LPIPS loss

[ivashmak]

When using the LPIPS loss, as exemplified in:
https://github.com/amundra15/livehand/blob/release/models/loss.py#L45
https://github.com/amundra15/livehand/blob/release/models/loss.py#L57

LPIPS expects images to be within the -1 to 1 range. Within the forward function, there is a normalize flag that is by default set to False. It normalizes the images to the -1 to 1 range:

        if normalize: # turn on this flag if input is [0,1] so it can be adjusted to [-1, +1]
            in0 = 2 * in0  - 1
            in1 = 2 * in1  - 1

However, in your case, images (rgb, rgb_sr, target) are within the 0 to 1 range. And you did not set normalize flat to True. Is this intentional, or was it perhaps overlooked?

[amundra15]

Hi @ivashmak , you are right. This was overlooked during the implementation.

[amundra15]

In case you experiment with correct normalization, I would be curious to look at the results.

[ivashmak]

Thanks for reply. Sure, I will let you know if I do this.