gemlib.mcmc.transform_sampling_algorithm

gemlib.mcmc.transform_sampling_algorithm#

gemlib.mcmc.transform_sampling_algorithm(bijectors, sampling_algorithm)#

Transform a sampling algorithm.

This wrapper transforms sampling_algorithm with respect to the probability measure on which it acts. transform_sampling_algorithm is particularly useful for unbounding parameter spaces in order to use algorithms such as Hamiltonian Monte Carlo or the No-U-Turn-Samplers (NUTS).

It does this by applying the change-of-variables formula, such that for \(Y = g(X)\),

\[f_Y(y)=f_X(g^{-1}(y))\left|\frac{\mathrm{d}g^{-1}(y)}{\mathrm{d}y}\right|\]
Parameters:

  • bijectors (Iterable) – a structure of TensorFlow Probability bijectors compatible with position

  • sampling_algorithm (SamplingAlgorithm) – a sampling algorithm

Returns:

A new SamplingAlgorithm representing a kernel working on the transformed space.

Examples

Instantiate a transformed hmc kernel:

import tensorflow_probability as tfp
from gemlib.mcmc import transform_sampling_algorithm
from gemlib.mcmc import hmc

kernel = transform_sampling_algorithm(
    bijectors=[tfp.bijectors.Exp(), tfp.bijectors.Exp()],
    sampling_algorithm=hmc(step_size=0.1, num_leapfrog_steps=16),
)
Contents