What's missing for 1.0?

[Datseris]

Hi there,
this awesome package has been around for a really long time and is a dependency of many Julia packages but still not in 1.0. From my experience, the interface has also been "stable" for a quite a while.

What is missing? What breaking changes are planned that keep this package obtaining its first stable version? Do you have a list of these somewhere? If not, can you please state them here, or somewhere else, to give the developers that use this package an idea?

[matbesancon]

@JuliaRegistrator register()

[JuliaRegistrator]

Registration pull request created: JuliaRegistries/General/26483

After the above pull request is merged, it is recommended that a tag is created on this repository for the registered package version.

This will be done automatically if the Julia TagBot GitHub Action is installed, or can be done manually through the github interface, or via:

git tag -a v0.24.7 -m "<description of version>" 530b9e0fe2d07e9a8082eaefdbc42fdcab3b7474
git push origin v0.24.7

[Datseris]

@matbesancon I'm not sure whether I should be able to deduce something here, but I can't.

[matbesancon]

@Datseris sorry no I just used the issue opened to register a version

[Datseris]

Oh okay. BTW you can also tag releases by commenting on commits. I typically find this cleaner on the long term, as the comment that tags the release is directly associated the commit this release should be done on.

[drrmmng]

Regarding the original question:
I believe this package needs more documentation.

For example, I was trying to figure out if I can backpropagate through rand with respect to the distribution parameters in Flux. That is relevant for things like variational inference. After working my way through the code I'm pretty convinced that this is possible, at least for the multivariate normal distribution.

Are there plans to improve the documentation and if so, how could I help?

[matbesancon]

I will close this issue as a duplicate of #880

I believe this package needs more documentation.

Completely agreed, more and better.

For example, I was trying to figure out if I can backpropagate through rand with respect to the distribution parameters in Flux

Through other functions yes, but through rand I am not sure this is doable? rand is not even a function in the mathematical sense, or if you consider rand(rng, d), it is mutating its first argument in a non-differentiable way

[matbesancon]

Are there plans to improve the documentation and if so, how could I help?

No plans yet, but whenever you see things that could be improved a PR is useful

[drrmmng]

Through other functions yes, but through rand I am not sure this is doable? rand is not even a function in the mathematical sense, or if you consider rand(rng, d), it is mutating its first argument in a non-differentiable way

Good point! What I was trying to do with this package is called the reparameterization trick.

In case of the MvNormal distribution that seems to be possible, however, my problem was that there is nothing like a batch dimension for distribution along the line of PyTorch's distributions. This made it impossible to create a batch of MvNormal distributions by passing a 2d array, which is necessary for CUDA arrays, since they discourage scalar access.