Samplers

Samplers and related classes are found in com.stripe.rainier.sampler.

SamplerConfig

Calls to sample() can optionally provide a custom configuration object that implements the following trait:

trait SamplerConfig {
  def iterations: Int
  def warmupIterations: Int
  def statsWindow: Int

  def stepSizeTuner(): StepSizeTuner
  def massMatrixTuner(): MassMatrixTuner
  def sampler(): Sampler
}

Your configuration can directly implement every method of this trait, or extend DefaultConfiguration and selectively override.

  • iterations is the number of samples per chain to keep, after the warmup period (default: 1000)
  • warmupIterations is the number of samples per chain to use for warmup (default: 1000)
  • statsWindow is the number of training samples to keep for diagnostics like average acceptance rate (default: 100)
  • stepSizeTuner() should create a new StepSizeTuner object. This can be:
    • new DualAvgTuner(targetAcceptRate) (the default is this with 0.8)
    • StaticStepSize(stepSize)
  • massMatrixTuner() should create a new MassMatrixTuner object. This can be:
    • new IdentityMassMatrix
    • new DiagonalMassMatrixTuner(initialWindowSize, windowExpansion, skipFirst, skipLast) (the default is new DiagonalMassMatrixTuner(50, 1.5, 50, 50))
    • new DenseMassMatrixTuner(initialWindowSize, windowExpansion, skipFirst, skipLast)
  • sampler() should create a new Sampler object. This can be:
    • new HMCSampler(nSteps)
    • new EHMCSampler(maxSteps, minSteps) (the default is new EHMCSampler(1024, 1))

For backwards compatibility with 0.3.0, you can also build configs using the old HMC and EHMC constructors below.

HMC

Hamiltonian Monte Carlo with dual averaging as per Gelman & Hoffman

HMC(warmupIterations: Int, iterations: Int, nSteps: Int)

  • warmupIterations are used to find the best step size
  • iterations produce usable samples
  • nSteps is the number of leap-frog steps used for integration

EHMC

Empirical HMC as per Wu et al

EHMC(warmupIterations: Int, iterations: Int, l0: Int = 10, k: Int = 1000)

  • warmupIterations are used to find the best step size
  • iterations produce usable samples
  • k is the number of iterations used to build an empirical distribution of steps to U-turn
  • l0 is the number of leap-frog steps used during this empirical phase
← VecBryant (2020) →