v0.13 deprecations

[CarloLucibello]

• Removing datasets (have been deprecated for a year or so)
• unexporting flatten and params
• fixes a doctest issue failing doc tests #1739
• removed deprecated getproperty for recurrent layers.
• deprecated weight keyword argument in Conv constructor

I guess the only controversial point is if we want to unexport params. That could cause some (maybe a lot) of breakage, although we have been using it mostly in a qualified way in docs and model-zoo examples. On one hand, I wish it was never exported (it's a generic and common name, and it clashes with the exported Distributions.params). On the other hand, I'm not sure it is worth causing breakage.

Replaces #1735, Fix #1739, fix #629, fix #1779

[mcabbott]

There's no way to make a 12.X release with a depwarn for Flux.params, right?

[mcabbott]

We could remove deprecations for Dense's getproperty, ref FluxML/Zygote.jl#1094 (comment) . And constructors too.

[CarloLucibello]

We could remove deprecations for Dense's getproperty, ref FluxML/Zygote.jl#1094 (comment) . And constructors too.

removed all v0.12 deprecations introduced more than 6 months ago

[CarloLucibello]

There's no way to make a 12.X release with a depwarn for Flux.params, right?

not that I know

[CarloLucibello]

@mcabbott there is also this

Flux.jl/src/losses/utils.jl

Line 26 in f8ebead

# This can be made an error in Flux v0.13, for now just a warning

We have to decide if we want to make it an error already. It's been there 4 months, maybe we can wait the next release.

[mcabbott]

Could go either way on that. As a middle path, can upgrade @warn to @error to make it louder?

Other things from a quick look today...

• params got a docstring 5 minutes ago Add docstring for params #1786 but this should say Flux.params

• Juno Remove Juno.jl dependency #1779 -- if making a breaking release, could we just kill this, and add back other progress bars later, if someone is inspired?

• Parallel Parallel edge-cases #1685 via Make Parallel Vararg #1698 or something lighter which just changes it to always call the combinator exactly once, nothing else.

• Channels with Pairs, Print channel dimensions of Dense like those of Conv #1658. I share your discomfort with having two ways, but do think this would be clearer. Perhaps we should just bite the bullet and move the , method to depreciations & out of documentation, even if we never delete it.

• Some layer constructors have accumulated redundant methods. Letting Conv((2,-99), 3=>4; weight=rand(2,3,4,5)) work without error doesn't sound like a good design, but removing it is breaking.

• Instead of replace Base.tanh with faster tanh #1272 we can now use NNlib.fast_tanh. Not certain whether making layers use this counts as a breaking change, changes should be tiny floating point effects. This is now Use faster activation functions #1837 .

[ToucheSir]

No. 1 [Edit: surely 2, Juno] is already covered by #1643, just needs a quick bump to resolve VCS conflicts.

[mkschleg]

I didn’t see this in the list, but I think we wanted to remove the getproperty for recur as well for v0.13.

[CarloLucibello]

Implemented two of the @mcabbott's suggestions

• Remove Juno.jl dependency Remove Juno.jl dependency #1779 -- if making a breaking release, could we just kill this, and add back other progress bars later, if someone is inspired?
• Some layer constructors have accumulated redundant methods. Letting Conv((2,-99), 3=>4; weight=rand(2,3,4,5)) work without error doesn't sound like a good design, but removing it is breaking.

@mcabbott you want to add the other stuff to the v0.13 milestone?

PS I don't know why CI is being stupid complaining about the Manifest

[mcabbott]

For other changes:

I had a commit trying out fast_tanh on top of #1761, which we should merge. Some quite large forward speedups, modest with the gradient though, IIRC.

Parallel would also be good to sort out, haven't looked lately.

Edit: not on the list above, but

• perhaps destructure & trainable Have destructure return only trainable params #1742 should get a look?

• perhaps we should switch Chain to use foldl as in Significant compile time latency in Flux with default optimization Zygote.jl#1126. There is a chance this may break 2nd derivatives; if we don't have any in the tests, then perhaps that's not a disaster -- it's a breaking change, and this will eventually be fixed. This is now Replace unrolled foldl used to evaluate Chain with a better one #1809.

Also, Juno is loaded here: I pushed 2 commits here, hope you don't mind.

[CarloLucibello]

ok I think this is good to go

[mcabbott]

• perhaps destructure & trainable Have destructure return only trainable params #1742 should get a look?

One possibly evil idea is just to remove destructure since it's a breaking release.

Then we can make a PR adding it back, which must pass reasonably quality tests, i.e. It must have clear semantics, described in the docs; it must have tests, covering just about everything, and some attempt to invent adversarial examples. If you must have the old one, don't upgrade, until the new one is ready.

[darsnack]

Isn't destructure critical for SciML stuff? Seems like a large "don't upgrade" umbrella. Instead of removing it, if we want to skip addressing it on this release, then let's just agree that a fast v0.14 is fine too (like @ToucheSir suggested on Slack).

[mcabbott]

Well in my evil plan this desire to upgrade will inspire SciML types who need it to chip in to figuring out the solid PR adding this functionality -- to approximately ChainRules's standards. Instead of the thing being grandfathered in and intractable to work on without upsetting someone. But possibly this is too weird.

Edit: And perhaps it should move, too: FluxML/Functors.jl#5

[mcabbott]

Since #1859 is merged, perhaps master is now sort-of 0.13-dev, and we should start merging things.

[ToucheSir]

Another PR I'd like to submit for consideration is #1778. Making that non-breaking is going to be more of a pain than the alternative.

[codecov-commenter]

Codecov Report

Merging #1751 (019f832) into master (9b21e2c) will increase coverage by 9.97%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1751      +/-   ##
==========================================
+ Coverage   74.50%   84.47%   +9.97%     
==========================================
  Files          28       21       -7     
  Lines        1706     1475     -231     
==========================================
- Hits         1271     1246      -25     
+ Misses        435      229     -206     

Impacted Files	Coverage Δ
src/deprecations.jl	0.00% <ø> (-22.23%)	⬇️
src/functor.jl	86.66% <ø> (ø)
src/layers/recurrent.jl	85.54% <ø> (+9.86%)	⬆️
src/optimise/optimisers.jl	93.64% <ø> (ø)
src/layers/basic.jl	80.53% <100.00%> (+3.60%)	⬆️
src/layers/conv.jl	80.54% <100.00%> (+0.43%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 9b21e2c...019f832. Read the comment docs.