forked from coder3000/SR-CapsNet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathconvnet.py
159 lines (128 loc) · 6.09 KB
/
convnet.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from modules import *
from utils import weights_init
class ConvNet(nn.Module):
def __init__(self, planes, cfg_data, num_caps, caps_size, depth, mode):
caps_size = 16
super(ConvNet, self).__init__()
channels, classes = cfg_data['channels'], cfg_data['classes']
self.num_caps = num_caps
self.caps_size = caps_size
self.depth = depth
self.mode = mode
self.layers = nn.Sequential(
nn.Conv2d(channels, planes, kernel_size=3, stride=1, padding=1, bias=False),
nn.BatchNorm2d(planes),
nn.ReLU(True),
nn.Conv2d(planes, planes*2, kernel_size=3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(planes*2),
nn.ReLU(True),
nn.Conv2d(planes*2, planes*2, kernel_size=3, stride=1, padding=1, bias=False),
nn.BatchNorm2d(planes*2),
nn.ReLU(True),
nn.Conv2d(planes*2, planes*4, kernel_size=3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(planes*4),
nn.ReLU(True),
nn.Conv2d(planes*4, planes*4, kernel_size=3, stride=1, padding=1, bias=False),
nn.BatchNorm2d(planes*4),
nn.ReLU(True),
nn.Conv2d(planes*4, planes*8, kernel_size=3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(planes*8),
nn.ReLU(True),
)
self.conv_layers = nn.ModuleList()
self.norm_layers = nn.ModuleList()
#========= ConvCaps Layers
for d in range(1, depth):
if self.mode == 'DR':
self.conv_layers.append(DynamicRouting2d(num_caps, num_caps, caps_size, caps_size, kernel_size=3, stride=1, padding=1))
nn.init.normal_(self.conv_layers[0].W, 0, 0.5)
elif self.mode == 'EM':
self.conv_layers.append(EmRouting2d(num_caps, num_caps, caps_size, kernel_size=3, stride=1, padding=1))
self.norm_layers.append(nn.BatchNorm2d(4*4*num_caps))
elif self.mode == 'SR':
self.conv_layers.append(SelfRouting2d(num_caps, num_caps, caps_size, caps_size, kernel_size=3, stride=1, padding=1, pose_out=True))
self.norm_layers.append(nn.BatchNorm2d(planes*num_caps))
else:
break
final_shape = 4
# DR
if self.mode == 'DR':
self.conv_pose = nn.Conv2d(8*planes, num_caps*caps_size, kernel_size=3, stride=1, padding=1, bias=False)
self.bn_pose = nn.BatchNorm2d(num_caps*caps_size)
self.fc = DynamicRouting2d(num_caps, classes, caps_size, caps_size, kernel_size=final_shape, padding=0)
# initialize so that output logits are in reasonable range (0.1-0.9)
nn.init.normal_(self.fc.W, 0, 0.1)
# EM
elif self.mode == 'EM':
self.conv_a = nn.Conv2d(8*planes, num_caps, kernel_size=3, stride=1, padding=1, bias=False)
self.conv_pose = nn.Conv2d(8*planes, num_caps*caps_size, kernel_size=3, stride=1, padding=1, bias=False)
self.bn_a = nn.BatchNorm2d(num_caps)
self.bn_pose = nn.BatchNorm2d(num_caps*caps_size)
self.fc = EmRouting2d(num_caps, classes, caps_size, kernel_size=final_shape, padding=0)
# SR
elif self.mode == 'SR':
self.conv_a = nn.Conv2d(8*planes, num_caps, kernel_size=3, stride=1, padding=1, bias=False)
self.conv_pose = nn.Conv2d(8*planes, num_caps*caps_size, kernel_size=3, stride=1, padding=1, bias=False)
self.bn_a = nn.BatchNorm2d(num_caps)
self.bn_pose = nn.BatchNorm2d(num_caps*caps_size)
self.fc = SelfRouting2d(num_caps, classes, caps_size, 1, kernel_size=final_shape, padding=0, pose_out=False)
# avg pooling
elif self.mode == 'AVG':
self.pool = nn.AvgPool2d(final_shape)
self.fc = nn.Linear(8*planes, classes)
# max pooling
elif self.mode == 'MAX':
self.pool = nn.MaxPool2d(final_shape)
self.fc = nn.Linear(8*planes, classes)
elif self.mode == 'FC':
self.conv_ = nn.Conv2d(8*planes, num_caps*caps_size, kernel_size=3, stride=1, padding=1, bias=False)
self.bn_ = nn.BatchNorm2d(num_caps*caps_size)
self.fc = nn.Linear(num_caps*caps_size*final_shape*final_shape, classes)
self.apply(weights_init)
def forward(self, x):
out = self.layers(x)
# DR
if self.mode == 'DR':
pose = self.bn_pose(self.conv_pose(out))
b, c, h, w = pose.shape
pose = pose.permute(0, 2, 3, 1).contiguous()
pose = squash(pose.view(b, h, w, self.num_caps, self.caps_size))
pose = pose.view(b, h, w, -1)
pose = pose.permute(0, 3, 1, 2)
for m in self.conv_layers:
pose = m(pose)
out = self.fc(pose)
out = out.view(b, -1, self.caps_size)
out = out.norm(dim=-1)
# EM
elif self.mode == 'EM':
a, pose = self.conv_a(out), self.conv_pose(out)
a, pose = torch.sigmoid(self.bn_a(a)), self.bn_pose(pose)
for m, bn in zip(self.conv_layers, self.norm_layers):
a, pose = m(a, pose)
pose = bn(pose)
a, _ = self.fc(a, pose)
out = a.view(a.size(0), -1)
# ours
elif self.mode == 'SR':
a, pose = self.conv_a(out), self.conv_pose(out)
a, pose = torch.sigmoid(self.bn_a(a)), self.bn_pose(pose)
for m, bn in zip(self.conv_layers, self.norm_layers):
a, pose = m(a, pose)
pose = bn(pose)
a, _ = self.fc(a, pose)
out = a.view(a.size(0), -1)
out = out.log()
elif self.mode == 'AVG' or self.mode =='MAX':
out = self.pool(out)
out = out.view(out.size(0), -1)
out = self.fc(out)
elif self.mode == 'FC':
out = F.relu(self.bn_(self.conv_(out)))
out = out.view(out.size(0), -1)
out = self.fc(out)
return out