src/visual_stream.py

import chainer

class ConvolutionBlock(chainer.Chain):
    def __init__(self, in_channels, out_channels):
        super(ConvolutionBlock, self).__init__(
            conv=chainer.links.Convolution2D(
                in_channels,
                out_channels,
                7,
                2,
                3,
                initialW=chainer.initializers.HeNormal(),
            ),
            bn_conv=chainer.links.BatchNormalization(out_channels),
        )

    def __call__(self, x):
        chainer.config.train = False

        h = self.conv(x)
        h = self.bn_conv(h)
        y = chainer.functions.relu(h)

        return y


class ResidualBlock(chainer.Chain):
    def __init__(self, in_channels, out_channels):
        super(ResidualBlock, self).__init__(
            res_branch2a=chainer.links.Convolution2D(
                in_channels,
                out_channels,
                3,
                pad=1,
                initialW=chainer.initializers.HeNormal(),
            ),
            bn_branch2a=chainer.links.BatchNormalization(out_channels),
            res_branch2b=chainer.links.Convolution2D(
                out_channels,
                out_channels,
                3,
                pad=1,
                initialW=chainer.initializers.HeNormal(),
            ),
            bn_branch2b=chainer.links.BatchNormalization(out_channels),
        )

    def __call__(self, x):
        chainer.config.train = False

        h = self.res_branch2a(x)
        h = self.bn_branch2a(h)
        h = chainer.functions.relu(h)
        h = self.res_branch2b(h)
        h = self.bn_branch2b(h)
        h = x + h
        y = chainer.functions.relu(h)

        return y


class ResidualBlockA:
    def __init__(self):
        pass

    def __call__(self):
        pass


class ResidualBlockB(chainer.Chain):
    def __init__(self, in_channels, out_channels):
        super(ResidualBlockB, self).__init__(
            res_branch1=chainer.links.Convolution2D(
                in_channels,
                out_channels,
                1,
                2,
                initialW=chainer.initializers.HeNormal(),
            ),
            bn_branch1=chainer.links.BatchNormalization(out_channels),
            res_branch2a=chainer.links.Convolution2D(
                in_channels,
                out_channels,
                3,
                2,
                1,
                initialW=chainer.initializers.HeNormal(),
            ),
            bn_branch2a=chainer.links.BatchNormalization(out_channels),
            res_branch2b=chainer.links.Convolution2D(
                out_channels,
                out_channels,
                3,
                pad=1,
                initialW=chainer.initializers.HeNormal(),
            ),
            bn_branch2b=chainer.links.BatchNormalization(out_channels),
        )

    def __call__(self, x):
        chainer.config.train = False

        temp = self.res_branch1(x)
        temp = self.bn_branch1(temp)
        h = self.res_branch2a(x)
        h = self.bn_branch2a(h)
        h = chainer.functions.relu(h)
        h = self.res_branch2b(h)
        h = self.bn_branch2b(h)
        h = temp + h
        y = chainer.functions.relu(h)

        return y


class ResNet18(chainer.Chain):
    def __init__(self):
        super(ResNet18, self).__init__(
            conv1_relu=ConvolutionBlock(3, 32),
            res2a_relu=ResidualBlock(32, 32),
            res2b_relu=ResidualBlock(32, 32),
            res3a_relu=ResidualBlockB(32, 64),
            res3b_relu=ResidualBlock(64, 64),
            res4a_relu=ResidualBlockB(64, 128),
            res4b_relu=ResidualBlock(128, 128),
            res5a_relu=ResidualBlockB(128, 256),
            res5b_relu=ResidualBlock(256, 256),
        )

    def __call__(self, x):
        chainer.config.train = False

        h = self.conv1_relu(x)
        h = chainer.functions.max_pooling_2d(h, 3, 2, 1)
        h = self.res2a_relu(h)
        h = self.res2b_relu(h)
        h = self.res3a_relu(h)
        h = self.res3b_relu(h)
        h = self.res4a_relu(h)
        h = self.res4b_relu(h)
        h = self.res5a_relu(h)
        h = self.res5b_relu(h)
        y = chainer.functions.average_pooling_2d(h, h.data.shape[2:])

        return y