2023ICLR_SFNet: Selective frequency network for image restoration

1. 在运行SFNet代码时，前后代码保持不变，运行两次结果发生变化, 把下面这段代码注掉就可以保持前后两次运行结果一致，不确定是否是nn.BatchNorm2d计算均值和方差导致

class dynamic_filter(nn.Module):
    def __init__(self, inchannels, mode, kernel_size=3, stride=1, group=8):
        super(dynamic_filter, self).__init__()
        self.stride = stride
        self.kernel_size = kernel_size
        self.group = group

        self.lamb_l = nn.Parameter(torch.zeros(inchannels), requires_grad=True)
        self.lamb_h = nn.Parameter(torch.zeros(inchannels), requires_grad=True)

        self.conv = nn.Conv2d(inchannels, group*kernel_size**2, kernel_size=1, stride=1, bias=False)
        self.bn = nn.BatchNorm2d(group*kernel_size**2)
        self.act = nn.Softmax(dim=-2)
        nn.init.kaiming_normal_(self.conv.weight, mode='fan_out', nonlinearity='relu')

        self.pad = nn.ReflectionPad2d(kernel_size//2)

        self.ap = nn.AdaptiveAvgPool2d((1, 1))
        self.modulate = SFconv(inchannels, mode)

    def forward(self, x):
        identity_input = x 
        low_filter = self.ap(x)
        low_filter = self.conv(low_filter)
        low_filter = self.bn(low_filter)     

        n, c, h, w = x.shape  
        x = F.unfold(self.pad(x), kernel_size=self.kernel_size).reshape(n, self.group, c//self.group, self.kernel_size**2, h*w)

        n,c1,p,q = low_filter.shape
        low_filter = low_filter.reshape(n, c1//self.kernel_size**2, self.kernel_size**2, p*q).unsqueeze(2)
       
        low_filter = self.act(low_filter)
    
        low_part = torch.sum(x * low_filter, dim=3).reshape(n, c, h, w)

        out_high = identity_input - low_part
        out = self.modulate(low_part, out_high)
        return out