PixelShuffle 实现模型上下采样

PixelShuffle 和 UnpixelShuffle 减少尺寸的同时也会改变通道数

1 32 128 128 -- UnpixelShuffle（2）-> 1 128 64 64

【torch.nn.PixelShuffle】和 【torch.nn.UnpixelShuffle】-优快云博客

效果还行 可惜硬件不支持

        if use_PixelShuffle:
            print('use_PixelShuffle')
        ##############################  PixelShuffle  ####################################
            self.down2 = nn.Sequential(
                nn.PixelUnshuffle(2) , #1 32 128 128 -> 1 128 64 64
                nn.Conv2d(feat_top*4, feat_mid, 1, stride=1, padding=0, bias=bias)
            )
            self.down4 = nn.Sequential(
                nn.PixelUnshuffle(2) ,
                nn.Conv2d(feat_mid*4, feat_bot, 1, stride=1, padding=0, bias=bias)
            )
            self.up2 = nn.Sequential(nn.Conv2d(feat_bot, feat_mid*4, 1, stride=1, padding=0, bias=bias),
                                      nn.PixelShuffle(2))
            self.up4 = nn.Sequential(nn.Conv2d(feat_mid, feat_top*4, 1, stride=1, padding=0, bias=bias),
                                      nn.PixelShuffle(2))
        #################################################################################