FVD seems not normal

[Alienge]

Nice Repo! I use this Repo to calculate the FVD from generating FFS dataset. But the FVD seems not normal! Any advices? I use the following code to calculate.

evaluator = fvd.cdfvd('videomae', ckpt_path=None)
evaluator.load_videos('ffs', data_type='stats_pkl', resolution=256, sequence_length=16)
evaluator.compute_fake_stats(evaluator.load_videos('./test/',data_type="video_folder"))
score = evaluator.compute_fvd_from_stats()
print(score)

The result is -1.9188070608923394e+28

[songweige]

Hi @Alienge, how many samples did you use to compute the FVD?

It seems that there is a numerical error where it could happen in the covariance matric inversion step as it didn't get even samples.

[Alienge]

I use the 2048 samples to compute the FVD. I also use the "i3d", it seems normal. I had check the value from scripts, abnormal seems come from $sqrt(cov1*cov2)$ code

[Alienge]

Hi @Alienge, how many samples did you use to compute the FVD?

It seems that there is a numerical error where it could happen in the covariance matric inversion step as it didn't get even samples.

Solution: change scipy == 1.12.0 to scipy ==1.10.0.

[songweige]

Thank you for digging into this and sharing what you found!!

I suspect this was because of a loose check in real->complex conversion for Schur decomposition, which only happens in 1.12.0:

https://github.com/scipy/scipy/commit/4751485b2996d8717af17d225091428d543acab3

It seems that they have improved this in the latest version 1.14.0.

https://github.com/scipy/scipy/commit/f56fe87a163cd1de957c2b891f265122fa3bbd41

I couldn't reproduce this error on my end. Would it be easy for you to test `1.14.0' as well? Thanks!! If it still causes an issue, we might want to switch to some more robust matrix square root function.

[Alienge]

I also test "scipy==1.14.0". It not appear the abnormal number. This problem is only happened in "scipy==1.12.0". If U want to test, I will upload the origin data to Google Drive to reproduce this problem.