tfp.experimental.mcmc.chees_rate_criterion

ChEES rate criterion.

tfp.experimental.mcmc.chees_rate_criterion(
    previous_state,
    proposed_state,
    accept_prob,
    trajectory_length,
    validate_args=False,
    experimental_shard_axis_names=None,
    experimental_reduce_chain_axis_names=None
)

This is just like chees_criterion, but normalized by the trajectory length:

ChEES rate = 1/4 E[(||x' - E[x]||**2 - ||x - E[x]||**2)**2 /
  trajectory_length]

Args
previous_state	(Possibly nested) floating point Tensor. The previous state of the HMC chain.
proposed_state	(Possibly nested) floating point Tensor. The proposed state of the HMC chain.
accept_prob	Floating Tensor. Probability of acceping the proposed state.
trajectory_length	Floating Tensor. Trajectory length.
validate_args	Whether to perform non-static argument validation.
experimental_shard_axis_names	A structure of string names indicating how members of the state are sharded.
experimental_reduce_chain_axis_names	A string or list of string names indicating which named chain axes to reduce over when computing the criterion.

Returns
chees_rate	The value of the ChEES rate criterion.