-
Notifications
You must be signed in to change notification settings - Fork 2
/
lossG2S.py
66 lines (52 loc) · 1.9 KB
/
lossG2S.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
# Copyright © NavInfo Europe 2021.
import torch
import math
class G2S(torch.nn.Module):
"""
GPS-to-Scale (G2S) loss
"""
def __init__(self, num_epochs, altitude_with_gps=True):
"""
Initializes G2S loss
Parameters
----------
num_epochs: Total number of epochs for which the network will be trained.
altitude_with_gps: if gps has altitude values.
"""
super(G2S, self).__init__()
self.num_epochs = num_epochs
self.altitude_with_gps = altitude_with_gps
def forward(self, gps, translation, epoch):
"""
Computes the g2s loss
Parameters
----------
gps: The input gps to the network loaded from the GPSDataloader class.
translation: The translation prediction made by the network of shape 12 x 3.
epoch: Current training epoch number indexed from 0.
"""
dis = {}
if self.altitude_with_gps:
dis["-1, 0"] = torch.norm(translation["-1, 0"], dim=1)
dis["0, 1"] = torch.norm(translation["0, 1"], dim=1)
else:
dis["-1, 0"] = torch.norm(translation["-1, 0"][:, [0, 2]], dim=1)
dis["0, 1"] = torch.norm(translation["0, 1"][:, [0, 2]], dim=1)
s1 = gps["-1, 0"].float() / dis["-1, 0"]
s2 = gps["0, 1"].float() / dis["0, 1"]
loss = torch.mean((s1 - 1) ** 2 + (s2 - 1) ** 2)
weight = self.inverse_exp_epoch_weight(epoch)
g2s_loss = weight * loss
scale = 0.5 * torch.mean(s1 + s2)
return {
'g2s_loss': g2s_loss,
'scale': scale
}
def inverse_exp_epoch_weight(self, epoch):
"""
Computes the dynamic weight for the G2S loss
Parameters
----------
epoch: The training epoch. The weight increases exponentially with the epochs.
"""
return math.exp(epoch - self.num_epochs + 1)