fld is nan

[SunnierLee]

When I was calculating the FLD, the result I got was NaN. What could be the reason for this?

[marcojira]

Hard to say, could be an issue with one of the train/test/gen sets (for example if one is very small). Do you have any more details on your use case?

[SunnierLee]

I use CelebA dataset, and the numbers of train/test/gen images are 145,064/19,962/60,000.

Could too bad generation quality cause nan? For my good baselines, the FLD is reasonable. I have a poor baseline, which generates CelebA image like this

[marcojira]

Sorry for the late reply. Yes potentially, if the features are too far from the reference sets.