use torch.split for flat_to_shape, since it's slightly faster than tensor slicing, and
update documentation for sdeint and sdeint_adjoint.
This PR is mostly trigger by #61, adding some minor changes that I had in my in mind for a while.
Even though logqp is removed from the documentation, I still plan to support logqp in the API, merely for good backward compatibility. The concrete plan there is to have a wrapper that creates a new set of f and g from existing ones in the SDE object passed in. That should come in a future PR and shouldn't be complicated as a change.
Minor changes
latent_sde.py
inexamples
,torch.split
forflat_to_shape
, since it's slightly faster than tensor slicing, andsdeint
andsdeint_adjoint
.This PR is mostly trigger by #61, adding some minor changes that I had in my in mind for a while.
Even though
logqp
is removed from the documentation, I still plan to supportlogqp
in the API, merely for good backward compatibility. The concrete plan there is to have a wrapper that creates a new set off
andg
from existing ones in the SDE object passed in. That should come in a future PR and shouldn't be complicated as a change.