Iteration over params(m) in explicit mode gives no gradient

[mcabbott]

The use of params within explicit-mode gradients has been broken by #2054 . Sadly there were no tests of this:

julia> m = (x=[1,2.0], y=[3.0]);

julia> gradient(m -> (sum(norm, Flux.params(m))), m)
((x = [0.4472135954999579, 0.8944271909999159], y = [1.0]),)  # before, Flux@0.13.5
(nothing,)  # after, Flux@0.13.6

julia> gradient(() -> (sum(norm, Flux.params(m))), Flux.params(m))
Grads(...)

julia> ans[m.x]  # unchanged
2-element Vector{Float64}:
 0.4472135954999579
 0.8944271909999159

(This form really is very annoying. Suggestions are fine but required textboxes seem to be solving a problem we didn't have. Also, it seems to break links to issues, like #2054 above. Edited to delete the boilerplate.)

[ToucheSir]

How much do we care about this working? I'm not sure I have the wherewithal to figure it out while not regressing Flux.

This form really is very annoying. Suggestions are fine but required textboxes seem to be solving a problem we didn't have...

We (or really github) provide an escape hatch. You can still open a blank issue via the " Don’t see your issue here? Open a blank issue." link at the bottom of the new issue page.

[mcabbott]

IDK, it's the officially documented way to add a regularisation term, http://fluxml.ai/Flux.jl/stable/models/regularisation/ .

Ideally that would move to something like FluxML/Optimisers.jl#57 . But if the goal is to provide a soft transition to explicit mode, so you can add the new things slowly during 0.13 rather than all-at-once to run on 0.14, then I think the old way ought to keep working a bit longer?

[ToucheSir]

Is there a smart rule we could write which avoids #2054? I think the main considerations would be how map trainable pullbacks to the original tree structure and the final params vector. Maybe maintaining an auxiliary tree structure with indices at the leaves?

[mcabbott]

This is surely possible, someone just has to write one more Functors thing which builds this gradient tree, as the rrule for params(m).