【上采样】——深度学习上采样方式

• 最近邻上采样
• 双线性上采样
• transposed convolutions
• 光流上采样SFNet
• learned upsamplehttps://arxiv.org/pdf/2011.06961.pdf
  code

class Upsample(nn.Module):
    def __init__(self, mode, channels=None):
        super(Upsample, self).__init__()
        self.interp = nn.functional.interpolate

        if mode == 'bilinear':
            self.align_corners = False
        else:
            self.align_corners = None

        if 'learned-3x3' in mode:
            # mimic a bilinear interpolation by nearest neigbor upscaling and
            # a following 3x3 conv. Only works as supposed when the
            # feature maps are upscaled by a factor 2.

            if mode == 'learned-3x3':
                self.pad = nn.ReplicationPad2d((1, 1, 1, 1))
                self.conv = nn.Conv2d(channels, channels, groups=channels,
                                      kernel_size=3, padding=0)
            # kernel that mimics bilinear interpolation, 这相当于是一个滤波器
            w = torch.tensor([[[
                [0.0625, 0.1250, 0.0625],
                [0.1250, 0.2500, 0.1250],
                [0.0625, 0.1250, 0.0625]
            ]]])

            self.conv.weight = torch.nn.Parameter(torch.cat([w] * channels))

            # set bias to zero
            with torch.no_grad():
                self.conv.bias.zero_()

            self.mode = 'nearest'
        else:
            # define pad and conv just to make the forward function simpler
            self.pad = nn.Identity()
            self.conv = nn.Identity()
            self.mode = mode

    def forward(self, x):
        size = (int(x.shape[2]*2), int(x.shape[3]*2))
        x = self.interp(x, size, mode=self.mode,
                        align_corners=self.align_corners)
        x = self.pad(x)
        x = self.conv(x)
        return x

整体流程：

• 首先用nearest上采样
• 然后对图像的边缘用边缘值做填充
• 最后做一个滤波，滤波器的参数是可以学习