Deprecate FastChain and sciml_train for v2.0 #794
Conversation
Superscedes #715
| nn = Lux.Chain(x -> x.^3, | ||
| Lux.Dense(2, 16, tanh), | ||
| Lux.Dense(16, 2)) | ||
| p_init, st = Lux.setup(rng, nn) | ||
| p_init = Lux.ComponentArray(p_init) |
There was a problem hiding this comment.
@Abhishek-1Bhatt did you test Lux with multiple shooting?
| @test loss_function(res.minimizer) < l1 | ||
| res = DiffEqFlux.sciml_train(loss_function, nODE.p, Optim.KrylovTrustRegion(), Optimization.AutoZygote(), maxiters = 100, callback=cb) | ||
| res = Optimization.solve(optprob, KrylovTrustRegion(), maxiters=100, callback=cb) | ||
| @test loss_function(res.minimizer) < l1 |
There was a problem hiding this comment.
We should probably also test second order Lux.
|
I think |
Yes that is true, but I don't want to mix too many changes at the same time. Since that's separate from DiffEqFlux itself, I'd like to hold off on that. But definitely we want to get rid of Flux.jl and Optim.jl optimizer usage. |
|
IMO, for a breaking release, it's better to resolve #744 first. and remove |
|
Yes that should be in the release too. I think it can just be removed and the tests will still pass? I guess just give it a try? If Newton and FFJORD pass then we're good. |
| import NNlib | ||
| import Lux | ||
| using Requires | ||
| using Cassette | ||
| @reexport using Flux |
There was a problem hiding this comment.
Now that we support Lux, reexporting Flux could make qualification errors in codes that only have Lux.
Error like:
WARNING: both Lux and Flux export "Chain"; uses of it in module Main must be qualified
ERROR: UndefVarError: Chain not definedThere was a problem hiding this comment.
Yeah we should remove the reexport. Add that to the list of things.
Co-authored-by: Abhishek Bhatt <46929125+Abhishek-1Bhatt@users.noreply.github.com>
|
I'm trying to just get this to mergeable now. The plan will be to hold breaking on master for a month (which we normally don't do, but we can make an exception for this big of a deprecation). We should do all of the things like remove the Zygote rule overloads (make sure they are upstreamed), setup LuxCore, update docs to not use OptimizationFlux, etc. and get this 2.0 fully ready. But this keeps being a runaway train of a PR so I want to get something that works, and then continue with improvements that can be done in parallel. |
One more spot to mark with Flux
Specify Flux in all layers
|
🎉 master now in v2.0 mode, slam it with the breaking changes you want. |
Superscedes #715