network/Transformer.py

import torch
import torch.nn as nn
import torch.nn.functional as F


class Transformer(nn.Module):
    def __init__(self):
        super(Transformer, self).__init__()
        #
        self.refpad01_1 = nn.ReflectionPad2d(3)
        self.conv01_1 = nn.Conv2d(3, 64, 7)
        self.in01_1 = InstanceNormalization(64)
        # relu
        self.conv02_1 = nn.Conv2d(64, 128, 3, 2, 1)
        self.conv02_2 = nn.Conv2d(128, 128, 3, 1, 1)
        self.in02_1 = InstanceNormalization(128)
        # relu
        self.conv03_1 = nn.Conv2d(128, 256, 3, 2, 1)
        self.conv03_2 = nn.Conv2d(256, 256, 3, 1, 1)
        self.in03_1 = InstanceNormalization(256)
        # relu

        ## res block 1
        self.refpad04_1 = nn.ReflectionPad2d(1)
        self.conv04_1 = nn.Conv2d(256, 256, 3)
        self.in04_1 = InstanceNormalization(256)
        # relu
        self.refpad04_2 = nn.ReflectionPad2d(1)
        self.conv04_2 = nn.Conv2d(256, 256, 3)
        self.in04_2 = InstanceNormalization(256)
        # + input

        ## res block 2
        self.refpad05_1 = nn.ReflectionPad2d(1)
        self.conv05_1 = nn.Conv2d(256, 256, 3)
        self.in05_1 = InstanceNormalization(256)
        # relu
        self.refpad05_2 = nn.ReflectionPad2d(1)
        self.conv05_2 = nn.Conv2d(256, 256, 3)
        self.in05_2 = InstanceNormalization(256)
        # + input

        ## res block 3
        self.refpad06_1 = nn.ReflectionPad2d(1)
        self.conv06_1 = nn.Conv2d(256, 256, 3)
        self.in06_1 = InstanceNormalization(256)
        # relu
        self.refpad06_2 = nn.ReflectionPad2d(1)
        self.conv06_2 = nn.Conv2d(256, 256, 3)
        self.in06_2 = InstanceNormalization(256)
        # + input

        ## res block 4
        self.refpad07_1 = nn.ReflectionPad2d(1)
        self.conv07_1 = nn.Conv2d(256, 256, 3)
        self.in07_1 = InstanceNormalization(256)
        # relu
        self.refpad07_2 = nn.ReflectionPad2d(1)
        self.conv07_2 = nn.Conv2d(256, 256, 3)
        self.in07_2 = InstanceNormalization(256)
        # + input

        ## res block 5
        self.refpad08_1 = nn.ReflectionPad2d(1)
        self.conv08_1 = nn.Conv2d(256, 256, 3)
        self.in08_1 = InstanceNormalization(256)
        # relu
        self.refpad08_2 = nn.ReflectionPad2d(1)
        self.conv08_2 = nn.Conv2d(256, 256, 3)
        self.in08_2 = InstanceNormalization(256)
        # + input

        ## res block 6
        self.refpad09_1 = nn.ReflectionPad2d(1)
        self.conv09_1 = nn.Conv2d(256, 256, 3)
        self.in09_1 = InstanceNormalization(256)
        # relu
        self.refpad09_2 = nn.ReflectionPad2d(1)
        self.conv09_2 = nn.Conv2d(256, 256, 3)
        self.in09_2 = InstanceNormalization(256)
        # + input

        ## res block 7
        self.refpad10_1 = nn.ReflectionPad2d(1)
        self.conv10_1 = nn.Conv2d(256, 256, 3)
        self.in10_1 = InstanceNormalization(256)
        # relu
        self.refpad10_2 = nn.ReflectionPad2d(1)
        self.conv10_2 = nn.Conv2d(256, 256, 3)
        self.in10_2 = InstanceNormalization(256)
        # + input

        ## res block 8
        self.refpad11_1 = nn.ReflectionPad2d(1)
        self.conv11_1 = nn.Conv2d(256, 256, 3)
        self.in11_1 = InstanceNormalization(256)
        # relu
        self.refpad11_2 = nn.ReflectionPad2d(1)
        self.conv11_2 = nn.Conv2d(256, 256, 3)
        self.in11_2 = InstanceNormalization(256)
        # + input

        ##------------------------------------##
        self.deconv01_1 = nn.ConvTranspose2d(256, 128, 3, 2, 1, 1)
        self.deconv01_2 = nn.Conv2d(128, 128, 3, 1, 1)
        self.in12_1 = InstanceNormalization(128)
        # relu
        self.deconv02_1 = nn.ConvTranspose2d(128, 64, 3, 2, 1, 1)
        self.deconv02_2 = nn.Conv2d(64, 64, 3, 1, 1)
        self.in13_1 = InstanceNormalization(64)
        # relu
        self.refpad12_1 = nn.ReflectionPad2d(3)
        self.deconv03_1 = nn.Conv2d(64, 3, 7)
        # tanh

    def forward(self, x):
        y = F.relu(self.in01_1(self.conv01_1(self.refpad01_1(x))))
        y = F.relu(self.in02_1(self.conv02_2(self.conv02_1(y))))
        t04 = F.relu(self.in03_1(self.conv03_2(self.conv03_1(y))))

        ##
        y = F.relu(self.in04_1(self.conv04_1(self.refpad04_1(t04))))
        t05 = self.in04_2(self.conv04_2(self.refpad04_2(y))) + t04

        y = F.relu(self.in05_1(self.conv05_1(self.refpad05_1(t05))))
        t06 = self.in05_2(self.conv05_2(self.refpad05_2(y))) + t05

        y = F.relu(self.in06_1(self.conv06_1(self.refpad06_1(t06))))
        t07 = self.in06_2(self.conv06_2(self.refpad06_2(y))) + t06

        y = F.relu(self.in07_1(self.conv07_1(self.refpad07_1(t07))))
        t08 = self.in07_2(self.conv07_2(self.refpad07_2(y))) + t07

        y = F.relu(self.in08_1(self.conv08_1(self.refpad08_1(t08))))
        t09 = self.in08_2(self.conv08_2(self.refpad08_2(y))) + t08

        y = F.relu(self.in09_1(self.conv09_1(self.refpad09_1(t09))))
        t10 = self.in09_2(self.conv09_2(self.refpad09_2(y))) + t09

        y = F.relu(self.in10_1(self.conv10_1(self.refpad10_1(t10))))
        t11 = self.in10_2(self.conv10_2(self.refpad10_2(y))) + t10

        y = F.relu(self.in11_1(self.conv11_1(self.refpad11_1(t11))))
        y = self.in11_2(self.conv11_2(self.refpad11_2(y))) + t11
        ##

        y = F.relu(self.in12_1(self.deconv01_2(self.deconv01_1(y))))
        y = F.relu(self.in13_1(self.deconv02_2(self.deconv02_1(y))))
        y = torch.tanh(self.deconv03_1(self.refpad12_1(y)))

        return y


class InstanceNormalization(nn.Module):
    def __init__(self, dim, eps=1e-9):
        super(InstanceNormalization, self).__init__()
        self.scale = nn.Parameter(torch.FloatTensor(dim))
        self.shift = nn.Parameter(torch.FloatTensor(dim))
        self.eps = eps
        self._reset_parameters()

    def _reset_parameters(self):
        self.scale.data.uniform_()
        self.shift.data.zero_()

    def __call__(self, x):
        n = x.size(2) * x.size(3)
        t = x.view(x.size(0), x.size(1), n)
        mean = torch.mean(t, 2).unsqueeze(2).unsqueeze(3).expand_as(x)
        # Calculate the biased var. torch.var returns unbiased var
        var = torch.var(t, 2).unsqueeze(2).unsqueeze(3).expand_as(x) * (
            (n - 1) / float(n)
        )
        scale_broadcast = self.scale.unsqueeze(1).unsqueeze(1).unsqueeze(0)
        scale_broadcast = scale_broadcast.expand_as(x)
        shift_broadcast = self.shift.unsqueeze(1).unsqueeze(1).unsqueeze(0)
        shift_broadcast = shift_broadcast.expand_as(x)
        out = (x - mean) / torch.sqrt(var + self.eps)
        out = out * scale_broadcast + shift_broadcast
        return out