-
Notifications
You must be signed in to change notification settings - Fork 0
/
model_6conv.py
165 lines (124 loc) · 5.43 KB
/
model_6conv.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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from Rot_DCF import Rot_DCF_Init, Rot_DCF, RotBN
from DCF import *
class Conv_Module(nn.Module):
def __init__(self, layer_type, feat_in, feat_out, first_rot=False, Ntheta=8,
K=3, K_a=5):
super(Conv_Module, self).__init__()
if layer_type == 'plain_conv':
self.conv = nn.Conv2d(feat_in, feat_out, kernel_size=5, padding=2)
self.bn = nn.BatchNorm2d(feat_out)
elif layer_type == 'rotdcf_conv':
if first_rot:
self.conv = Rot_DCF_Init(feat_out, feat_in, kernel_size=5, Ntheta=Ntheta,
K=K, K_a=K_a)
else:
self.conv = Rot_DCF(feat_out, feat_in, kernel_size=5, Ntheta=Ntheta,
K=K, K_a=K_a)
self.bn = RotBN(feat_out)
else:
self.conv = Conv_DCF(feat_in, feat_out, kernel_size=5, padding=2, num_bases=K, mode='mode0_1')
self.bn = nn.BatchNorm2d(feat_out)
def forward(self, x):
x = self.conv(x)
x = F.relu(x, inplace=True)
x = self.bn(x)
return x
class MNIST_CNN_Net(nn.Module):
def __init__(self, M=32):
super(MNIST_CNN_Net, self).__init__()
self.conv_layers = nn.Sequential(
Conv_Module('plain_conv', feat_in=1, feat_out=M),
Conv_Module('plain_conv', feat_in=M, feat_out=int(1.5*M)),
nn.AvgPool2d(kernel_size=2, stride=2),
Conv_Module('plain_conv', feat_in=int(1.5*M), feat_out=2*M),
Conv_Module('plain_conv', feat_in=2*M, feat_out=2*M),
nn.AvgPool2d(kernel_size=2, stride=2),
Conv_Module('plain_conv', feat_in=2*M, feat_out=3*M),
Conv_Module('plain_conv', feat_in=3*M, feat_out=4*M),
nn.AvgPool2d(kernel_size=2, stride=2),
)
# self.adapt_avg_pool = nn.AdaptiveAvgPool2d(output_size=1)
num_params = np.sum([param.numel() for param in self.parameters()])
self.fc1 = nn.Linear(3*3*4*M, 64)
self.fc2 = nn.Linear(64, 10)
self.dropout = nn.Dropout(0.5)
print('\nCreate CNN Net with num_features {}, Total Params: {}k'.\
format(M, num_params/1000))
print(self)
def forward(self, x):
bs, c, H, W = x.shape
x = self.conv_layers(x)
x = x.view(bs, -1)
x = F.relu(self.fc1(x), inplace=True)
x = self.dropout(x)
x = self.fc2(x)
return x
class MNIST_DCF_Net(nn.Module):
def __init__(self, M=32, K=5):
super(MNIST_DCF_Net, self).__init__()
self.K = K
self.conv_layers = nn.Sequential(
Conv_Module('dcf_conv', feat_in=1, feat_out=M),
Conv_Module('dcf_conv', feat_in=M, feat_out=int(1.5*M)),
nn.AvgPool2d(kernel_size=2, stride=2),
Conv_Module('dcf_conv', feat_in=int(1.5*M), feat_out=2*M),
Conv_Module('dcf_conv', feat_in=2*M, feat_out=2*M),
nn.AvgPool2d(kernel_size=2, stride=2),
Conv_Module('dcf_conv', feat_in=2*M, feat_out=3*M),
Conv_Module('dcf_conv', feat_in=3*M, feat_out=4*M),
nn.AvgPool2d(kernel_size=2, stride=2),
)
# self.adapt_avg_pool = nn.AdaptiveAvgPool2d(output_size=1)
num_params = np.sum([param.numel() for param in self.parameters()])
self.fc1 = nn.Linear(3*3*4*M, 64)
self.fc2 = nn.Linear(64, 10)
self.dropout = nn.Dropout(0.5)
print('\nCreate DCF Net with num_features {}, K {}, Total Params: {}k'.\
format(M, self.K, num_params/1000))
print(self)
def forward(self, x):
bs, c, H, W = x.shape
x = self.conv_layers(x)
x = x.view(bs, -1)
x = F.relu(self.fc1(x), inplace=True)
x = self.dropout(x)
x = self.fc2(x)
return x
class MNIST_RotDCF_Net(nn.Module):
def __init__(self, M=8, Ntheta=8, K=3, K_a=5):
super(MNIST_RotDCF_Net, self).__init__()
self.conv_layers = nn.Sequential(
Conv_Module('rotdcf_conv', 1, M, Ntheta=Ntheta, K=K, K_a=K_a, first_rot=True),
Conv_Module('rotdcf_conv', M, int(1.5*M), Ntheta=Ntheta, K=K, K_a=K_a),
nn.AvgPool2d(kernel_size=2, stride=2),
Conv_Module('rotdcf_conv', int(1.5*M), 2*M, Ntheta=Ntheta, K=K, K_a=K_a),
Conv_Module('rotdcf_conv', 2*M, 2*M, Ntheta=Ntheta, K=K, K_a=K_a),
nn.AvgPool2d(kernel_size=2, stride=2),
Conv_Module('rotdcf_conv', 2*M, 3*M, Ntheta=Ntheta, K=K, K_a=K_a),
Conv_Module('rotdcf_conv', 3*M, 4*M, Ntheta=Ntheta, K=K, K_a=K_a),
nn.AvgPool2d(kernel_size=2, stride=2),
)
num_params = np.sum([param.numel() for param in self.parameters()])
self.fc1 = nn.Linear(3*3*Ntheta*4*M, 64)
self.fc2 = nn.Linear(64, 10)
self.dropout = nn.Dropout(0.5)
print('\nCreate Rot-DCF Net with num_features {}, Ntheta {}, K {}, K_a {}; Total Params: {}k'.\
format(M, Ntheta, K, K_a, num_params/1000))
print(self)
def forward(self, x):
bs, c, H, W = x.shape
x = self.conv_layers(x)
x = x.view(bs, -1)
x = F.relu(self.fc1(x), inplace=True)
x = self.dropout(x)
x = self.fc2(x)
return x
if __name__ == '__main__':
x = torch.randn(4, 1, 28, 28)
net = MNIST_Net(M=8, Ntheta=16)
x = net(x)
print(x.shape)