-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrainer.py
217 lines (178 loc) · 8.46 KB
/
trainer.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
208
209
210
211
212
213
214
215
216
217
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import time
import torch.optim as optim
from tqdm import tqdm
class Trainer(nn.Module):
def __init__(self, model, train_dataloader, multi_scale,
only_recon_epoch=15, dual_transform='flip'):
super().__init__()
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
self.model = model
self.train_dataloader = train_dataloader
self.multi_scale = multi_scale
self.only_recon_epoch = only_recon_epoch
self.dual_transform =dual_transform
self.to(self.device)
def init_optimzer(self, lr=0.0005, betas=(0.9, 0.999)):
self.lr = lr
self.optimizer = optim.Adam(self.model.parameters(), lr=lr, betas=betas)
def adjust_learning_rate(self, optimizer, epoch):
if epoch > 199:
self.lr = 0.00005
print('lr:', self.lr)
for param_group in optimizer.param_groups:
param_group['lr'] = self.lr
def warp(self, x, disp):
"""
warp an image/tensor (im2) back to im1, according to the optical flow
x: [B, C, H, W] (im2)
flo: [B, 2, H, W] flow
"""
B, C, H, W = x.size()
# mesh grid
xx = torch.arange(0, W).view(1, -1).repeat(H, 1)
yy = torch.arange(0, H).view(-1, 1).repeat(1, W)
xx = xx.view(1, 1, H, W).repeat(B, 1, 1, 1)
yy = yy.view(1, 1, H, W).repeat(B, 1, 1, 1)
grid = torch.cat((xx, yy), 1).float()
if x.is_cuda:
vgrid = grid.cuda()
else:
vgrid = grid
vgrid[:,:1,:,:] = vgrid[:,:1,:,:] - disp
# scale grid to [-1,1]
vgrid[:, 0, :, :] = 2.0 * vgrid[:, 0, :, :] / max(W - 1, 1) - 1.0
vgrid[:, 1, :, :] = 2.0 * vgrid[:, 1, :, :] / max(H - 1, 1) - 1.0
vgrid = vgrid.permute(0, 2, 3, 1)
output = F.grid_sample(x, vgrid, align_corners=True)
return output
def SSIM(self, x, y):
C1 = 0.01 ** 2
C2 = 0.03 ** 2
mu_x = F.avg_pool2d(x, 3, 1, 0)
mu_y = F.avg_pool2d(y, 3, 1, 0)
#(input, kernel, stride, padding)
sigma_x = F.avg_pool2d(x ** 2, 3, 1, 0) - mu_x ** 2
sigma_y = F.avg_pool2d(y ** 2, 3, 1, 0) - mu_y ** 2
sigma_xy = F.avg_pool2d(x * y , 3, 1, 0) - mu_x * mu_y
SSIM_n = (2 * mu_x * mu_y + C1) * (2 * sigma_xy + C2)
SSIM_d = (mu_x ** 2 + mu_y ** 2 + C1) * (sigma_x + sigma_y + C2)
SSIM = SSIM_n / SSIM_d
return torch.clamp((1 - SSIM) / 2, 0, 1)
def gradient(self, pred):
D_dy = pred[:, :, 1:, :] - pred[:, :, :-1, :]
D_dx = pred[:, :, :, 1:] - pred[:, :, :, :-1]
return D_dx, D_dy
def compute_grad2_smoothness_loss(self, flo, image, beta):
"""
Calculate the image-edge-aware second-order smoothness loss
"""
img_grad_x, img_grad_y = self.gradient(image)
weights_x = torch.exp(-10.0 * torch.mean(torch.abs(img_grad_x), 1, keepdim=True))
weights_y = torch.exp(-10.0 * torch.mean(torch.abs(img_grad_y), 1, keepdim=True))
dx, dy = self.gradient(flo)
dx2, dxdy = self.gradient(dx)
dydx, dy2 = self.gradient(dy)
return (torch.mean(beta*weights_x[:,:, :, 1:]*torch.abs(dx2)) + torch.mean(beta*weights_y[:, :, 1:, :]*torch.abs(dy2))) / 2.0
def reconstruction_loss(self, x, y):
ssim = torch.mean(self.SSIM(x, y))
l1 = torch.mean(torch.abs(x - y))
return 0.85*ssim + 0.15*l1
def set_loss_weight(self, smth_loss_weight=10, lr_loss_weight=0.01 ):
self.smth_loss_weight = smth_loss_weight
self.lr_loss_weight = lr_loss_weight
def compute_loss(self, imgL, imgR, epoch):
# estimate left disparity
if self.multi_scale:
l_disps = self.model(imgL, imgR)
else:
l_disps = self.model(imgL, imgR)
# estimate right disparity
if self.dual_transform == 'flip':
imgL_ast = imgL.flip(3)
imgR_ast = imgR.flip(3)
elif self.dual_transform == 'rotate':
imgL_ast = imgL.flip(2).flip(3)
imgR_ast = imgR.flip(2).flip(3)
if self.multi_scale:
r_disps_ast = self.model(imgR_ast, imgL_ast)
else:
r_disps_ast = self.model(imgR_ast, imgL_ast)
loss = 0
for l_disp,r_disp_ast in zip(l_disps, r_disps_ast):
# compute left reconstruction loss
recon_imgL = self.warp(imgR, l_disp)
left_recon_loss = self.reconstruction_loss(
recon_imgL[:,:,:, 75:575], imgL[:,:,:, 75:575])
# compte right reconstruction loss
recon_imgR_ast = self.warp(imgL_ast, r_disp_ast)
if self.dual_transform == 'flip':
recon_imgR = recon_imgR_ast.flip(3)
r_disp = r_disp_ast.flip(3)
elif self.dual_transform == 'rotate':
recon_imgR = recon_imgR_ast.flip(2).flip(3)
r_disp = r_disp_ast.flip(2).flip(3)
right_recon_loss = self.reconstruction_loss(
recon_imgR[:,:,:, 0:500], imgR[:,:,:, 0:500])
loss += left_recon_loss + right_recon_loss
# compute total loss
if epoch > self.only_recon_epoch:
# compute left disparity smoothness loss
left_smth_loss = self.compute_grad2_smoothness_loss(l_disp/20, imgL, 1.0)
# compute right disparity smoothness loss
right_smth_loss = self.compute_grad2_smoothness_loss(r_disp/20, imgR, 1.0)
# compute left-right consistency loss
r2l_disp = self.warp(r_disp, l_disp)
if self.dual_transform == 'flip':
l_disp_ast = l_disp.flip(3)
elif self.dual_transform == 'rotate':
l_disp_ast = l_disp.flip(2).flip(3)
l2r_disp_ast = self.warp(l_disp_ast, r_disp_ast)
if self.dual_transform == 'flip':
l2r_disp = l2r_disp_ast.flip(3)
elif self.dual_transform == 'rotate':
l2r_disp = l2r_disp_ast.flip(2).flip(3)
lr_right_loss = torch.mean(torch.abs(l_disp[:, :, :, 75:575] - r2l_disp[:, :, :, 75:575]))
lr_left_loss = torch.mean(torch.abs(r_disp[:, :, :, 0:500] - l2r_disp[:, :, :, 0:500]))
lr_loss = lr_left_loss + lr_right_loss
loss += self.smth_loss_weight * (left_smth_loss + right_smth_loss) + self.lr_loss_weight * lr_loss
return loss
def fit(self, start_epoch ,epochs, cpt_dir, save_freq=1):
print('start training')
start_full_time = time.time()
epoch_pbar = tqdm(range(start_epoch, epochs))
for epoch in epoch_pbar:
print(f'This is {epoch}-th epoch')
self.model.train()
total_train_loss = 0
self.adjust_learning_rate(self.optimizer, epoch)
## training ##
iter_pbar = tqdm(self.train_dataloader)
for batch_idx, (imgL_crop, imgR_crop) in enumerate(iter_pbar):
imgL_crop, imgR_crop = imgL_crop.to(self.device), imgR_crop.to(self.device)
self.optimizer.zero_grad()
loss = self.compute_loss(imgL_crop, imgR_crop, epoch)
loss.backward()
self.optimizer.step()
loss = loss.item()
iter_pbar.set_description('E:{}|loss{:.3f}'.format(epoch, loss))
total_train_loss += loss
total_train_loss /= len(self.train_dataloader)
epoch_pbar.set_description('E{}|loss{:.3f}'.format(epoch, total_train_loss))
## save checkpoint ##
if epoch % save_freq == 0:
os.makedirs(cpt_dir, exist_ok=True)
total_loss_str = str(total_train_loss).replace('.','_')[0:7]
cpt_path = os.path.join(cpt_dir, f'checkpoint_{epoch}_{total_loss_str}.cpt')
torch.save({
'epoch': epoch,
'state_dict': self.model.state_dict(),
'optimizer': self.optimizer.state_dict(),
'train_loss': total_train_loss,
}, cpt_path)
print(f'Checkpoint saved to {cpt_path}')
print('Done :)')
print('full training time: %.2f hours' % ((time.time() - start_full_time) / 3600))