RSVP for your your local TensorFlow Everywhere event today!

tfp.experimental.mcmc.simple_heuristic_tuning

Tune the number of steps and scaling of one mutation.

tfp.experimental.mcmc.simple_heuristic_tuning(
    num_steps, log_scalings, log_accept_prob, optimal_accept=0.234,
    target_accept_prob=0.99, name=None
)

This is a simple heuristic for tuning the number of steps of the next mutation, as well as the scaling of a transition kernel (e.g., step size in HMC, scale of a Normal proposal in RWMH) using the acceptance probability from the previous mutation stage in SMC.

num_steps The initial number of steps for the next mutation, to be tune.
log_scalings The log of the scale of the proposal kernel
log_accept_prob The log of the acceptance ratio from the last mutation.
optimal_accept Optimal acceptance ratio for a Transitional Kernel. Default value is 0.234 (Optimal for Random Walk Metropolis kernel).
target_accept_prob Target acceptance probability at the end of one mutation step. Default value: 0.99
name Python str name prefixed to Ops created by this function. Default value: None.

num_steps The number of steps for the next mutation.
new_log_scalings The log of the scale of the proposal kernel for the next mutation.