Meesum Qazalbash
Meesum Qazalbash
An observation I wanted to share. I don't know the reason behind this behaviour. Memory footprint drastically (in GBs) increases when we slightly increase the number of global steps. This...
What is the solution?
@fehiepsi, can you look into these changes? --- @juanitorduz you helped us a lot in solving #2066, you might be interested in these changes too.
Only errors left here are coming from the statments `constrain_obj: ConstraintT = ConstraintClass(...)`, when `ConstraintClass` uses `NonScalarArray` type. Because `ConstraintT` has `NumLike` and expects similar from `ConstraintClass`. Same problem can...
Hi @fehiepsi, I am a little busy with my grad school applications, can I update you after Dec 15?
The current implementation of `TransformedDistribution` transforms the `log_prob`, `sample`, `cdf` and `icdf` methods, but it doesn't transform the support. Either we should add this functionality of transforming the support or...
I understood that I was just explaining my thoughts!
@dylanhmorris IMO, something like this would solve both problems! ```python def support(self, value: ArrayLike) -> ArrayLike: mask = jnp.ones_like(value, dtype=bool) y = value for transform in reversed(self.transforms): x = transform.inv(y)...
> Thanks @Qazalbash! I thought we can merge your changes and iterate on later PRs. Turns out that there are many non-trivial changes so it would be nice to split...
I will second that because without progress bar it is hard to know to which extend the process will run.