-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodel.py
196 lines (183 loc) · 7.59 KB
/
model.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
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
'''Custom ResNet with multi scale residual attention.'''
# import pdb
import torch
import torch.nn as nn
import torch.nn.functional as F
class ResBlock(nn.Module):
''' Single residual block with variable number of conv layers.
Attention can be turned on or off.
'''
def __init__(self, init_feats, feats, num_layers, attention=True,
downsample=False):
'''
Args:
init_feats (int): number of input channels in first layer
feats (int): number of feature channels
num_layers (int): number of layers in the block
attention (bool): To set attention on/off
downsample (bool): To set downsampling on/off
'''
super(ResBlock, self).__init__()
self.attention = attention
self.downsample = downsample
self.attn_scaler = nn.Parameter(torch.Tensor([1]))
if self.downsample:
stride = 2
self.down_conv = nn.Conv2d(init_feats, feats, kernel_size=1,
stride=2, bias=False)
else:
stride = 1
layers = []
for layer_num in range(num_layers):
conv_layer = nn.Conv2d(init_feats, feats, kernel_size=3,
stride=stride, padding=1, bias=False)
stride = 1
init_feats = feats
norm_layer = nn.BatchNorm2d(feats)
relu = nn.ReLU(inplace=True)
layers.append(conv_layer)
layers.append(norm_layer)
if layer_num < (num_layers - 1):
layers.append(relu)
self.conv_block = nn.Sequential(*layers)
def forward(self, x):
if isinstance(x, tuple):
x = x[0]
identity = x
out = self.conv_block(x)
if self.downsample:
identity = self.down_conv(identity)
if self.attention:
attn = torch.sigmoid(out)
attn_out = self.attn_scaler*identity*attn
# out = torch.sigmoid(out)
# out = identity*out
res = identity+out+attn_out
else:
res = identity+out
res = F.relu(res)
return res, attn
class AuxClassifier(nn.Module):
def __init__(self, in_channels, num_classes):
super(AuxClassifier, self).__init__()
self.conv0 = nn.Conv2d(in_channels, in_channels*4, kernel_size=3,
stride=4, padding=1)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.final = nn.Linear(in_channels*4, num_classes)
def forward(self, x):
x = F.relu(self.conv0(x))
x = self.avgpool(x)
x = torch.flatten(x, 1)
out = self.final(x)
return out
class MARL(nn.Module):
def __init__(self, in_channels, num_blocks, num_layers, num_classes=2,
num_feats=64, downsample_freq=1):
'''
Args:
in_channels (int): number of channels in input
num_blocks (int): number of residual block
num_layers (int): number of layers in each block
num_classes (int): number of classes in classification
num_feats (int): number of feature channels in first layer
downsample_freq (int): number of blocks after which to downsample.
'''
super(MARL, self).__init__()
self.downsample_freq = downsample_freq
self.conv1 = nn.Conv2d(in_channels, num_feats, kernel_size=7,
stride=2, padding=3, bias=False)
self.bn1 = nn.BatchNorm2d(num_feats)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layers = []
self.num_blocks = num_blocks
init_feats = num_feats
downsample = False
for block_num in range(num_blocks):
block = ResBlock(init_feats, num_feats, num_layers,
downsample=downsample, attention=True)
init_feats = num_feats
self.layers.append(block.cuda())
downsample = False
cond = (block_num + 1 != num_blocks)
if ((block_num+1) % self.downsample_freq == 0) and cond:
num_feats *= 2
downsample = True
self.main_arch = nn.Sequential(*self.layers)
self.attn_conv = nn.Conv2d(960, 512, kernel_size=3, stride=4)
# self.attn_conv = nn.Conv2d(1920, 512, kernel_size=3, stride=4)
self.semifinal = nn.Conv2d(512, 512, kernel_size=3, stride=1,
padding=1)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.final = nn.Linear(num_feats, num_classes)
self.aux_classifier = AuxClassifier(128, num_classes)
# self.weights = nn.Parameter(torch.ones(num_blocks, 128, 128)).cuda()
def forward(self, x):
attn_map_list = []
conicity_sum = 0
x = F.relu(self.bn1(self.conv1(x)))
x = self.maxpool(x)
for num, block in enumerate(self.main_arch, 1):
x, attn_map = block(x)
if self.training and num == (self.num_blocks // 2):
aux = self.aux_classifier(x)
conicity = self.get_conicity(attn_map)
conicity_sum += conicity
# attn_map = F.interpolate(attn_map, (52, 56),
# align_corners=False, mode='bilinear')
attn_map_list.append(attn_map)
mid_size = attn_map_list[(len(attn_map_list)//2) - 1].shape
for idx, attn_map in enumerate(attn_map_list):
attn_map_list[idx] = F.interpolate(attn_map,
(mid_size[2], mid_size[3]),
align_corners=False,
mode='bilinear')
# import pdb
# pdb.set_trace()
stacked_attn_map = torch.cat(attn_map_list, 1)
ms_attn = torch.sigmoid(self.attn_conv(stacked_attn_map))
x = ms_attn*x
x = self.semifinal(x)
x = self.avgpool(x)
x = torch.flatten(x, 1)
out = self.final(x)
if self.training:
return out, aux, conicity_sum
else:
return out, conicity_sum
def get_conicity(self, attn_map):
atm = 0
attn_map = torch.reshape(attn_map, (-1, attn_map.shape[1],
attn_map.shape[2]
* attn_map.shape[3]))
# taking each channel as vector
mean_vec = torch.mean(attn_map, 1).unsqueeze(1)
# for i in range(attn_map.shape[1]):
# atm += F.cosine_similarity(attn_map[:, i], mean_vec)
# conicity = atm.float()/(i+1)
atm = F.cosine_similarity(attn_map, mean_vec, 2)
conicity = torch.mean(atm, 1)
return conicity
class ARL(MARL):
def __init__(self, in_channels, num_blocks, num_layers, num_classes=2,
num_feats=64, downsample_freq=1):
super(ARL, self).__init__(
in_channels, num_blocks, num_layers, num_classes, num_feats=64,
downsample_freq=1
)
def forward(self, x):
conicity_sum = 0
x = F.relu(self.bn1(self.conv1(x)))
x = self.maxpool(x)
for num, block in enumerate(self.main_arch, 1):
x, attn_map = block(x)
if self.training and num == (self.num_blocks // 2):
aux = self.aux_classifier(x)
conicity = self.get_conicity(attn_map)
conicity_sum += conicity
x = self.avgpool(x)
x = torch.flatten(x, 1)
out = self.final(x)
if self.training:
return out, aux, conicity_sum
else:
return out, conicity_sum