More restrictive @non_differentiable rand

[nmheim]

Addresses #262. I am really not sure if this is a good approach because there are quite a few different rand methods...
This is what works with the use cases I have. What are the most common uses of rand that should be non-differentiable?

[oxinabox]

Why not just add:

@non_differentiable rand()
@non_differentiable rand(::Any, ::Any, ::Any, ::Any, ::Any, ::Any)

?

rand(::Any...) is in the list of things to add when we have varargs support.
#253

It could also be added with varargs directly, without using the macro.

[willtebbutt]

@oxinabox see #262?

[oxinabox]

Oh i was confused by the title.

[oxinabox]

I think all the Int should be Integer
to match:

[69] rand(X, d::Integer, dims::Integer...) in Random at /usr/local/src/julia/julia-1.5/usr/share/julia/stdlib/v1.5/Random/src/Random.jl:283

[oxinabox]

Can we add a test that makes sure
rrule and frule for something that we don't want to have mark as @non_differentiable returns nothing (i.e. no rule)

[oxinabox]

I think we might also in particular want:

@non_differentiable rand(::AbstractRNG, ::Random.Sampler)
@non_differentiable rand(::AbstractRNG, ::Tuple)

[nmheim]

Let me know what you think about my take on testing the new rules. I am really not sure if this is how it should be done. Especially because the rrules and frules return Zero or DoesNotExist...

[oxinabox]

Nice.
A few suggestions,
once those are addressed and the version number bumped we can merge this and tag a release

[willtebbutt]

Just a couple of style things.

Once @oxinabox 's suggestions and these are attended to, please bump the patch version and we'll merge when tests pass :)

Thanks for the contribution -- it's much appreciated!

[nmheim]

Great, thanks for the suggestions! I think the tests are failing because of something unrelated to this PR though... they also fail in #261 because of something getindex related...

[willtebbutt]

Great, thanks for the suggestions! I think the tests are failing because of something unrelated to this PR though... they also fail in #261 because of something getindex related...

Oh, yeah, don't worry about the Zygote integration tests -- Zygote master is currently broken.

[nmheim]

Actually, I am not sure about the comment on the test of the "differentiable" cases.
I used rand(ones(2,2)) as a test because we don't need an additional dependency like this, but I don't know why rand of an array works at all...
A better test case would be to use e.g. a distribution. Should I add that?

[oxinabox]

A better test case would be to use e.g. a distribution. Should I add that?

Yes.
Or maybe a fake distribution by defining our own struct,
and putting it in the same position a distribution argument would go.
(that way we avoid the test dependency)

[nmheim]

I am sorry, but I am a bit clueless how to test the rules on fields of a struct if they don't appear in the rand function call... this is probably not how it should be done...

[oxinabox]

I am going to add my suggestions and then if i didn't break CI merge this

[nmheim]

Awesome, thank you! :)