path: root/modules/esrgan_model_arch.py

# this file is adapted from https://github.com/victorca25/iNNfer

import math
import functools
import torch
import torch.nn as nn
import torch.nn.functional as F


####################
# RRDBNet Generator
####################

class RRDBNet(nn.Module):
    def __init__(self, in_nc, out_nc, nf, nb, nr=3, gc=32, upscale=4, norm_type=None,
            act_type='leakyrelu', mode='CNA', upsample_mode='upconv', convtype='Conv2D',
            finalact=None, gaussian_noise=False, plus=False):
        super(RRDBNet, self).__init__()
        n_upscale = int(math.log(upscale, 2))
        if upscale == 3:
            n_upscale = 1

        self.resrgan_scale = 0
        if in_nc % 16 == 0:
            self.resrgan_scale = 1
        elif in_nc != 4 and in_nc % 4 == 0:
            self.resrgan_scale = 2

        fea_conv = conv_block(in_nc, nf, kernel_size=3, norm_type=None, act_type=None, convtype=convtype)
        rb_blocks = [RRDB(nf, nr, kernel_size=3, gc=32, stride=1, bias=1, pad_type='zero',
            norm_type=norm_type, act_type=act_type, mode='CNA', convtype=convtype,
            gaussian_noise=gaussian_noise, plus=plus) for _ in range(nb)]
        LR_conv = conv_block(nf, nf, kernel_size=3, norm_type=norm_type, act_type=None, mode=mode, convtype=convtype)

        if upsample_mode == 'upconv':
            upsample_block = upconv_block
        elif upsample_mode == 'pixelshuffle':
            upsample_block = pixelshuffle_block
        else:
            raise NotImplementedError('upsample mode [{:s}] is not found'.format(upsample_mode))
        if upscale == 3:
            upsampler = upsample_block(nf, nf, 3, act_type=act_type, convtype=convtype)
        else:
            upsampler = [upsample_block(nf, nf, act_type=act_type, convtype=convtype) for _ in range(n_upscale)]
        HR_conv0 = conv_block(nf, nf, kernel_size=3, norm_type=None, act_type=act_type, convtype=convtype)
        HR_conv1 = conv_block(nf, out_nc, kernel_size=3, norm_type=None, act_type=None, convtype=convtype)

        outact = act(finalact) if finalact else None

        self.model = sequential(fea_conv, ShortcutBlock(sequential(*rb_blocks, LR_conv)),
            *upsampler, HR_conv0, HR_conv1, outact)

    def forward(self, x, outm=None):
        if self.resrgan_scale == 1:
            feat = pixel_unshuffle(x, scale=4)
        elif self.resrgan_scale == 2:
            feat = pixel_unshuffle(x, scale=2)
        else:
            feat = x

        return self.model(feat)


class RRDB(nn.Module):
    """
    Residual in Residual Dense Block
    (ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks)
    """

    def __init__(self, nf, nr=3, kernel_size=3, gc=32, stride=1, bias=1, pad_type='zero',
            norm_type=None, act_type='leakyrelu', mode='CNA', convtype='Conv2D',
            spectral_norm=False, gaussian_noise=False, plus=False):
        super(RRDB, self).__init__()
        # This is for backwards compatibility with existing models
        if nr == 3:
            self.RDB1 = ResidualDenseBlock_5C(nf, kernel_size, gc, stride, bias, pad_type,
                    norm_type, act_type, mode, convtype, spectral_norm=spectral_norm,
                    gaussian_noise=gaussian_noise, plus=plus)
            self.RDB2 = ResidualDenseBlock_5C(nf, kernel_size, gc, stride, bias, pad_type,
                    norm_type, act_type, mode, convtype, spectral_norm=spectral_norm,
                    gaussian_noise=gaussian_noise, plus=plus)
            self.RDB3 = ResidualDenseBlock_5C(nf, kernel_size, gc, stride, bias, pad_type,
                    norm_type, act_type, mode, convtype, spectral_norm=spectral_norm,
                    gaussian_noise=gaussian_noise, plus=plus)
        else:
            RDB_list = [ResidualDenseBlock_5C(nf, kernel_size, gc, stride, bias, pad_type,
                                              norm_type, act_type, mode, convtype, spectral_norm=spectral_norm,
                                              gaussian_noise=gaussian_noise, plus=plus) for _ in range(nr)]
            self.RDBs = nn.Sequential(*RDB_list)

    def forward(self, x):
        if hasattr(self, 'RDB1'):
            out = self.RDB1(x)
            out = self.RDB2(out)
            out = self.RDB3(out)
        else:
            out = self.RDBs(x)
        return out * 0.2 + x


class ResidualDenseBlock_5C(nn.Module):
    """
    Residual Dense Block
    The core module of paper: (Residual Dense Network for Image Super-Resolution, CVPR 18)
    Modified options that can be used:
        - "Partial Convolution based Padding" arXiv:1811.11718
        - "Spectral normalization" arXiv:1802.05957
        - "ICASSP 2020 - ESRGAN+ : Further Improving ESRGAN" N. C. 
            {Rakotonirina} and A. {Rasoanaivo}
    """

    def __init__(self, nf=64, kernel_size=3, gc=32, stride=1, bias=1, pad_type='zero',
            norm_type=None, act_type='leakyrelu', mode='CNA', convtype='Conv2D',
            spectral_norm=False, gaussian_noise=False, plus=False):
        super(ResidualDenseBlock_5C, self).__init__()

        self.noise = GaussianNoise() if gaussian_noise else None
        self.conv1x1 = conv1x1(nf, gc) if plus else None

        self.conv1 = conv_block(nf, gc, kernel_size, stride, bias=bias, pad_type=pad_type,
            norm_type=norm_type, act_type=act_type, mode=mode, convtype=convtype,
            spectral_norm=spectral_norm)
        self.conv2 = conv_block(nf+gc, gc, kernel_size, stride, bias=bias, pad_type=pad_type,
            norm_type=norm_type, act_type=act_type, mode=mode, convtype=convtype,
            spectral_norm=spectral_norm)
        self.conv3 = conv_block(nf+2*gc, gc, kernel_size, stride, bias=bias, pad_type=pad_type,
            norm_type=norm_type, act_type=act_type, mode=mode, convtype=convtype,
            spectral_norm=spectral_norm)
        self.conv4 = conv_block(nf+3*gc, gc, kernel_size, stride, bias=bias, pad_type=pad_type,
            norm_type=norm_type, act_type=act_type, mode=mode, convtype=convtype,
            spectral_norm=spectral_norm)
        if mode == 'CNA':
            last_act = None
        else:
            last_act = act_type
        self.conv5 = conv_block(nf+4*gc, nf, 3, stride, bias=bias, pad_type=pad_type,
            norm_type=norm_type, act_type=last_act, mode=mode, convtype=convtype,
            spectral_norm=spectral_norm)

    def forward(self, x):
        x1 = self.conv1(x)
        x2 = self.conv2(torch.cat((x, x1), 1))
        if self.conv1x1:
            x2 = x2 + self.conv1x1(x)
        x3 = self.conv3(torch.cat((x, x1, x2), 1))
        x4 = self.conv4(torch.cat((x, x1, x2, x3), 1))
        if self.conv1x1:
            x4 = x4 + x2
        x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
        if self.noise:
            return self.noise(x5.mul(0.2) + x)
        else:
            return x5 * 0.2 + x


####################
# ESRGANplus
####################

class GaussianNoise(nn.Module):
    def __init__(self, sigma=0.1, is_relative_detach=False):
        super().__init__()
        self.sigma = sigma
        self.is_relative_detach = is_relative_detach
        self.noise = torch.tensor(0, dtype=torch.float)

    def forward(self, x):
        if self.training and self.sigma != 0:
            self.noise = self.noise.to(x.device)
            scale = self.sigma * x.detach() if self.is_relative_detach else self.sigma * x
            sampled_noise = self.noise.repeat(*x.size()).normal_() * scale
            x = x + sampled_noise
        return x 

def conv1x1(in_planes, out_planes, stride=1):
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)


####################
# SRVGGNetCompact
####################

class SRVGGNetCompact(nn.Module):
    """A compact VGG-style network structure for super-resolution.
    This class is copied from https://github.com/xinntao/Real-ESRGAN
    """

    def __init__(self, num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu'):
        super(SRVGGNetCompact, self).__init__()
        self.num_in_ch = num_in_ch
        self.num_out_ch = num_out_ch
        self.num_feat = num_feat
        self.num_conv = num_conv
        self.upscale = upscale
        self.act_type = act_type

        self.body = nn.ModuleList()
        # the first conv
        self.body.append(nn.Conv2d(num_in_ch, num_feat, 3, 1, 1))
        # the first activation
        if act_type == 'relu':
            activation = nn.ReLU(inplace=True)
        elif act_type == 'prelu':
            activation = nn.PReLU(num_parameters=num_feat)
        elif act_type == 'leakyrelu':
            activation = nn.LeakyReLU(negative_slope=0.1, inplace=True)
        self.body.append(activation)

        # the body structure
        for _ in range(num_conv):
            self.body.append(nn.Conv2d(num_feat, num_feat, 3, 1, 1))
            # activation
            if act_type == 'relu':
                activation = nn.ReLU(inplace=True)
            elif act_type == 'prelu':
                activation = nn.PReLU(num_parameters=num_feat)
            elif act_type == 'leakyrelu':
                activation = nn.LeakyReLU(negative_slope=0.1, inplace=True)
            self.body.append(activation)

        # the last conv
        self.body.append(nn.Conv2d(num_feat, num_out_ch * upscale * upscale, 3, 1, 1))
        # upsample
        self.upsampler = nn.PixelShuffle(upscale)

    def forward(self, x):
        out = x
        for i in range(0, len(self.body)):
            out = self.body[i](out)

        out = self.upsampler(out)
        # add the nearest upsampled image, so that the network learns the residual
        base = F.interpolate(x, scale_factor=self.upscale, mode='nearest')
        out += base
        return out


####################
# Upsampler
####################

class Upsample(nn.Module):
    r"""Upsamples a given multi-channel 1D (temporal), 2D (spatial) or 3D (volumetric) data.
    The input data is assumed to be of the form
    `minibatch x channels x [optional depth] x [optional height] x width`.
    """

    def __init__(self, size=None, scale_factor=None, mode="nearest", align_corners=None):
        super(Upsample, self).__init__()
        if isinstance(scale_factor, tuple):
            self.scale_factor = tuple(float(factor) for factor in scale_factor)
        else:
            self.scale_factor = float(scale_factor) if scale_factor else None
        self.mode = mode
        self.size = size
        self.align_corners = align_corners

    def forward(self, x):
        return nn.functional.interpolate(x, size=self.size, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners)

    def extra_repr(self):
        if self.scale_factor is not None: