models.py

import torch
from torch import nn
import torchvision
import math


class ConvolutionalBlock(nn.Module):
    """
    A convolutional block, comprising convolutional, BN, activation layers.
    """

    def __init__(self, in_channels, out_channels, kernel_size, stride=1, batch_norm=False, activation=None):
        """
        :param in_channels: number of input channels
        :param out_channels: number of output channe;s
        :param kernel_size: kernel size
        :param stride: stride
        :param batch_norm: include a BN layer?
        :param activation: Type of activation; None if none
        """
        super(ConvolutionalBlock, self).__init__()

        if activation is not None:
            activation = activation.lower()
            assert activation in {'prelu', 'leakyrelu', 'tanh'}

        # A container that will hold the layers in this convolutional block
        layers = list()

        # A convolutional layer
        layers.append(
            nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride,
                      padding=kernel_size // 2))

        # A batch normalization (BN) layer, if wanted
        if batch_norm is True:
            layers.append(nn.BatchNorm2d(num_features=out_channels))

        # An activation layer, if wanted
        if activation == 'prelu':
            layers.append(nn.PReLU())
        elif activation == 'leakyrelu':
            layers.append(nn.LeakyReLU(0.2))
        elif activation == 'tanh':
            layers.append(nn.Tanh())

        # Put together the convolutional block as a sequence of the layers in this container
        self.conv_block = nn.Sequential(*layers)

    def forward(self, input):
        """
        Forward propagation.
        :param input: input images, a tensor of size (N, in_channels, w, h)
        :return: output images, a tensor of size (N, out_channels, w, h)
        """
        output = self.conv_block(input)  # (N, out_channels, w, h)

        return output


class SubPixelConvolutionalBlock(nn.Module):
    """
    A subpixel convolutional block, comprising convolutional, pixel-shuffle, and PReLU activation layers.
    """

    def __init__(self, kernel_size=3, n_channels=64, scaling_factor=2):
        """
        :param kernel_size: kernel size of the convolution
        :param n_channels: number of input and output channels
        :param scaling_factor: factor to scale input images by (along both dimensions)
        """
        super(SubPixelConvolutionalBlock, self).__init__()

        # A convolutional layer that increases the number of channels by scaling factor^2, followed by pixel shuffle and PReLU
        self.conv = nn.Conv2d(in_channels=n_channels, out_channels=n_channels * (scaling_factor ** 2),
                              kernel_size=kernel_size, padding=kernel_size // 2)
        # These additional channels are shuffled to form additional pixels, upscaling each dimension by the scaling factor
        self.pixel_shuffle = nn.PixelShuffle(upscale_factor=scaling_factor)
        self.prelu = nn.PReLU()

    def forward(self, input):
        """
        Forward propagation.
        :param input: input images, a tensor of size (N, n_channels, w, h)
        :return: scaled output images, a tensor of size (N, n_channels, w * scaling factor, h * scaling factor)
        """
        output = self.conv(input)  # (N, n_channels * scaling factor^2, w, h)
        output = self.pixel_shuffle(output)  # (N, n_channels, w * scaling factor, h * scaling factor)
        output = self.prelu(output)  # (N, n_channels, w * scaling factor, h * scaling factor)

        return output


class ResidualBlock(nn.Module):
    """
    A residual block, comprising two convolutional blocks with a residual connection across them.
    """

    def __init__(self, kernel_size=3, n_channels=64):
        """
        :param kernel_size: kernel size
        :param n_channels: number of input and output channels (same because the input must be added to the output)
        """
        super(ResidualBlock, self).__init__()

        # The first convolutional block
        self.conv_block1 = ConvolutionalBlock(in_channels=n_channels, out_channels=n_channels, kernel_size=kernel_size,
                                              batch_norm=True, activation='PReLu')

        # The second convolutional block
        self.conv_block2 = ConvolutionalBlock(in_channels=n_channels, out_channels=n_channels, kernel_size=kernel_size,
                                              batch_norm=True, activation=None)

    def forward(self, input):
        """
        Forward propagation.
        :param input: input images, a tensor of size (N, n_channels, w, h)
        :return: output images, a tensor of size (N, n_channels, w, h)
        """
        residual = input  # (N, n_channels, w, h)
        output = self.conv_block1(input)  # (N, n_channels, w, h)
        output = self.conv_block2(output)  # (N, n_channels, w, h)
        output = output + residual  # (N, n_channels, w, h)

        return output


class SRResNet(nn.Module):
    """
    The SRResNet, as defined in the paper.
    """

    def __init__(self, large_kernel_size=9, small_kernel_size=3, n_channels=64, n_blocks=16, scaling_factor=4):
        """
        :param large_kernel_size: kernel size of the first and last convolutions which transform the inputs and outputs
        :param small_kernel_size: kernel size of all convolutions in-between, i.e. those in the residual and subpixel convolutional blocks
        :param n_channels: number of channels in-between, i.e. the input and output channels for the residual and subpixel convolutional blocks
        :param n_blocks: number of residual blocks
        :param scaling_factor: factor to scale input images by (along both dimensions) in the subpixel convolutional block
        """
        super(SRResNet, self).__init__()

        # Scaling factor must be 2, 4, or 8
        scaling_factor = int(scaling_factor)
        assert scaling_factor in {2, 4, 8}, "The scaling factor must be 2, 4, or 8!"

        # The first convolutional block
        self.conv_block1 = ConvolutionalBlock(in_channels=3, out_channels=n_channels, kernel_size=large_kernel_size,
                                              batch_norm=False, activation='PReLu')

        # A sequence of n_blocks residual blocks, each containing a skip-connection across the block
        self.residual_blocks = nn.Sequential(
            *[ResidualBlock(kernel_size=small_kernel_size, n_channels=n_channels) for i in range(n_blocks)])

        # Another convolutional block
        self.conv_block2 = ConvolutionalBlock(in_channels=n_channels, out_channels=n_channels,
                                              kernel_size=small_kernel_size,
                                              batch_norm=True, activation=None)

        # Upscaling is done by sub-pixel convolution, with each such block upscaling by a factor of 2
        n_subpixel_convolution_blocks = int(math.log2(scaling_factor))
        self.subpixel_convolutional_blocks = nn.Sequential(
            *[SubPixelConvolutionalBlock(kernel_size=small_kernel_size, n_channels=n_channels, scaling_factor=2) for i
              in range(n_subpixel_convolution_blocks)])

        # The last convolutional block
        self.conv_block3 = ConvolutionalBlock(in_channels=n_channels, out_channels=3, kernel_size=large_kernel_size,
                                              batch_norm=False, activation='Tanh')

    def forward(self, lr_imgs):
        """
        Forward prop.
        :param lr_imgs: low-resolution input images, a tensor of size (N, 3, w, h)
        :return: super-resolution output images, a tensor of size (N, 3, w * scaling factor, h * scaling factor)
        """
        output = self.conv_block1(lr_imgs)  # (N, 3, w, h)
        residual = output  # (N, n_channels, w, h)
        output = self.residual_blocks(output)  # (N, n_channels, w, h)
        output = self.conv_block2(output)  # (N, n_channels, w, h)
        output = output + residual  # (N, n_channels, w, h)
        output = self.subpixel_convolutional_blocks(output)  # (N, n_channels, w * scaling factor, h * scaling factor)
        sr_imgs = self.conv_block3(output)  # (N, 3, w * scaling factor, h * scaling factor)

        return sr_imgs