XNet/lib/modules/frcab.py

import torch
import torch.nn as nn


class FFTRCAB(nn.Module):

    def __init__(self, dim):
        super(FFTRCAB, self).__init__()

        self.CBG3x3 = nn.Sequential(
            nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, bias=False),
            nn.LeakyReLU(0.1, inplace=True),
            nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, bias=False),
        )

        self.avg_pool = nn.AdaptiveAvgPool2d(1)

        self.xc_aEnhance = nn.Sequential(
            nn.Conv2d(dim, dim // 2 + 1, 1, 1, 0),
            nn.LeakyReLU(0.1, inplace=True),
            nn.Conv2d(dim // 2 + 1, dim // 2 + 1, 1, 1, 0),
        )

        self.xc_pEnhance = nn.Sequential(
            nn.Conv2d(dim, dim // 2 + 1, 1, 1, 0),
            nn.LeakyReLU(0.1, inplace=True),
            nn.Conv2d(dim // 2 + 1, dim // 2 + 1, 1, 1, 0),
        )

    def forward(self, x):
        x_conv = self.CBG3x3(x)
        x_conv = x_conv.to(torch.float32)

        x_pool = self.avg_pool(x_conv)

        xc_a = self.xc_aEnhance(x_pool)
        xc_p = self.xc_pEnhance(x_pool)

        x_fft = torch.fft.rfft2(x_pool, dim=1, norm="ortho")
        x_a = torch.abs(x_fft)
        x_p = torch.angle(x_fft)

        xa_enh = x_a * xc_a
        xp_enh = x_p * xc_p

        xa = xa_enh * torch.cos(xp_enh)
        xp = xa_enh * torch.sin(xp_enh)
        x_comp = torch.complex(xa, xp)

        xc = torch.fft.irfft2(x_comp, dim=1, norm="ortho")

        x_out = x_conv * xc

        return x_out + x


if __name__ == "__main__":
    input_tensor = torch.randn(3, 64, 128, 128)

    fft_rcab = FFTRCAB(64)

    output_tensor = fft_rcab(input_tensor)

    print(output_tensor.shape)