search

WillSuen / LostGANs

51 stars 36 forks source link

FID score #13

Closed wb-finalking closed 3 years ago

wb-finalking commented 3 years ago

Hello! I am very interested in your work. And I use your pretrained model netGv2_coco128.pth for testing. But my metric scores(IS: 14.31, FID 30.71) are not match yours in paper (IS: 14.21, FID: 24.76). I want to know why there is a big difference between my FID score and yours. I use the evaluate code mentioned in https://github.com/WillSuen/LostGANs/issues/3 and generate 5 images for each layout. Looking forward to your reply.

WillSuen commented 3 years ago

Hi thanks for your interests. How many images did you use for evaluate FID? I use all about 3000 images, so the FID is calculated over about 3k vs 15k. If still have the issue, could you please upload the generated images and let me have a check

wb-finalking commented 3 years ago

Hi thanks for your interests. How many images did you use for evaluate FID? I use all about 3000 images, so the FID is calculated over about 3k vs 15k. If still have the issue, could you please upload the generated images and let me have a check

Sorry, its my fault to modify the code for image saving, which would result in information loss. When I use misc.imsave for image saving, FID score can reach to 24.8.