reproduction problem

[zhao2000hhh]

When I retrained with your training code and configuration, I couldn't produce a model as good as yours。 Specifically, I tried on the celeba-HQ dataset. mask ftom https://nv-adlr.github.io/publication/partialconv-inpainting train on 20-30% subset test on the whole mask set my model： psnr=28.5719, val: ssim=0.8786, val: lpips=0.0945 your model： PSNR = 29.581334 SSIM= 0.893 LPIPS=0.08094 Do you have any special training techniques？ Looking forward to your reply.

[NiFangBaAGe]

Thanks for your interest in our work. The masks for experiments are generated from https://nv-adlr.github.io/publication/partialconv-inpainting. We dilate or erode them to ensure the mask ratio is not more than 25% (following HiFill).

[zhao2000hhh]

Is the testing also conducted on the 25%mask？

[NiFangBaAGe]

Yes, all the training and testing masks follow this setting.

[zhao2000hhh]

thank you reply. By the way, could you share me the 25% mask?

[zhao2000hhh]

looking forward to your reply. sos

[NiFangBaAGe]

The mask is available here. Hope it can help you.

[zhao2000hhh]

good luck！