How to apply this function to 1channel image?

[schliezer]

Thank you so much for your contribution! I came across a problem when comparing two 1channel images. The traceback threw out a 'non-singleton dimension dismatch' at line 154 'return (inp - self.shift) / self.scale'. Later I learned that this function works for 3channel images, but I am not sure if that is the reason in my case (input/target shape(B,C,H,W): (4,1,256,256)), if it is, then how can I convert to 3channel? Thanks!

[richzhang]

Repeat the channel 3x to make it an RGB image

[iris0329]

Hi @richzhang
Thanks for your contribution, Could you give me some ideas on how to change shift and scale if I want to apply it to my own data, which is a 5-channel image?

[richzhang]

The method is for RGB images. It somewhat makes sense for "1 channel" grayscale images, but doesn't make sense to run for a 5 channel image.

[iris0329]

I see now, thanks a lot

[Mally-cj]

I am courious about how to determine the value of self.shift and self.scale . And as I comparing two 1channel images，can I delet the ScalingLayer？What happen if I do?

[Mally-cj]

Thank you so much for your contribution! I came across a problem when comparing two 1channel images. The traceback threw out a 'non-singleton dimension dismatch' at line 154 'return (inp - self.shift) / self.scale'. Later I learned that this function works for 3channel images, but I am not sure if that is the reason in my case (input/target shape(B,C,H,W): (4,1,256,256)), if it is, then how can I convert to 3channel? Thanks!

I also meet this question and I have solve it.
As I use LPIPSWithDiscriminator to encounter this problem，I find that adding a line of code can solve it in this folling code can solve it. Because LPIPS only calculates a value of 1 * 1, the input can be a tensor of batch size * 1 * size * size. Therefore, the reason for the error is that it affects the code of other parts.