UserWarning: nn.Upsample is deprecated. about pytorch-gan HOT 4 CLOSED

Haven't had time to add this yet. Have to decide if I want to add this custom layer to each implementation or if I should add a common python file with custom layers. If you want to get rid of this message you can use this Upsample module.

class Upsample(nn.Module):
    """ nn.Upsample is deprecated """

    def __init__(self, scale_factor, mode="nearest"):
        super(Upsample, self).__init__()
        self.scale_factor = scale_factor
        self.mode = mode

    def forward(self, x):
        x = F.interpolate(x, scale_factor=self.scale_factor, mode=self.mode)
        return x

from pytorch-gan.

Actually this is not an error, it's just a warning. You don't need to worry about it.

If you want, you can just change the torch.nn.Upsample() layer for the function torch.nn.functional.interpolate() and choose the same interpolation method to have same behavior, e.g., change torch.nn.Upsample(mode='nearest', scale_factor=2)(x_in) to torch.nn.functional.interpolate(x_in, mode='nearest', scale_factor=2).

The problem is that nn.function.interpolate does not have bicubic interpolation implementation. In this case I transform the tensor back to numpy array and use any existing implemantion for it (like the PIL.Image.resize), then transform it back to Tensor (not sure if it is intelligent, but does the trick). There is a nice dicussion about this change here.

from pytorch-gan.

Just like @alsombra said it's just a warning, so you can run the implementations without problems despite the message you got. That said, I will probably write a wrapper for torch.nn.functional.interpolate soon which will replace the nn.Upsample layers.

from pytorch-gan.

Thank you so much l got it.

from pytorch-gan.