XNOR-Net/net.py at master · rarilurelo/XNOR-Net · GitHub

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
import chainer
import chainer.functions as F
import chainer.links as L
#import link_binary_linear
import link_binary_xnor_convolution
import bst_xnor


#class MnistBWCNN(chainer.Chain):
#    """ An example of binary CNN for MNIST dataset."""
#    def __init__(self):
#        super(MnistCNN, self).__init__(
#            c1=link_binary_convolution.BinaryConvolution2D(1, 32, 5),
#            b1=L.BatchNormalization(32),
#            c2=link_binary_convolution.BinaryConvolution2D(32, 64, 5),
#            b2=L.BatchNormalization(64),
#            c3=link_binary_convolution.BinaryConvolution2D(64, 128, 5),
#            b3=L.BatchNormalization(128),
#            l1=link_binary_linear.BinaryLinear(128*16*16, 256),
#            b4=L.BatchNormalization(256),
#            l2=link_binary_linear.BinaryLinear(256, 10),
#            b5=L.BatchNormalization(10)
#        )
#        self.train = True
#
#    def __call__(self, x):
#        h1 = bst.bst(self.b1(self.c1(x), test=not self.train))
#        h2 = bst.bst(self.b2(self.c2(h1), test=not self.train))
#        h3 = bst.bst(self.b3(self.c3(h2), test=not self.train))
#        h4 = bst.bst(self.b4(self.l1(h3), test=not self.train))
#        return self.b5(self.l2(h4), test=not self.train)

class Cifar10BWCNN(chainer.Chain):
    """ binary CNN fot Cifar10"""
    def __init__(self):
        super(Cifar10BWCNN, self).__init__(
            c1=link_binary_xnor_convolution.BinaryXnorConvolution2D(3, 32, 5),
            b1=L.BatchNormalization(32),
            c2=link_binary_xnor_convolution.BinaryXnorConvolution2D(32, 64, 5, pad=2),
            b2=L.BatchNormalization(64),
            l1=L.Linear(64*9*9, 256),
            b3=L.BatchNormalization(256),
            l2=L.Linear(256, 10),
            b4=L.BatchNormalization(10)
        )
        self.train = True

    def __call__(self, x):
        h1 = F.max_pooling_2d(self.b1(self.c1(x), test=not self.train), ksize=4, stride=2, pad=2)
        h2 = F.max_pooling_2d(self.b2(self.c2(h1), test=not self.train), ksize=4, stride=2, pad=2)
        h3 = F.relu(self.b3(self.l1(h2)))
        return self.b4(self.l2(h3))