Better adaptation configuration interface

[xukai92]

Related:
#448
#485 (comment)

[yebai]

Another related issue: #447

[yebai]

@mohamed82008 can you give this a try?

[trappmartin]

This might also be related #482

[yebai]

Perhaps @xukai92 can post a set of design goals for this issue.

[mohamed82008]

The current signature of sample is:

function sample(model::Function, alg::T;
                                chunk_size=CHUNKSIZE,               # set temporary chunk size
                                save_state=false,                   # flag for state saving
                                resume_from=nothing,                # chain to continue
                                reuse_spl_n=0,                      # flag for spl re-using
                                adapt_conf=STAN_DEFAULT_ADAPT_CONF  # adapt configuration
                               ) where T<:Hamiltonian

from

Turing.jl/src/samplers/hmc.jl

Line 87 in 7cb84a5

function sample(model::Function, alg::T;

.

The parameter adapt_conf is passed as the second argument to Sampler so that's the warmup sample size according to #448 (comment). This is already available as a kwarg but the name may be obscure.

Other sample functions are:

Turing.jl/src/samplers/ipmcmc.jl

Line 95 in 166c2fb

sample(model::Function, alg::IPMCMC) = begin

Turing.jl/src/samplers/gibbs.jl

Line 52 in 7cb84a5

sample(model::Function, alg::Gibbs;

Turing.jl/src/samplers/is.jl

Line 38 in 7cb84a5

sample(model::Function, alg::IS) = begin

Turing.jl/src/samplers/mh.jl

Line 111 in 7cb84a5

function sample(model::Function, alg::MH;

Turing.jl/src/samplers/pgibbs.jl

Line 82 in 7cb84a5

sample(model::Function, alg::PG;

Turing.jl/src/samplers/pmmh.jl

Line 104 in 7cb84a5

sample(model::Function, alg::PMMH;

Turing.jl/src/samplers/smc.jl

Line 72 in 7cb84a5

function sample(model::Function, alg::SMC)

Turing.jl/src/samplers/support/stan-interface.jl

Line 15 in 7cb84a5

sample(mf::T, ss::Stan.Sample) where {T<:Function} = sample(mf, ss.num_samples, ss.num_warmup, ss.save_warmup, ss.thin, ss.adapt, ss.alg)

With the exception of the last one, all take model and alg as the first 2 arguments. But not all have the same kwargs. I will see how I can improve the interface. Hints are welcome!

[yebai]

Hi @mohamed82008, thanks for looking into this. The main goals of this issue are

• Implement our own Adapt and Hmc related types. Turing currently re-uses the ones from Stan.jl, which can be found here

• Clean up the WarnUpManager type (code). Perhaps merge with type Adapt.

• Clean up the sample function so that they have consistent keyword arguments.

[yebai]

Closed by #793