Move Libtask to an extension #75
Conversation
|
We probably need to think about the handling of the rng in the non-libtask case |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Pull Request Test Coverage Report for Build 6448601669
💛 - Coveralls |
Codecov ReportAttention:
Additional details and impacted files@@ Coverage Diff @@
## master #75 +/- ##
==========================================
- Coverage 96.32% 95.60% -0.72%
==========================================
Files 7 8 +1
Lines 381 410 +29
==========================================
+ Hits 367 392 +25
- Misses 14 18 +4
☔ View full report in Codecov by Sentry. |
|
|
||
| Base.copy(model::LibtaskModel) = LibtaskModel(model.f, copy(model.ctask)) | ||
|
|
||
| const LibtaskTrace{R} = AdvancedPS.Trace{<:LibtaskModel,R} |
There was a problem hiding this comment.
I don't think we can access this LibtaskModel externally. Hence, users can not use this libtask extension even when Libtask is loaded. To fix this, we can consider overloading a function defined in AdvancedPS.
| const LibtaskTrace{R} = AdvancedPS.Trace{<:LibtaskModel,R} | |
| const LibtaskTrace{R} = AdvancedPS.Trace{<:LibtaskModel,R} | |
| AdvancedPS.Trace(Libtask, f, args; __rng__ = Random.default_rng()) = AdvancedPS.Trace(LibtaskModel(f, args...), rng=__rng__) |
There was a problem hiding this comment.
I think it's not possible (for general Julia versions - maybe at all?) to dispatch on modules. One will be able to use them e.g. as type parameters only in 1.10: JuliaLang/julia#47749
There was a problem hiding this comment.
It's possible to access types by retrieving the extension module with get_extension. But that does not lead to a good API IMO and also adding accessible types is not the intended use case of extensions.
There was a problem hiding this comment.
Thanks, @devmotion -- that's very helpful to know. However, in this case, we can dispatch any type. For example:
| const LibtaskTrace{R} = AdvancedPS.Trace{<:LibtaskModel,R} | |
| const LibtaskTrace{R} = AdvancedPS.Trace{<:LibtaskModel,R} | |
| AdvancedPS.Trace(Val{:libtask}, f, args; __rng__ = Random.default_rng()) = AdvancedPS.Trace(LibtaskModel(f, args...), rng=__rng__) |
There was a problem hiding this comment.
I wanted to hide all the internals away from the user, not sure they should know about what Trace is. The main package implements sample(::AbstractStateSpace) and the extension overloads it with sample(::AbstractModel).
This works only if you load Libtask (and enable the extension):
struct Model <: AbstractMCMC.AbstractModel end
function (model::Model)(rng::AbstractRNG)
x = rand(...)
end
sample(rng, Model(), PG(100), 1_000)
test/container.jl
Outdated
| # # Test task copy version of trace | ||
| # trng = AdvancedPS.TracedRNG() | ||
| # tmodel = AdvancedPS.GenericModel(f2, trng) | ||
| # tr = AdvancedPS.Trace(tmodel, trng) |
There was a problem hiding this comment.
With the additional constructor proposed in https://github.com/TuringLang/AdvancedPS.jl/pull/75/files#r1284737656, we should be able to uncomment these tests.
|
Excellent work -- I left some suggestions above. |
FredericWantiez
force-pushed
the
fred/extension
branch
from
631900a to
006c4fd
Compare
|
Just realized while fixing some of tests that the AdvancedPS.jl/src/container.jl Line 227 in 72a1e55 update_ref! should probably dispatch on the type of the sampler somehow.
|
|
Something like this would actually would use resample the ancestor particle for both chains_pg = sample(model, pg, Nₛ);
chains_pgas = sample(model, pgas, Nₛ); |
Not sure I follow here, can you clarity? |
FredericWantiez
force-pushed
the
fred/extension
branch
from
006c4fd to
806c10a
Compare
Still a draft but we can track the progress on moving Libtask to its own extension