Compute FID with 50k training samples?

[Baoren1996]

Hi,
I not sure why you compute FID with 50k training sample rather than 10k test samples as other papers do? will this mislead to a much better FID than it should be?

[PeterouZh]

Sorry for the late reply. It is more common to calculate the FID between all training data and 50k generated data. Please refer to this site https://github.com/NVlabs/stylegan2-ada-pytorch#quality-metrics