About the LPIPS evaluation

[zhiyuanyou]

Hi～

Thanks for your great work!

I have a small question about the LPIPS evaluation.

Actually, you do the LPIPS evaluation with evaluate_lpips() in: https://github.com/phizaz/diffae/blob/master/experiment.py#L606. The function evaluate_lpips() is defined in:
https://github.com/phizaz/diffae/blob/master/metrics.py#L44.

Then in evaluate_lpips(), I track the input imgs with the following paths:

• https://github.com/phizaz/diffae/blob/master/metrics.py#L86
• https://github.com/phizaz/diffae/blob/master/diffusion/base.py#L691
• https://github.com/phizaz/diffae/blob/master/diffusion/base.py#L650
• https://github.com/phizaz/diffae/blob/master/diffusion/base.py#L307

I find that, finally, in p_mean_variance(), you pass x (which equals imgs and is the original image) directly into model.forward() (See: https://github.com/phizaz/diffae/blob/master/diffusion/base.py#L307).

Here I am confused, since we should input x_T instead of x (the original image) into model.forward() for evaluation.

Could you please explain it for me?

Thanks in advance.

[string-ellipses]

Hi! Thanks for your discovery. I have the same confusion. And may I ask if you have resolved it?