search

jingyuanli001 / RFR-Inpainting

The source code for CVPR 2020 accepted paper "Recurrent Feature Reasoning for Image Inpainting"
MIT License
355 stars 76 forks source link

CUB200 training result issue #38

Closed cccxyc closed 3 years ago

cccxyc commented 3 years ago

image image hi there I trained bird dataset for 800000 iters，yet the results remain blur. Does this have somthing todo with my mask setting? I opened mask with mode='RGB' and deleted the code 'masks = torch.cat([masks]*3, dim = 1)'

jingyuanli001 commented 3 years ago

I would assume this is due to some issues in the input mask format. Opening the mask/image with different mode is dangerous as it changes the range of value, e.g. the mask image could either be within range 0~1 (gray scale) or range 0~255 (8 bit). This then influence the values of the output images. One rule-of-thumb is to ensure that the masks used in generating composed image in model.py have only 0s and 1s in it. Otherwise things could be wired.

cccxyc commented 3 years ago

will, i did set mask.jpg with only 0 and 1 pixel. the celeba dataset generates better results at samilar iters. i guess it depends on the complicity of trainiing images