-
Notifications
You must be signed in to change notification settings - Fork 0
/
loss_functions.py
158 lines (134 loc) · 5.9 KB
/
loss_functions.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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from IPython import embed
from inverse_warp import inverse_warp
def photometric_reconstruction_loss(tgt_img, ref_imgs, intrinsics, depth,
explainability_mask, pose, uncert, uncert_loss_epsilon=0.001, use_uncert=False,
rotation_mode='euler', padding_mode='zeros'):
def one_scale(depth, explainability_mask, uncertainity):
# embed()
assert(pose.size(1) == len(ref_imgs))
assert(explainability_mask is None or depth.size()[2:] == explainability_mask.size()[2:])
reconstruction_loss = 0
b, _, h, w = depth.size()
downscale = tgt_img.size(2)/h
# print("tgt_: ", tgt_img.size())
# print("ref_: ", ref_imgs[0].size())
# print("depth_: ", depth.size())
tgt_img_scaled = F.interpolate(tgt_img, (h, w), mode='area')
ref_imgs_scaled = [F.interpolate(ref_img, (h, w), mode='area') for ref_img in ref_imgs]
intrinsics_scaled = torch.cat((intrinsics[:, 0:2]/downscale, intrinsics[:, 2:]), dim=1)
# print("tgt: ", tgt_img_scaled.size())
# print("ref: ", ref_imgs_scaled[0].size())
# print("depth: ", depth.size())
warped_imgs = []
diff_maps = []
for i, ref_img in enumerate(ref_imgs_scaled):
current_pose = pose[:, i]
ref_img_warped, valid_points = inverse_warp(ref_img, depth[:,0], current_pose,
intrinsics_scaled, rotation_mode, padding_mode)
## valid_points is a tensor that contains bool values of size (4, 128, 416) to (4, 16, 52)
diff = (tgt_img_scaled - ref_img_warped) * valid_points.unsqueeze(1).float()
# embed()
if explainability_mask is not None:
diff = diff * explainability_mask[:,i:i+1].expand_as(diff) #???
reconstruction_loss += diff.abs().mean()
elif use_uncert:
diff = diff / (torch.log(1 + torch.exp(uncertainity + uncert_loss_epsilon)))
diff_mean = diff.abs().mean()
reg_term = torch.log(1 + torch.exp(uncertainity)).abs().mean()
reconstruction_loss += diff_mean + reg_term
# print("Index: ", i, " Errors: ", diff_mean, " Regularization term: ", reg_term)
else:
reconstruction_loss += diff.abs().mean()
# embed()
warped_imgs.append(ref_img_warped[0])
diff_maps.append(diff[0])
return reconstruction_loss, warped_imgs, diff_maps
warped_results, diff_results = [], []
if type(explainability_mask) not in [tuple, list]:
explainability_mask = [explainability_mask]
if type(depth) not in [list, tuple]:
depth = [depth]
if type(uncert) not in [list, tuple]:
uncert = [uncert]
# embed()
total_loss = 0
uncert_level_coeff = torch.ones(len(uncert))
for d, mask, u, u_coeff in zip(depth, explainability_mask, uncert, uncert_level_coeff):
loss, warped, diff = one_scale(d, mask, u)
if use_uncert:
total_loss += u_coeff * loss
else:
total_loss += loss
warped_results.append(warped)
diff_results.append(diff)
return total_loss, warped_results, diff_results
def explainability_loss(mask):
'''
Calculates Binary Cross Entropy Loss between a ones tensor and explainability mask
'''
if type(mask) not in [tuple, list]:
mask = [mask]
loss = 0
for mask_scaled in mask:
ones_var = torch.ones_like(mask_scaled)
loss += nn.functional.binary_cross_entropy(mask_scaled, ones_var)
return loss
def smooth_loss(pred_map):
'''
Smoothing loss, calculated using gradients in x and y direction, followed by
'''
def gradient(pred):
D_dy = pred[:, :, 1:] - pred[:, :, :-1]
D_dx = pred[:, :, :, 1:] - pred[:, :, :, :-1]
return D_dx, D_dy
if type(pred_map) not in [tuple, list]:
pred_map = [pred_map]
loss = 0
weight = 1.
for scaled_map in pred_map:
dx, dy = gradient(scaled_map)
dx2, dxdy = gradient(dx)
dydx, dy2 = gradient(dy)
loss += (dx2.abs().mean() + dxdy.abs().mean() + dydx.abs().mean() + dy2.abs().mean())*weight
weight /= 2.3
return loss
@torch.no_grad()
def compute_depth_errors(gt, pred, crop=True):
'''
Computes Relative Absolute Error and Relative Squared Error for Depth
'''
abs_rel, sq_rel = 0, 0
batch_size = gt.size(0)
if crop:
crop_mask = gt[0] != gt[0]
y1,y2 = int(0.40810811 * gt.size(1)), int(0.99189189 * gt.size(1))
x1,x2 = int(0.03594771 * gt.size(2)), int(0.96405229 * gt.size(2))
crop_mask[y1:y2,x1:x2] = 1
for current_gt, current_pred in zip(gt, pred):
valid = (current_gt > 0) & (current_gt < 80)
if crop:
valid = valid & crop_mask
if valid.sum() == 0:
continue
valid_gt = current_gt[valid]
valid_pred = current_pred[valid].clamp(1e-3, 80)
valid_pred = valid_pred * torch.median(valid_gt)/torch.median(valid_pred)
abs_rel += torch.mean(torch.abs(valid_gt - valid_pred) / valid_gt)
sq_rel += torch.mean(((valid_gt - valid_pred)**2) / valid_gt)
return [abs_rel.item()/batch_size , sq_rel.item()/batch_size]
@torch.no_grad()
def compute_pose_error(gt, pred):
'''
Computes Absolute Trajectory Error (ATE)
'''
for (current_gt, current_pred) in zip(gt, pred):
assert(current_gt.shape[0] == current_pred.shape[0])
seq_length = current_gt.shape[0]
scale_factor = torch.sum(current_gt[..., -1] * current_pred[..., -1]) / torch.sum(current_pred[..., -1] ** 2)
ATE = torch.norm((current_gt[..., -1] - scale_factor * current_pred[..., -1]).reshape(-1)).cpu().numpy()
return ATE/seq_length