.seed(int(time.time()))
for data in dataloader:
img1 = data["img1"].permute(0, 3, 1, 2).float().to(cuda)
o_img2 = data["img2"].permute(0, 3, 1, 2).float().to(cuda)
# imsave('show_dataset/img1.png', (img1.clamp(0,1)[:, :]*255).byte().cpu().permute(0, 2, 3, 1).numpy()[0, :, :, ::-1])
# imsave('show_dataset/img2.png', (o_img2.clamp(0,1)[:, :, 32:32+128, 32:32+128]*255).byte().cpu().permute(0, 2, 3, 1).numpy()[0, :, :, ::-1])
# exit(0)
img2 = o_img2.clone()
gt = data["gt"].reshape(-1, 2, 2, 2).to(cuda)
result = []
loss = 0
for level in range(3):
net = nets[level]
flow = net(img1[:, :, 32:32+128, 32:32+128], img2[:, :, 32:32+128, 32:32+128], level)
result.append(flow.clone())
loss_level = F.l1_loss(flow.permute(0, 2, 3, 1), gt)
writer.add_scalar('loss_%d' % level, loss_level.item())
loss += loss_level
if level != 2:
B = img1.shape[0]
grid = gen_grid(2, 2, 32, 128, 32, 128, B).to(cuda)
grid_flow = grid + flow.permute(0, 2, 3, 1)
homo = kornia.geometry.find_homography_dlt(grid_flow.reshape(B, -1, 2).contiguous(), grid.reshape(B, -1, 2).contiguous())
img2 = kornia.geometry.warp_perspective(img2, homo, (192, 192)).detach()
ori_grid = gen_grid(2, 2, 32, 128, 32, 128, B, device=cuda)
grid = ori_grid + gt
grid3 = torch.cat([grid.reshape(B, -1, 2), torch.ones(B, 4, 1, device=cuda)], dim=2)
grid3 = (homo @ grid3.transpose(1, 2)).transpose(1, 2)
grid2 = grid3[:, :, :2] / grid3[:, :, 2:]
gt = (grid2.reshape(B, 2, 2, 2) - ori_grid).detach()
print('level%d:' % level, loss_level)
for optimizer in optimizers:
optimizer.zero_grad()
loss.backward()
for optimizer in optimizers:
optimizer.step()
print(loss.item())
if train_idx % 50 == 0:
id = 0
checkpoint_dir = os.path.join('results/checkpoints/%s' % args.name)
for net in nets:
torch.save(net.state_dict(), os.path.join(checkpoint_dir, 'net%d_latest.pt' % (id)))
torch.save(net.state_dict(), os.path.join(checkpoint_dir, 'net%d_epoch_%d.pt' % (id, epoch)))
id += 1
id = 0
for optim in optimizers:
torch.save(optim.state_dict(), os.path.join(checkpoint_dir, 'optim%d_latest.pt' % (id)))
torch.save(optim.state_dict(), os.path.join(checkpoint_dir, 'optim%d_epoch_%d.pt' % (id, epoch)))
id += 1
state = {"epoch_idx": epoch, "train_idx": train_idx}
torch.save(state, os.path.join(checkpoint_dir, 'state.pt'))
img2 = o_img2.clone()[:1]
fig = plt.figure()
for level in range(3):
net = nets[level]
W, H = 192, 192
B = 1
flow = result[level][:1]
grid = gen_grid(2, 2, 32, 128, 32, 128, B).to(cuda)
grid_flow = grid + flow.permute(0, 2, 3, 1)
homo = kornia.geometry.find_homography_dlt(grid_flow.reshape(B, -1, 2).contiguous(), grid.reshape(B, -1, 2).contiguous())
img2 = kornia.geometry.warp_perspective(img2, homo, (192, 192))
plt.subplot(int('14%d' % (level+1)))
show_img1 = (img1.permute(0, 2, 3, 1)[0]).detach().cpu().numpy()[:, :, ::-1]
show_img2 = (img2.permute(0, 2, 3, 1)[0]).detach().cpu().numpy()[:, :, ::-1]
plt.imshow(show_img1 / 2 + show_img2 / 2)
# img2 = o_img2.clone()[:1]
# flow = gt.to(cuda)[:1].permute(0, 3, 1, 2)
# grid = gen_grid(2, 2, 32, 128, 32, 128, B).to(cuda)
# grid_flow = grid + flow.permute(0, 2, 3, 1)
# homo = kornia.geometry.find_homography_dlt(grid_flow.reshape(B, -1, 2).contiguous(), grid.reshape(B, -1, 2).contiguous())
# img2 = kornia.geometry.warp_perspective(img2, homo, (192, 192))
# plt.subplot(int('23%d' % (level+1+3)))
# show_img1 = (img1.permute(0, 2, 3, 1)[0]).detach().cpu().numpy()[:, :, ::-1]
# show_img2 = (img2.permute(0, 2, 3, 1)[0]).detach().cpu().numpy()[:, :, ::-1]
# plt.imshow(show_img1 / 2 + show_img2 / 2)
plt.savefig('test.jpg')
writer.add_figure("fig_epoch%d" % (epoch), fig, train_idx)
plt.close('all')
train_idx += 1
# f = open('overfit_wo.txt', 'w')
# for e in loss_list:
# f.write('%f\n' % e)