FID is a negative value

[santiago-parodi]

I'm attempting to compute the FID between two datasets, each containing 192 samples. I used the following command:

python -m pytorch_fid --dims 192 ./real/ ./fake/

However, I'm getting a negative FID value:

FID: -4.808164533187664e-09

Any ideas of what might be happening? Thanks in advance!

[dome272]

I saw this sometimes as well. I think its numerical instabilities. When your ground_truth and fake_data are super close it would end up as 0 in theory, but due to limited numerical precision it goes slightly negative. You can see its -4e-8, which is super small.

[egoist945402376]

I saw this sometimes as well. I think its numerical instabilities. When your ground_truth and fake_data are super close it would end up as 0 in theory, but due to limited numerical precision it goes slightly negative. You can see its -4e-8, which is super small.

Thanks