search

WendongZh / SPL

[IJCAI'21] Code for Context-Aware Image Inpainting with Learned Semantic Priors,
53 stars 4 forks source link

About load the pretrained TR_resNet? #4

Closed crazygirl1992 closed 3 years ago

crazygirl1992 commented 3 years ago

Hi,i run your code and i find a problem, It is what to choice the TR_ReNet? There are different models,and when i choice the 448*448 on VOC,the errors happen. 图片

WendongZh commented 3 years ago

Thanks for your interest!

As we mentioned in the ReadMe, we use the TResNet_L architecture pretrained on OpenImage dataset.

If you want to change the pretrained model, you can select the same architecture (TResNet_L). If you want to use other architecture such as TResNet_XL on VOC, additional modifications should be considered such as feature map channels in our distill head and (maybe) other particular layers in TResNet_XL with the same output size.