remove Datasets + additional deprecations

[CarloLucibello]

Remove Flux's datasets in favor of data providers in the julia ecosystem.

[CarloLucibello]

@DhairyaLGandhi are you ok with this change?

[DhairyaLGandhi]

Best to deprecate no?

[CarloLucibello]

problem is each dataset is a module, I searched for a while but I didn't find how to deprecate a module

[DhairyaLGandhi]

Hmm, not sure but we might end up with add depwarns to the methods that people would call for the datasets?

[CarloLucibello]

bors try

[bors]

try

Build succeeded:

• ci/gitlab/gitlab.com

[DhairyaLGandhi]

Many thanks to @KristofferC !!

 julia> module Foo
           module Bar
           end
           
           Base.@deprecate_binding Bazz Bar
        end
 Main.Foo
 
 julia> Foo.Bazz
 WARNING: Foo.Bazz is deprecated, use Bar instead.
   likely near REPL[2]:1
 Main.Foo.Bar

[DhairyaLGandhi]

We should avoid having this list here for now and it should be the synced with the list in the website. And in general we should also do a once over the packages as well

[CarloLucibello]

and it should be the synced with the list in the website. And in general we should also do a once over the packages as well

I don't understand

[CarloLucibello]

bors r+

[bors]

Build succeeded:

• ci/gitlab/gitlab.com

[DhairyaLGandhi]

Why did this not follow the deprecations route? It makes breaking API changes without the deprecations themselves allowing users to continue using the functionality.

[CarloLucibello]

I did introduce informative errors pointing users to MLDatasets, it should be very easy for them to switch. Also, recent examples in model-zoo have been relying on MLDatasets for a long time.
I can make a new PR with a softer deprecation path by importing MLDatasets here and rerouting its function to match the previous Flux api. I prefer not to though, it is a lot of work to do here, while users can overcome the change pretty easily

[DhairyaLGandhi]

I'd prefer a warning about them being removed and using the datasets we had for now, and removing them in a couple of releases.

Agree on not adding MLDatasets as a dep