Laurent Mazare

Thanks for the PR, it seems that PARQUET_2_6 though is only available in arrow 6.0 and not in arrow 5 for which we currently have most of our production usage....

The old glib bit is deprecated, I'll clean that up. The currently active part is under c_api (plus the ppx) and should work with arrow 4.0.0 and 4.0.1 (the currently...

Sure, I've made a PR to update the opam version, you can follow it [here](https://github.com/ocaml/opam-repository/pull/21573).

Thanks for the suggestion. This indeed seems like a nice thing to add - the underlying torch primitives for sampling should already exist in the rust wrappers, e.g. [normal](https://docs.rs/tch/0.0.8/tch/struct.Tensor.html#method.normal) so...

Do you need to backprop through your KL divergence ? If that's the case I'm not sure that rv could be used out of the box but maybe I'm missing...

Re NotImplementedError, indeed that's the base class for distributions in pytorch so nothing is implemented here and the various methods are implemented in the derived classes like 'Normal'. The distribution...

Just to mention that I added a variational auto-encoder example to tch-rs. This inctludes a KL divergence loss [here](https://github.com/LaurentMazare/tch-rs/blob/master/examples/vae/main.rs#L60). It's certainly very far away from what a nice `distributions` api...

@vegapit yes that's mostly about adding such functions and probably some traits for the various distributions. You can see the implementation for the normal distribution in the python api [here](https://pytorch.org/docs/stable/_modules/torch/distributions/normal.html#Normal)....

Btw the error function is also available in tch-rs https://docs.rs/tch/0.1.0/tch/struct.Tensor.html#method.erf (as the rust low level wrappers are automatically generated, we mostly get these for free)

Yes that kind of trait would indeed probably be useful, @vegapit do you think this would cover your use case?