【CNN】——双线性插值和conv_transpose2d

conv transpose2d
code from:mmflow

双线性插值转成反卷积操作

class Upsample(nn.Module):
    """Upsampling module.

    Args:
        scale_factor (int): Scale factor of upsampling.
        channels (int): Number of channels of conv_transpose2d.
    """

    def __init__(self, scale_factor: int, channels: int) -> None:
        super().__init__()
        self.kernel_size = 2 * scale_factor - scale_factor % 2
        self.stride = scale_factor
        self.pad = math.ceil((scale_factor - 1) / 2.)
        self.channels = channels
        self.register_buffer('weight', self.bilinear_upsampling_filter())

    # caffe::BilinearFilter
    def bilinear_upsampling_filter(self) -> torch.Tensor:
        """Generate the weights for caffe::BilinearFilter.

        Returns:
            Tensor: The weights for caffe::BilinearFilter
        """
        f = math.ceil(self.kernel_size / 2.)
        c = (2 * f - 1 - f % 2) / 2. / f
        weight = torch.zeros(self.kernel_size**2)
        for i in range(self.kernel_size**2):
            x = i % self.kernel_size
            y = (i / self.kernel_size) % self.kernel_size
            weight[i] = (1 - abs(x / f - c)) * (1 - abs(y / f - c))
        return weight.view(1, 1, self.kernel_size,
                           self.kernel_size).repeat(self.channels, 1, 1, 1)

    def forward(self, data: torch.Tensor) -> torch.Tensor:
        """Forward function for upsample.

        Args:
            data (Tensor): The input data.

        Returns:
            Tensor: The upsampled data.
        """
        return F.conv_transpose2d(
            data,
            self.weight,
            stride=self.stride,
            padding=self.pad,
            groups=self.channels)