AHMC

Creates a sampler that uses the AdvancedHMC framework to construct an Hamiltonian Monte Carlo NUTS sampler.

The user must specify the metric they want to use. Typically we recommend DiagEuclideanMetric as a reasonable starting place. The other options are chosen to match the Stan languages defaults and should provide a good starting point. Please see the AdvancedHMC docs for more information.

Notes

For autodiff the must provide a Val(::Symbol) that specifies the AD backend. Currently, we use LogDensityProblemsAD.

Fields

• metric: AdvancedHMC metric to use

• integrator: AdvancedHMC integrator Defaults to AdvancedHMC.Leapfrog

• trajectory: HMC trajectory sampler Defaults to AdvancedHMC.MultinomialTS

• termination: HMC termination condition Defaults to AdvancedHMC.StrictGeneralisedNoUTurn

• adaptor: Adaptation strategy for mass matrix and stepsize Defaults to AdvancedHMC.StanHMCAdaptor

• targetacc: Target acceptance rate for all trajectories on the tree Defaults to 0.85

• init_buffer: The number of steps for the initial tuning phase. Defaults to 75 which is the Stan default

• term_buffer: The number of steps for the final fast step size adaptation Default if 50 which is the Stan default

• window_size: The number of steps to tune the covariance before the first doubling Default is 25 which is the Stan default

• autodiff: autodiff backend see LogDensitProblemsAD.jl for possible backends. The default is Zygote which is appropriate for high dimensional problems.

Disk

Type that specifies to save the HMC results to disk.

Fields

• name: Path of the directory where the results will be saved. If the path does not exist it will be automatically created.

• stride: The output stride, i.e. every stride steps the MCMC output will be dumped to disk.

Memory

Stored the HMC samplers in memory or ram.

load_table(out::DiskOutput, indices::Union{Base.Colon, UnitRange, StepRange}=Base.Colon(); table="samples")
load_table(out::String, indices::Union{Base.Colon, UnitRange, StepRange}=Base.Colon(); table="samples")

The the results from a HMC run saved to disk. To read in the output the user can either pass the resulting out object, or the path to the directory that the results were saved, i.e. the path specified in DiskStore.

Arguments

• out::Union{String, DiskOutput}: If out is a string is must point to the direct that the DiskStore pointed to. Otherwise it is what is directly returned from sample.
• indices: The indices of the that you want to load into memory. The default is to load the entire table.

Keyword Arguments

• table: A string specifying the table you wish to read in. There are two options: "samples" which corresponds to the actual MCMC chain, and stats which corresponds to additional information about the sampler, e.g., the log density of each sample and tree statistics.

AbstractMCMC.sample(post::Comrade.Posterior,
                    sampler::AHMC,
                    nsamples;
                    initial_params=nothing,
                    saveto::Union{Memory, Disk}=Memory(),
                    kwargs...)

Samples the posterior post using the AdvancedHMC sampler specified by AHMC. This will run the sampler for nsamples.

To initialize the chain the user can set initial_params to Vector{NamedTuple} whose elements are the starting locations for each of the nchains. If no starting location is specified nchains random samples from the prior will be chosen for the starting locations.

With saveto the user can optionally specify whether to store the samples in memory MemoryStore or save directly to disk with DiskStore(filename, stride). The stride controls how often the samples are dumped to disk. In addition is saving to disk the user can also resume a previous run by setting restart=true.

For possible kwargs please see the AdvancedHMC.jl docs

This returns a PosteriorSamples object.

Notes

This will automatically transform the posterior to the flattened unconstrained space.

AbstractMCMC.sample(post::Comrade.Posterior,
                    sampler::AHMC,
                    parallel::AbstractMCMC.AbstractMCMCEnsemble,
                    nsamples,
                    nchainsl;
                    initial_params=nothing,
                    kwargs...)

Samples the posterior post using the AdvancedHMC sampler specified by AHMC. This will sample nchains copies of the posterior using the parallel scheme. Each chain will be sampled for nsamples.

To initialize the chain the user can set initial_params to Vector{NamedTuple} whose elements are the starting locations for each of the nchains. If no starting location is specified nchains random samples from the prior will be chosen for the starting locations.

For possible kwargs please see the AdvancedHMC.jl docs

This returns a PosteriorSamples object indexed as iteration × chain.

Notes

This will automatically transform the posterior to the flattened unconstrained space.