-
Notifications
You must be signed in to change notification settings - Fork 7
/
Mutual_Regularization_Loss.py
181 lines (138 loc) · 8.32 KB
/
Mutual_Regularization_Loss.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
# -*- coding: utf-8 -*-
from __future__ import print_function
import time
import pprint
import torch
import os
import torch.nn as nn
from torch.autograd import Variable
from torch.nn.parameter import Parameter
from torch.nn.init import xavier_normal
import torch.nn.functional as F
import torch.optim as optim
# from torchvision import datasets, transforms
class FocalLoss(nn.Module):
def __init__(self,gamma=2.0,alpha=0.25, name='FocalLoss'):
super(FocalLoss, self).__init__()
self.gamma = gamma
self.alpha = alpha
def forward(self, y_pred, y_true):
gt = y_true
heat = y_pred
logpt = - F.binary_cross_entropy(heat, gt,reduction='none')
pt = torch.exp(logpt)
focal_loss = -( (1-pt)**self.gamma ) * logpt * (self.alpha*gt+(1-self.alpha)*(1-gt))
return focal_loss
class IntraConsistencyLoss(nn.Module):
#intracons_loss = self.intracons_loss_func((a_heat_gt, s_heat_gt, e_heat_gt), (a_heat, s_heat, e_heat))
def __init__(self, alpha, name='IntraConsistencyLoss'):
super(IntraConsistencyLoss, self).__init__()
self.alpha = alpha
def forward(self, y_true, y_pred):
a_gt, s_gt, e_gt = y_true
a_heat, s_heat, e_heat = y_pred
def _intra_consistency_loss(heat, gt):
# heat_fea -> [N,window,1]
# gt -> [N,window,1]
# mask -> [N,window,1]
# a = np.array([[0,1,1,1,1,0,0,0]]) shape: (1,8)
# matmul(a.T, a)
# Out:
# [0, 0, 0, 0, 0, 0, 0, 0] [0, 0, 0, 0, 0, 0, 0, 0] [0, 1, 1, 1, 1, 0, 0, 0] [0, 0, 0, 0, 0, 1, 1, 1]
# [0, 1, 1, 1, 1, 0, 0, 0] [0, 0, 1, 1, 1, 0, 0, 0] [1, 0, 0, 0, 0, 1, 1, 1] [0, 0, 0, 0, 0, 0, 0, 0]
# [0, 1, 1, 1, 1, 0, 0, 0] [0, 1, 0, 1, 1, 0, 0, 0] [1, 0, 0, 0, 0, 1, 1, 1] [0, 0, 0, 0, 0, 0, 0, 0]
# [0, 1, 1, 1, 1, 0, 0, 0] ----> [0, 1, 1, 0, 1, 0, 0, 0] [1, 0, 0, 0, 0, 1, 1, 1] [0, 0, 0, 0, 0, 0, 0, 0]
# [0, 1, 1, 1, 1, 0, 0, 0] [0, 1, 1, 1, 0, 0, 0, 0] & [1, 0, 0, 0, 0, 1, 1, 1] [0, 0, 0, 0, 0, 0, 0, 0]
# [0, 0, 0, 0, 0, 0, 0, 0] [0, 0, 0, 0, 0, 0, 0, 0] [0, 1, 1, 1, 1, 0, 0, 0] [1, 0, 0, 0, 0, 0, 1, 1]
# [0, 0, 0, 0, 0, 0, 0, 0] [0, 0, 0, 0, 0, 0, 0, 0] [0, 1, 1, 1, 1, 0, 0, 0] [1, 0, 0, 0, 0, 1, 0, 1]
# [0, 0, 0, 0, 0, 0, 0, 0] [0, 0, 0, 0, 0, 0, 0, 0] M_gt_1 [0, 1, 1, 1, 1, 0, 0, 0] M_gt_2 [1, 0, 0, 0, 0, 1, 1, 0] M_gt_3
N = gt.shape[0]
window = gt.shape[1]
distance = (heat.unsqueeze(-1)-heat.unsqueeze(1)).abs() # [N,window,window]
# gt -> [N,window,1]
gt_1 = gt.unsqueeze(-1).unsqueeze(1) # [N,1,window,1]
gt_2 = gt.unsqueeze(1).unsqueeze(1) # [N,1,1,window]
eye_mat = torch.diag_embed(torch.ones_like(gt))
M_gt_1 = F.relu(torch.matmul(gt_1, gt_2) - eye_mat.unsqueeze(1)) # [N,1,window,window]
M_gt_2 = (gt_1 - gt_2) * (gt_1 - gt_2) # [N,1,window,window]
M_gt_3 = torch.ones_like(M_gt_1) - eye_mat.unsqueeze(1) - M_gt_1 - M_gt_2 # [N,1,window,window]
pairs_1 = M_gt_1.sum(dim=(1,2,3))+1 # [N]
pairs_2 = M_gt_2.sum(dim=(1,2,3)) + 1 # [N]
pairs_3 = M_gt_3.sum(dim=(1,2,3)) + 1 # [N]
consistency_1 = (distance.unsqueeze(1) * M_gt_1).sum(dim=(1,2,3)) / pairs_1 # [N]
consistency_2 = 1 - (distance.unsqueeze(1) * M_gt_2).sum(dim=(1,2,3)) / pairs_2 # [N]
consistency_3 = (distance.unsqueeze(1) * M_gt_3).sum(dim=(1,2,3)) / pairs_3 # [N]
consistency_loss = consistency_1 + consistency_2 + consistency_3
return consistency_loss
intra_loss_action = _intra_consistency_loss(a_heat, a_gt)
intra_loss_start = _intra_consistency_loss(s_heat, s_gt)
intra_loss_end = _intra_consistency_loss(e_heat, e_gt)
intra_loss = self.alpha * (intra_loss_action + intra_loss_start + intra_loss_end)
return intra_loss
class InterConsistencyLoss(nn.Module):
def __init__(self, alpha, name='InterConsistencyLoss'):
super(InterConsistencyLoss, self).__init__()
self.alpha = alpha
def forward(self, y_true, y_pred):
a_gt, s_gt, e_gt = y_true
a_heat, s_heat, e_heat = y_pred
def _inter_consistency_loss(action_heat, start_heat, end_heat):
# action_heat -> [N,window]
# start_heat -> [N,window]
# end_heat -> [N,window]
diff = torch.cat([action_heat[:,1:,]-action_heat[:,:-1,], action_heat[:,-1:,]-action_heat[:,-2:-1,]], 1) # [N,window]
# diff = diff / tf.reduce_max(diff, [1,2]) # [N,window,1]
zeros_tmp = torch.zeros_like(diff)
diff_1 = torch.where(diff>=0, diff,zeros_tmp) #tf.where(tf.greater_equal(diff, 0), diff, tf.zeros_like(diff))
diff_0 = - torch.where(diff<=0, diff,zeros_tmp) #tf.where(tf.less_equal(diff, 0), diff, tf.zeros_like(diff))
# start_heat = start_heat / tf.reduce_max(start_heat, [1,2])
# end_heat = end_heat / tf.reduce_max(end_heat, [1,2])
start_diff_consistency = (diff_1 - start_heat).abs()
end_diff_consistency = (diff_0 - end_heat).abs()
consistency_loss = end_diff_consistency + start_diff_consistency
return consistency_loss
inter_loss = self.alpha * (_inter_consistency_loss(a_heat, s_heat, e_heat))
return inter_loss
class Creatloss_base(nn.Module):
def __init__(self, args):
super(Creatloss_base, self).__init__()
self.args=args
self.clas_loss_func = nn.BCELoss(reduction='none')
#self.clas_loss_func = FocalLoss()
self.regr_loss_func = nn.SmoothL1Loss(reduction='none')
self.intracons_loss_func = IntraConsistencyLoss(alpha=self.args.inner_alpha)
self.intercons_loss_func = InterConsistencyLoss(alpha=self.args.external_alpha)
def clas_reweight(self,ori_loss,heat_gt,mask):
nmask = mask - heat_gt
pos_num = heat_gt.sum(1)
neg_num = nmask.sum(1)
pos_loss = (ori_loss * heat_gt).sum(1)/ (pos_num+1e-7)
neg_loss = (ori_loss * nmask).sum(1)/ (neg_num+1e-7)
reweight_loss = 0.5 * (pos_loss + neg_loss)
return reweight_loss #size: batch_size
def forward(self, action_heat,start_heat,end_heat,start_bias,end_bias,target):
a_heat, s_heat, e_heat, s_bias, e_bias = action_heat,start_heat,end_heat,start_bias,end_bias
a_heat_gt = target[:,0,:]
s_heat_gt = target[:,1,:]
e_heat_gt = target[:,2,:]
s_bias_gt = target[:,3,:]
e_bias_gt = target[:,4,:]
mask = target[:,5,:]
a_clas_loss = self.clas_loss_func(a_heat, a_heat_gt)
s_clas_loss = self.clas_loss_func(s_heat, s_heat_gt)
e_clas_loss = self.clas_loss_func(e_heat, e_heat_gt)
#clas_loss = (a_clas_loss + s_clas_loss + e_clas_loss) * mask
# 使用reweighting过的分类loss
clas_loss = self.clas_reweight(a_clas_loss,a_heat_gt,mask) \
+ self.clas_reweight(s_clas_loss,s_heat_gt,mask) \
+ self.clas_reweight(e_clas_loss,e_heat_gt,mask)
s_regr_loss = self.regr_loss_func(s_bias, s_bias_gt) * s_heat_gt
e_regr_loss = self.regr_loss_func(e_bias, e_bias_gt) * e_heat_gt
#regr_loss = s_regr_loss.sum(1)/s_heat_gt.sum(1) + e_regr_loss.sum(1)/e_heat_gt.sum(1)
regr_loss = (s_regr_loss + e_regr_loss) * mask
intracons_loss = self.intracons_loss_func((a_heat_gt, s_heat_gt, e_heat_gt), (a_heat, s_heat, e_heat))
intercons_loss = self.intercons_loss_func((a_heat_gt, s_heat_gt, e_heat_gt), (a_heat, s_heat, e_heat)).mean(1)
cons_loss = intracons_loss + intercons_loss
#loss = clas_loss.mean(1) + 0.05 * regr_loss + cons_loss
loss = clas_loss + regr_loss.mean(1) + cons_loss
return loss, a_clas_loss, s_clas_loss, e_clas_loss, regr_loss, intracons_loss, intercons_loss