octave conv cannot support input at any size

[Stinky-Tofu]

If the network is downsampled n times, the input size must be 2^a (a >= n), other sizes are not supported, such as 600x1000.

[d-li14]

Do you mean the tensor will not restore to its original size via upsampling if the side of the tensor is odd? If so, manually padding one pixel may be one possible solution. Hopefully, it will not have much negative influence.

[Stinky-Tofu]

@d-li14 是的，我写了一点代码解决了

class padToSameSize(nn.Module):
    def __init__(self):
        super(padToSameSize, self).__init__()

    def forward(self, lTensor, rTensor):
        hwOfLTensor = np.array(lTensor.size()[2:], dtype=int)
        hwOfRtensor = np.array(rTensor.size()[2:], dtype=int)
        maxHW = np.max([hwOfLTensor, hwOfRtensor], axis=0)
        padHWOfLTensor = maxHW - hwOfLTensor
        padHWOfRTensor = maxHW - hwOfRtensor
        lTensor = F.pad(lTensor, pad=[int(padHWOfLTensor[1]), 0, int(padHWOfLTensor[0]), 0])
        rTensor = F.pad(rTensor, pad=[int(padHWOfRTensor[1]), 0, int(padHWOfRTensor[0]), 0])
        return lTensor, rTensor

[d-li14]

@Stinky-Tofu Good job, exactly an effective workaround here!