search

karpathy / pytorch-normalizing-flows

Normalizing flows in PyTorch. Current intended use is education not production.
845 stars 98 forks source link

2D -> ND #3

Open sadegh-aa opened 4 years ago

sadegh-aa commented 4 years ago

Thanks for sharing a fairly complete repo of NFs. What kind of assumptions the code considers in the case of 2D that does not hold in ND? I have a dataset of 100D and the loss is nan all the time.

sadegh-aa commented 4 years ago

I guess this is related to the ActNorm which becomes nan. Also there is relation between the batch size and the dimension of the data. Do you have any comment on that?

kfzyqin commented 4 years ago

For nan, you may try decreasing batch size.