-
Notifications
You must be signed in to change notification settings - Fork 19
/
onemodel_sg-one.py
156 lines (111 loc) · 5.39 KB
/
onemodel_sg-one.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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from models.vgg import vgg_sg as vgg
class OneModel(nn.Module):
def __init__(self, args):
super(OneModel, self).__init__()
self.netB = vgg.vgg16(pretrained=True, use_decoder=True)
self.classifier_6 = nn.Sequential(
nn.Conv2d(128, 128, kernel_size=3, dilation=1, padding=1), #fc6
nn.ReLU(inplace=True)
)
self.exit_layer = nn.Conv2d(128, 2, kernel_size=1, padding=1)
# self.bce_logits_func = nn.BCEWithLogitsLoss()
self.bce_logits_func = nn.CrossEntropyLoss()
self.loss_func = nn.BCELoss()
self.cos_similarity_func = nn.CosineSimilarity()
self.triplelet_func = nn.TripletMarginLoss(margin=2.0)
def forward(self, anchor_img, pos_img, neg_img, pos_mask):
outA_pos, outA_side = self.netB(pos_img)
_, _, mask_w, mask_h = pos_mask.size()
outA_pos = F.upsample(outA_pos, size=(mask_w, mask_h), mode='bilinear')
vec_pos = torch.sum(torch.sum(outA_pos*pos_mask, dim=3), dim=2)/torch.sum(pos_mask)
outB, outB_side= self.netB(anchor_img)
# tmp_seg = outB * vec_pos.unsqueeze(dim=2).unsqueeze(dim=3)
vec_pos = vec_pos.unsqueeze(dim=2).unsqueeze(dim=3)
tmp_seg = self.cos_similarity_func(outB, vec_pos)
exit_feat_in = outB_side * tmp_seg.unsqueeze(dim=1)
outB_side_6 = self.classifier_6(exit_feat_in)
outB_side = self.exit_layer(outB_side_6)
return outB, tmp_seg, vec_pos, outB_side
def forward_5shot_avg(self, anchor_img, pos_img_list, pos_mask_list):
vec_pos_sum = 0.
for i in range(5):
pos_img = pos_img_list[i]
pos_mask = pos_mask_list[i]
pos_img = self.warper_img(pos_img)
pos_mask = self.warper_img(pos_mask)
outA_pos, _ = self.netB(pos_img)
_, _, mask_w, mask_h = pos_mask.size()
outA_pos = F.upsample(outA_pos, size=(mask_w, mask_h), mode='bilinear')
vec_pos = torch.sum(torch.sum(outA_pos*pos_mask, dim=3), dim=2)/torch.sum(pos_mask)
vec_pos_sum += vec_pos
vec_pos = vec_pos_sum/5.0
outB, outB_side = self.netB(anchor_img)
# tmp_seg = outB * vec_pos.unsqueeze(dim=2).unsqueeze(dim=3)
vec_pos = vec_pos.unsqueeze(dim=2).unsqueeze(dim=3)
tmp_seg = self.cos_similarity_func(outB, vec_pos)
exit_feat_in = outB_side * tmp_seg.unsqueeze(dim=1)
outB_side_6 = self.classifier_6(exit_feat_in)
outB_side = self.exit_layer(outB_side_6)
return outB, outA_pos, vec_pos, outB_side
def warper_img(self, img):
img_tensor = torch.Tensor(img).cuda()
img_var = Variable(img_tensor)
img_var = torch.unsqueeze(img_var, dim=0)
return img_var
def forward_5shot_max(self, anchor_img, pos_img_list, pos_mask_list):
outB_side_list = []
for i in range(5):
pos_img = pos_img_list[i]
pos_mask = pos_mask_list[i]
pos_img = self.warper_img(pos_img)
pos_mask = self.warper_img(pos_mask)
outA_pos, _ = self.netB(pos_img)
_, _, mask_w, mask_h = pos_mask.size()
outA_pos = F.upsample(outA_pos, size=(mask_w, mask_h), mode='bilinear')
vec_pos = torch.sum(torch.sum(outA_pos*pos_mask, dim=3), dim=2)/torch.sum(pos_mask)
outB, outB_side = self.netB(anchor_img)
# tmp_seg = outB * vec_pos.unsqueeze(dim=2).unsqueeze(dim=3)
vec_pos = vec_pos.unsqueeze(dim=2).unsqueeze(dim=3)
tmp_seg = self.cos_similarity_func(outB, vec_pos)
exit_feat_in = outB_side * tmp_seg.unsqueeze(dim=1)
outB_side_6 = self.classifier_6(exit_feat_in)
outB_side = self.exit_layer(outB_side_6)
outB_side_list.append(outB_side)
return outB, outA_pos, vec_pos, outB_side_list
def get_loss(self, logits, query_label):
outB, outA_pos, vec_pos, outB_side = logits
b, c, w, h = query_label.size()
outB_side = F.upsample(outB_side, size=(w, h), mode='bilinear')
outB_side = outB_side.permute(0,2,3,1).view(w*h, 2)
query_label = query_label.view(-1)
loss_bce_seg = self.bce_logits_func(outB_side, query_label.long())
loss = loss_bce_seg
return loss, 0,0
def get_pred_5shot_max(self, logits, query_label):
outB, outA_pos, vec_pos, outB_side_list = logits
w, h = query_label.size()[-2:]
res_pred = None
for i in range(5):
outB_side = outB_side_list[i]
outB_side = F.upsample(outB_side, size=(w, h), mode='bilinear')
out_side = F.softmax(outB_side, dim=1).squeeze()
values, pred = torch.max(out_side, dim=0)
if res_pred is None:
res_pred = pred
else:
res_pred = torch.max(pred, res_pred)
return values, res_pred
def get_pred(self, logits, query_image):
outB, outA_pos, vec_pos, outB_side = logits
w, h = query_image.size()[-2:]
outB_side = F.upsample(outB_side, size=(w, h), mode='bilinear')
out_softmax = F.softmax(outB_side, dim=1).squeeze()
values, pred = torch.max(out_softmax, dim=0)
return out_softmax, pred