Allow usage of AbstractSampler

[torfjelde]

This is a very rough idea of how we could go about supporting AbstractMCMC.AbstractSampler to reduce the necessary glue code.

With this, the following works:

julia> using Revise, Turing

julia> @model demo() = x ~ Normal()
demo (generic function with 2 methods)

julia> model = demo()
DynamicPPL.Model{typeof(demo), (), (), (), Tuple{}, Tuple{}, DynamicPPL.DefaultContext}(demo, NamedTuple(), NamedTuple(), DynamicPPL.DefaultContext())

julia> sampler = Turing.Inference.initialize_nuts(model);  # NOTE: This has been moved to the tests for now.

julia> chain = sample(model, externalsampler(sampler), 1000; nadapts=500, discard_initial=500)
Sampling 100%|██████████████████████████████████████████████████████████████████████████████████████| Time: 0:00:00
Chains MCMC chain (1000×14×1 Array{Float64, 3}):

Iterations        = 501:1:1500
Number of chains  = 1
Samples per chain = 1000
Wall duration     = 0.16 seconds
Compute duration  = 0.16 seconds
parameters        = x
internals         = lp, n_steps, is_accept, acceptance_rate, log_density, hamiltonian_energy, hamiltonian_energy_error, max_hamiltonian_energy_error, tree_depth, numerical_error, step_size, nom_step_size, is_adapt

Summary Statistics
  parameters      mean       std      mcse   ess_bulk   ess_tail      rhat   ess_per_sec 
      Symbol   Float64   Float64   Float64    Float64    Float64   Float64       Float64 

           x    0.0692    0.9664    0.0528   339.0005   512.5947    0.9991     2105.5929

Quantiles
  parameters      2.5%     25.0%     50.0%     75.0%     97.5% 
      Symbol   Float64   Float64   Float64   Float64   Float64 

           x   -1.7927   -0.5420    0.0395    0.7099    1.9460

[torfjelde]

As demonstrated in #2011 it seems supporting this properly is somewhat non-trivial, and so for the moment I've decided to just introduce a simple SamplerWrapper <: InferenceAlgorithm which wraps a AbstractMCMC.AbstractSampler but inherits then inherits everything else that comes with being an InferenceAlgorithm.

The annoying thing (other than it being ugly) is that the user has to wrap the AbstractMCMC.AbstractSampler themselves, since otherwise we run into ambuigity issues as before. The positives: any sampler using LogDensityFunction is now usable and integrates "nicely" with Turing.

[coveralls]

Pull Request Test Coverage Report for Build 5457548017

Warning: This coverage report may be inaccurate.

This pull request's base commit is no longer the HEAD commit of its target branch. This means it includes changes from outside the original pull request, including, potentially, unrelated coverage changes.

• For more information on this, see Tracking coverage changes with pull request builds.
• To avoid this issue with future PRs, see these Recommended CI Configurations.
• For a quick fix, rebase this PR at GitHub. Your next report should be accurate.

Details

• 0 of 32 (0.0%) changed or added relevant lines in 2 files are covered.
• 1 unchanged line in 1 file lost coverage.
• Overall coverage remained the same at 0.0%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/inference/Inference.jl	0	1	0.0%
src/contrib/inference/abstractmcmc.jl	0	31	0.0%

Files with Coverage Reduction	New Missed Lines	%
src/inference/Inference.jl	1	0.0%

Totals
Change from base Build 5455388390:	0.0%
Covered Lines:	0
Relevant Lines:	1458

---

💛 - Coveralls

[codecov]

Codecov Report

Patch and project coverage have no change.

Comparison is base (0dee0f8) 0.00% compared to head (b0503e5) 0.00%.

Additional details and impacted files

@@          Coverage Diff           @@
##           master   #2008   +/-   ##
======================================
  Coverage    0.00%   0.00%           
======================================
  Files          21      22    +1     
  Lines        1426    1458   +32     
======================================
- Misses       1426    1458   +32     

Impacted Files	Coverage Δ
src/Turing.jl	0.00% <ø> (ø)
src/contrib/inference/abstractmcmc.jl	0.00% <0.00%> (ø)
src/inference/Inference.jl	0.00% <0.00%> (ø)

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.

[torfjelde]

I believe the tests are sometimes failing simply because of numerical issues.

Some of the test models, e.g. demo_dot_assume_observe_index_literal, have one variance parameter for every observation. This means that the MLE estimate will have the variance parameter s be 0. Most of the time this seems to work just fine, i.e. LBFGS finds something close to 0, but sometimes it fails.

[devmotion]

@torfjelde Isn't this obsolete now that #2026 was merged?

[yebai]

I think so. @JaimeRZP, can you do a follow-up PR to unify these Transition types?

[torfjelde]

Yes it is, which I was aware of; me and Jaime wanted to wait until that had been merged before merging this PR.
I was planning to incorporate those changes into this PR before merging. We're also missing a version-bump.

I'd appreciate it if we left merging of a PR to the person who opened it, unless otherwise explicitly stated. In particular now when it's just a matter of days before I'll be back in full development capacity again. This has happened quite few times now :/

Also, it's not like this PR needs to be merged to be able to develop other functionality; it's easy enough to just depend on the branch directly.

[yebai]

I was under the impression that @JaimeRZP needs this to be released. Sorry for the rush -- hopefully, it didn't break anything! We can always add more changes in a follow-up PR.

[torfjelde]

Ah gotcha. @JaimeRZP you can always just do ]add Turing#torfjelde/allow-abstractsampler-draft if you want to try out recent developments. And if you want to develop features based on this branch, just create a branch based on this PR and then continue from there, as you did with #2028 :)