-
Notifications
You must be signed in to change notification settings - Fork 7
/
geometry_optimizer.py
177 lines (139 loc) · 6.27 KB
/
geometry_optimizer.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
import argparse, sys, os, csv, time, datetime
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import torch.utils.data as data
import torch.nn.functional as F
import cv2 as cv
from tqdm import tqdm
from path import Path
from sequence_io import SequenceIO
from models import *
from loss import Loss
from models.layers import disp_to_depth
from warper import Warper, pose_vec2mat, inverse_pose
import warnings
warnings.filterwarnings('ignore')
class GeometryOptimizer:
def __init__(self, opt):
self.opt = opt
global device
device = torch.device(opt.cuda)
self.output_dir = Path(opt.output_dir)/opt.name
(self.output_dir/'depths').makedirs_p()
self.seq_io = SequenceIO(opt)
print('=> sequence length = {}'.format(len(self.seq_io)))
warper = Warper(opt, self.seq_io.get_intrinsic(True)).to(device)
# model
self.load_model()
# Loss
self.loss_function = Loss(opt, warper, self.seq_io)
def load_model(self):
opt = self.opt
input_channel = 3
output_channel = 1
dispnet_encoder = ResnetEncoder(18, True, input_channel).to(device)
dispnet_decoder = DepthDecoder(dispnet_encoder.num_ch_enc, opt.scales,
num_output_channels=output_channel,
h=opt.height, w=opt.width).to(device)
dispnet_encoder.train()
dispnet_decoder.train()
self.dispnet = {'encoder': dispnet_encoder, 'decoder': dispnet_decoder}
posenet_encoder = ResnetEncoder(18, True, input_channel).to(device)
posenet_decoder = PoseDecoder(posenet_encoder.num_ch_enc, 1, output_channel).to(device)
posenet_encoder.train()
posenet_decoder.train()
self.posenet = {'encoder': posenet_encoder, 'decoder': posenet_decoder}
self.optim_params = [
{'params': dispnet_encoder.parameters(), 'initial_lr': opt.learning_rate},
{'params': dispnet_decoder.parameters(), 'initial_lr': opt.learning_rate},
{'params': posenet_encoder.parameters(), 'initial_lr': opt.learning_rate},
{'params': posenet_decoder.parameters(), 'initial_lr': opt.learning_rate}
]
self.optimizer = optim.Adam(self.optim_params, betas=(0.9, 0.99))
def run(self):
opt = self.opt
print('=> optimize depths for each frame')
self.n_iter = 0
snippet_len = 1 + int(np.ceil((len(self.seq_io) - opt.batch_size) / (opt.batch_size - max(opt.intervals))))
self.pbar = tqdm(total=opt.init_num_epochs + opt.num_epochs*snippet_len)
# batch for initialization
self.batch_idx = -1
begin, end = self.get_batch_indices(self.batch_idx)
init_batch_items = self.load_batch(begin, end)
self.loss_function.preprocess_minibatch_weights(init_batch_items)
for ep in range(opt.init_num_epochs):
depths = self.optimize_snippet(init_batch_items)
if opt.num_epochs == 0: return
for self.batch_idx in range(snippet_len):
begin, end = self.get_batch_indices(self.batch_idx)
batch_items = self.load_batch(begin, end)
self.loss_function.preprocess_minibatch_weights(batch_items)
for ep in range(opt.num_epochs):
items = self.optimize_snippet(batch_items)
# save depth & pose results
if end == 0: end = None
self.save_results(items, begin, end)
self.pbar.close()
def get_batch_indices(self, batch_idx):
self.batch_idx = batch_idx
if batch_idx <= 0:
begin = 0
self.prefix = 0
else:
begin = self.end - max(self.opt.intervals)
self.prefix = max(self.opt.intervals)
end = min(begin + self.opt.batch_size, len(self.seq_io))
self.begin = begin
self.end = end
return begin, end
def load_batch(self, begin, end):
batch = self.seq_io.load_snippet(begin, end, load_flow=True)
for k in batch.keys():
batch[k] = batch[k].to(device)
return batch
def optimize_snippet(self, items):
h, w = self.opt.height, self.opt.width
d_features = self.dispnet['encoder'](items['imgs'][self.prefix:])
d_outputs = self.dispnet['decoder'](d_features)
depths = [d_outputs['disp', s] for s in self.opt.scales]
depths = [depth * self.opt.max_depth + self.opt.min_depth for depth in depths]
if self.prefix > 0:
depths = [torch.cat([self.fix_depths[s][-self.prefix:], depths[s]], 0) for s in self.opt.scales]
items['depths'] = depths
p_features = self.posenet['encoder'](items['imgs'][max(0, self.prefix-1):])
poses = self.posenet['decoder'](p_features)
poses = pose_vec2mat(poses, self.opt.rotation_mode)
poses = inverse_pose(poses[0].view(-1, 4, 4)).expand_as(poses) @ poses
try: poses = self.poses[-1].expand_as(poses).to(device) @ poses
except: pass
if self.prefix > 0:
poses = torch.cat([self.poses[-self.prefix:].to(device), poses[1:]], 0)
items['poses'] = poses
items['poses_inv'] = inverse_pose(items['poses'])
loss_items = self.loss_function(items)
self.optimizer.zero_grad()
loss_items['full'].backward()
self.optimizer.step()
self.n_iter += 1
self.pbar.update(1)
return items
def save_results(self, items, start_idx, end):
# depth
save_indices = list(range(start_idx, end))
self.fix_depths = [items['depths'][s][-self.prefix:].detach() for s in self.opt.scales]
self.seq_io.save_depths(items['depths'], save_indices)
# pose
poses = items['poses'].cpu().detach()
try: self.poses = torch.cat([self.poses[:-self.prefix], poses], dim=0)
except: self.poses = poses
self.seq_io.save_poses(self.poses)
# mask
err_masks = items['err_mask'].cpu().detach()
self.seq_io.save_errors(err_masks, save_indices[:items['err_mask'].size(0)])
if __name__ == '__main__':
import options
opt = options.Options().parse()
geo_optim = GeometryOptimizer(opt)
geo_optim.run()