tfp.mcmc.default_exchange_proposed_fn

View source on GitHub

Default exchange proposal function, for replica exchange MC.

tfp.mcmc.default_exchange_proposed_fn(prob_exchange)

With probability prob_exchange propose combinations of replica for exchange. When exchanging, create combinations of adjacent replicas in Replica Exchange Monte Carlo

exchange_fn = default_exchange_proposed_fn(prob_exchange=0.5)
exchange_proposed = exchange_fn(num_replica=3)

exchange_proposed.eval()
==> [[0, 1]]  # 1 exchange, 0 <--> 1

exchange_proposed.eval()
==> []  # 0 exchanges

Args:

• prob_exchange: Scalar Tensor giving probability that any exchanges will be generated.

Returns:

• default_exchange_proposed_fn_: Python callable which take a number of replicas (a Python integer), and return combinations of replicas for exchange as an [n, 2] integer Tensor, 0 <= n <= num_replica // 2, with unique values in the set {0, ..., num_replica}.