forked from lelimite4444/BridgeDepthFlow
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
194 lines (162 loc) · 11.4 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
import torch
import torchvision
import torchvision.transforms as transforms
import torch.utils.data as data
import numpy as np
from torch.autograd import Variable
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import argparse
import random
from PIL import Image
import matplotlib.pyplot as plt
import cv2
from models.MonodepthModel import *
from models.PWC_net import *
from models.PWC_net import PWCDCNet
from utils.scene_dataloader import *
from utils.utils import *
from models.networks.submodules import *
from models.networks.resample2d_package.resample2d import Resample2d
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument('--model_name', type=str, help='model name', default='monodepth')
parser.add_argument('--data_path', type=str, help='path to the data', required=True)
parser.add_argument('--filenames_file', type=str, help='path to the filenames text file', required=True)
parser.add_argument('--input_height', type=int, help='input height', default=256)
parser.add_argument('--input_width', type=int, help='input width', default=512)
parser.add_argument('--batch_size', type=int, help='batch size', default=2)
parser.add_argument('--num_epochs', type=int, help='number of epochs', default=80)
parser.add_argument('--learning_rate', type=float, help='initial learning rate', default=1e-4)
parser.add_argument('--lr_loss_weight', type=float, help='left-right consistency weight', default=0.5)
parser.add_argument('--alpha_image_loss', type=float, help='weight between SSIM and L1 in the image loss', default=0.85)
parser.add_argument('--disp_gradient_loss_weight', type=float, help='disparity smoothness weigth', default=0.1)
parser.add_argument('--num_threads', type=int, help='number of threads to use for data loading', default=8)
parser.add_argument('--checkpoint_path', type=str, help='path to a specific checkpoint to load', default='')
parser.add_argument('--type_of_2warp', type=int, help='2warp type', default=0)
args = parser.parse_args()
return args
args = get_args()
if args.model_name == 'monodepth':
net = MonodepthNet().cuda()
elif args.model_name == 'pwc':
net = pwc_dc_net().cuda()
args.input_width = 832
left_image_1, left_image_2, right_image_1, right_image_2 = get_kitti_cycle_data(args.filenames_file, args.data_path)
CycleLoader = torch.utils.data.DataLoader(
myCycleImageFolder(left_image_1, left_image_2, right_image_1, right_image_2, True, args),
batch_size = args.batch_size, shuffle = True, num_workers = args.num_threads, drop_last = False)
optimizer = optim.Adam(net.parameters(), lr = args.learning_rate)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[4, 7, 10, 13], gamma=0.5)
for epoch in range(args.num_epochs):
scheduler.step()
for batch_idx, (left_image_1, left_image_2, right_image_1, right_image_2) in enumerate(CycleLoader, 0):
optimizer.zero_grad()
former = torch.cat((left_image_2, left_image_1, right_image_1, left_image_1), 0)
latter = torch.cat((right_image_2, left_image_2, right_image_2, right_image_1), 0)
left_pyramid = make_pyramid(former, 4)
right_pyramid = make_pyramid(latter, 4)
model_input = Variable(torch.cat((former, latter), 1).cuda())
model_input_2 = Variable(torch.cat((latter, former), 1).cuda())
if args.model_name == 'monodepth':
disp_est_scale, disp_est = net(model_input)
disp_est_scale_2, disp_est_2 = net(model_input_2)
elif args.model_name == 'pwc':
disp_est_scale = net(model_input)
disp_est = [torch.cat((disp_est_scale[i][:,0,:,:].unsqueeze(1) / disp_est_scale[i].shape[3],
disp_est_scale[i][:,1,:,:].unsqueeze(1) / disp_est_scale[i].shape[2]), 1) for i in range(4)]
disp_est_scale_2 = net(model_input_2)
disp_est_2 = [torch.cat((disp_est_scale_2[i][:,0,:,:].unsqueeze(1) / disp_est_scale_2[i].shape[3],
disp_est_scale_2[i][:,1,:,:].unsqueeze(1) / disp_est_scale_2[i].shape[2]), 1) for i in range(4)]
border_mask = [create_border_mask(left_pyramid[i], 0.1) for i in range(4)]
fw_mask = []
bw_mask = []
for i in range(4):
fw, bw, diff_fw, diff_bw = get_mask(disp_est_scale[i], disp_est_scale_2[i], border_mask[i])
fw += 1e-3
bw += 1e-3
fw[[0,1,6,7]] = fw[[0,1,6,7]] * 0 + 1
bw[[0,1,6,7]] = bw[[0,1,6,7]] * 0 + 1
fw_detached = fw.clone().detach()
bw_detached = bw.clone().detach()
fw_mask.append(fw_detached)
bw_mask.append(bw_detached)
#reconstruction from right to left
left_est = [Resample2d()(right_pyramid[i], disp_est_scale[i]) for i in range(4)]
l1_left = [torch.abs(left_est[i] - left_pyramid[i]) * fw_mask[i] for i in range(4)]
l1_reconstruction_loss_left = [torch.mean(l1_left[i]) / torch.mean(fw_mask[i]) for i in range(4)]
ssim_left = [SSIM(left_est[i] * fw_mask[i], left_pyramid[i] * fw_mask[i]) for i in range(4)]
ssim_loss_left = [torch.mean(ssim_left[i]) / torch.mean(fw_mask[i]) for i in range(4)]
image_loss_left = [args.alpha_image_loss * ssim_loss_left[i] +
(1 - args.alpha_image_loss) * l1_reconstruction_loss_left[i] for i in range(4)]
image_loss = image_loss_left[0] + image_loss_left[1] + image_loss_left[2] + image_loss_left[3]
disp_loss = [cal_grad2_error(disp_est_scale[i] / 20, left_pyramid[i], 1.0) for i in range(4)]
disp_gradient_loss = disp_loss[0] + disp_loss[1] + disp_loss[2] + disp_loss[3]
#reconstruction from left to right
right_est = [Resample2d()(left_pyramid[i], disp_est_scale_2[i]) for i in range(4)]
l1_right = [torch.abs(right_est[i] - right_pyramid[i]) * bw_mask[i] for i in range(4)]
l1_reconstruction_loss_right = [torch.mean(l1_right[i]) / torch.mean(bw_mask[i]) for i in range(4)]
ssim_right = [SSIM(right_est[i] * bw_mask[i], right_pyramid[i] * bw_mask[i]) for i in range(4)]
ssim_loss_right = [torch.mean(ssim_right[i]) / torch.mean(bw_mask[i]) for i in range(4)]
image_loss_right = [args.alpha_image_loss * ssim_loss_right[i] +
(1 - args.alpha_image_loss) * l1_reconstruction_loss_right[i] for i in range(4)]
image_loss_2 = image_loss_right[0] + image_loss_right[1] + image_loss_right[2] + image_loss_right[3]
disp_loss_2 = [cal_grad2_error(disp_est_scale_2[i] / 20, right_pyramid[i], 1.0) for i in range(4)]
disp_gradient_loss_2 = disp_loss_2[0] + disp_loss_2[1] + disp_loss_2[2] + disp_loss_2[3]
#LR consistency
right_to_left_disp = [- Resample2d()(disp_est_2[i], disp_est_scale[i]) for i in range(4)]
left_to_right_disp = [- Resample2d()(disp_est[i], disp_est_scale_2[i]) for i in range(4)]
lr_left_loss = [torch.mean(torch.abs(right_to_left_disp[i][[0,1,6,7]] - disp_est[i][[0,1,6,7]])) for i in range(4)]
lr_right_loss = [torch.mean(torch.abs(left_to_right_disp[i][[0,1,6,7]] - disp_est_2[i][[0,1,6,7]])) for i in range(4)]
lr_loss = sum(lr_left_loss + lr_right_loss)
loss = image_loss + image_loss_2 + 10 * (disp_gradient_loss + disp_gradient_loss_2) + args.lr_loss_weight * lr_loss
"""
##########################################################################################
# #
# batch 7,8 mask for the direction of the reconstruction #
# forward L_t ------------> R_t #
# | | mask : L_t+1 ---> L_t ---> R_t #
# 3,4 | | 5,6 mask_2 : L_t+1 ---> R_t+1 ---> R_t #
# | | mask_3 : R_t+1 ---> R_t ---> L_t #
# v v mask_4 : R_t+1 ---> L_t+1 ---> L_t #
# L_t+1 ----------> R_t+1 mask_5 : R_t ---> L_t ---> L_t+1 #
# 1,2 #
# #
##########################################################################################
"""
if args.type_of_2warp == 1:
mask_4 = [fw_mask[i][[2,3]] for i in range(4)]
warp2_est_4 = [Resample2d()(left_est[i][[0,1]], disp_est_scale[i][[2,3]]) for i in range(4)]
loss += 0.1 * sum([warp_2(warp2_est_4[i], left_pyramid[i][[6,7]], mask_4[i], args) for i in range(4)])
mask_5 = [bw_mask[i][[2,3]] for i in range(4)]
warp2_est_5 = [Resample2d()(left_est[i][[6,7]], disp_est_scale_2[i][[2,3]]) for i in range(4)]
loss += 0.1 * sum([warp_2(warp2_est_5[i], left_pyramid[i][[0,1]], mask_5[i], args) for i in range(4)])
elif args.type_of_2warp == 2:
mask = [Resample2d()(fw_mask[i][[2,3]], disp_est_scale_2[i][[0,1]]) for i in range(4)]
warp2_est = [Resample2d()(left_est[i][[2,3]], disp_est_scale_2[i][[6,7]]) for i in range(4)]
loss += 0.1 * sum([warp_2(warp2_est[i], right_pyramid[i][[6,7]], mask[i], args) for i in range(4)])
mask_3 = [Resample2d()(fw_mask[i][[4,5]], disp_est_scale[i][[0,1]]) for i in range(4)]
warp2_est_3 = [Resample2d()(left_est[i][[4,5]], disp_est_scale[i][[6,7]]) for i in range(4)]
loss += 0.1 * sum([warp_2(warp2_est_3[i], left_pyramid[i][[6,7]], mask_3[i], args) for i in range(4)])
elif args.type_of_2warp == 3:
mask = [Resample2d()(fw_mask[i][[2,3]], disp_est_scale_2[i][[0,1]]) for i in range(4)]
warp2_est = [Resample2d()(left_est[i][[2,3]], disp_est_scale_2[i][[6,7]]) for i in range(4)]
loss += 0.1 * sum([warp_2(warp2_est[i], right_pyramid[i][[6,7]], mask[i], args) for i in range(4)])
mask_2 = [fw_mask[i][[4,5]] for i in range(4)]
warp2_est_2 = [Resample2d()(right_est[i][[0,1]], disp_est_scale[i][[4,5]]) for i in range(4)]
loss += 0.1 * sum([warp_2(warp2_est_2[i], right_pyramid[i][[6,7]], mask_2[i], args) for i in range(4)])
loss.backward()
if args.model_name == 'monodepth':
print("Epoch :", epoch)
print("Batch Index :", batch_idx)
print(net.conv1.weight.grad[0,0,0,0])
elif args.model_name == 'pwc':
print("Epoch :", epoch)
print("Batch Index :", batch_idx)
print(net.conv1a[0].weight.grad[0,0,0,0])
optimizer.step()
if epoch % 1 == 0:
state = {'epoch': epoch, 'state_dict': net.state_dict(), 'optimizer': optimizer.state_dict(), 'scheduler': scheduler}
torch.save(state, args.checkpoint_path + "model_epoch" + str(epoch))
print("The model of epoch ", epoch, "has been saved.")