Description
Hello, your paper is really good!
But for fid50k_ full.There are some problems with the evaluation index of full.
"we report the FID between 50k generated and all real images" in your paper.
However, Pokemon has only more than 800 data sets, and the number of generated images is greater than the number of real images. Is this comparison really accurate? Although the initial literature did not mention this issue[1]. However, this document mentions that generating 50k images will lack diversity[2].
We try to study this problem in full dataset(<1000 and >800), half dataset(500) and 100 dataset vs same number generated images，We find that FID decreases as the number of images increases. Of course, generated 50k images vs full dataset(<1000 and >800)is least.The evaluation method we use is [3].
[1]Heusel M , Ramsauer H , Unterthiner T , et al. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium[J]. 2017.
[2]Konstantin Shmelkov, Cordelia Schmid, Karteek Alahari.How good is my GAN[C].ECCV,2018,218-234.
[3]https://github.com/mSEitzer/pytorch-fid