-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdiscriminative.py
181 lines (130 loc) · 5.76 KB
/
discriminative.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
from torch.nn.modules.loss import _Loss
from torch.autograd import Variable
import torch
import numpy as np
def calculate_means(pred, gt, n_objects, max_n_objects, usegpu):
"""pred: bs, height * width, n_filters
gt: bs, height * width, n_instances"""
bs, n_loc, n_filters = pred.size()
n_instances = gt.size(2)
pred_repeated = pred.unsqueeze(2).expand(
bs, n_loc, n_instances, n_filters) # bs, n_loc, n_instances, n_filters
# bs, n_loc, n_instances, 1
gt_expanded = gt.unsqueeze(3)
pred_masked = pred_repeated * gt_expanded
means = []
for i in range(bs):
_n_objects_sample = n_objects[i]
# n_loc, n_objects, n_filters
_pred_masked_sample = pred_masked[i, :, : _n_objects_sample]
# n_loc, n_objects, 1
_gt_expanded_sample = gt_expanded[i, :, : _n_objects_sample]
_mean_sample = _pred_masked_sample.sum(
0) / _gt_expanded_sample.sum(0) # n_objects, n_filters
if (max_n_objects - _n_objects_sample) != 0:
n_fill_objects = int(max_n_objects - _n_objects_sample)
_fill_sample = torch.zeros(n_fill_objects, n_filters)
if usegpu:
_fill_sample = _fill_sample.cuda()
_fill_sample = Variable(_fill_sample)
_mean_sample = torch.cat((_mean_sample, _fill_sample), dim=0)
means.append(_mean_sample)
means = torch.stack(means)
# means = pred_masked.sum(1) / gt_expanded.sum(1)
# # bs, n_instances, n_filters
return means
def calculate_variance_term(pred, gt, means, n_objects, delta_v, norm=2):
"""pred: bs, height * width, n_filters
gt: bs, height * width, n_instances
means: bs, n_instances, n_filters"""
bs, n_loc, n_filters = pred.size()
n_instances = gt.size(2)
# bs, n_loc, n_instances, n_filters
means = means.unsqueeze(1).expand(bs, n_loc, n_instances, n_filters)
# bs, n_loc, n_instances, n_filters
pred = pred.unsqueeze(2).expand(bs, n_loc, n_instances, n_filters)
# bs, n_loc, n_instances, n_filters
gt = gt.unsqueeze(3).expand(bs, n_loc, n_instances, n_filters)
_var = (torch.clamp(torch.norm((pred - means), norm, 3) -
delta_v, min=0.0) ** 2) * gt[:, :, :, 0]
var_term = 0.0
for i in range(bs):
_var_sample = _var[i, :, :n_objects[i]] # n_loc, n_objects
_gt_sample = gt[i, :, :n_objects[i], 0] # n_loc, n_objects
var_term += torch.sum(_var_sample) / torch.sum(_gt_sample)
var_term = var_term / bs
return var_term
def calculate_distance_term(means, n_objects, delta_d, norm=2, usegpu=True):
"""means: bs, n_instances, n_filters"""
bs, n_instances, n_filters = means.size()
dist_term = 0.0
for i in range(bs):
_n_objects_sample = int(n_objects[i])
if _n_objects_sample <= 1:
continue
_mean_sample = means[i, : _n_objects_sample, :] # n_objects, n_filters
means_1 = _mean_sample.unsqueeze(1).expand(
_n_objects_sample, _n_objects_sample, n_filters)
means_2 = means_1.permute(1, 0, 2)
diff = means_1 - means_2 # n_objects, n_objects, n_filters
_norm = torch.norm(diff, norm, 2)
margin = 2 * delta_d * (1.0 - torch.eye(_n_objects_sample))
if usegpu:
margin = margin.cuda()
margin = Variable(margin)
_dist_term_sample = torch.sum(
torch.clamp(margin - _norm, min=0.0) ** 2)
_dist_term_sample = _dist_term_sample / \
(_n_objects_sample * (_n_objects_sample - 1))
dist_term += _dist_term_sample
dist_term = dist_term / bs
return dist_term
def calculate_regularization_term(means, n_objects, norm):
"""means: bs, n_instances, n_filters"""
bs, n_instances, n_filters = means.size()
reg_term = 0.0
for i in range(bs):
_mean_sample = means[i, : n_objects[i], :] # n_objects, n_filters
_norm = torch.norm(_mean_sample, norm, 1)
reg_term += torch.mean(_norm)
reg_term = reg_term / bs
return reg_term
def discriminative_loss(input, target, n_objects,
max_n_objects, delta_v, delta_d, norm, usegpu):
"""input: bs, n_filters, fmap, fmap
target: bs, n_instances, fmap, fmap
n_objects: bs"""
alpha = beta = 1.0
gamma = 0.001
bs, n_filters, height, width = input.size()
n_instances = target.size(1)
input = input.permute(0, 2, 3, 1).contiguous().view(
bs, height * width, n_filters)
target = target.permute(0, 2, 3, 1).contiguous().view(
bs, height * width, n_instances)
cluster_means = calculate_means(
input, target, n_objects, max_n_objects, usegpu)
var_term = calculate_variance_term(
input, target, cluster_means, n_objects, delta_v, norm)
dist_term = calculate_distance_term(
cluster_means, n_objects, delta_d, norm, usegpu)
reg_term = calculate_regularization_term(cluster_means, n_objects, norm)
loss = alpha * var_term + beta * dist_term + gamma * reg_term
return loss
class DiscriminativeLoss(_Loss):
def __init__(self, delta_var, delta_dist, norm,
size_average=True, reduce=True, usegpu=True):
super(DiscriminativeLoss, self).__init__(size_average)
self.reduce = reduce
assert self.size_average
assert self.reduce
self.delta_var = float(delta_var)
self.delta_dist = float(delta_dist)
self.norm = int(norm)
self.usegpu = usegpu
assert self.norm in [1, 2]
def forward(self, input, target, n_objects, max_n_objects):
torch.no_grad(target)
return discriminative_loss(input, target, n_objects, max_n_objects,
self.delta_var, self.delta_dist, self.norm,
self.usegpu)