pyro-ppl / numpyro / 42598976056

93%

push

github

fix(infer): take at least one leapfrog step in HMC (#2416)

When the adapted step size grows past trajectory_length, HMC runs a single step of length trajectory_length, so the acceptance probability stops depending on the step size. If that acceptance is above target_accept_prob, dual averaging keeps raising the step size until it is clipped at finfo.max. trajectory_length / step_size then underflows to 0, num_steps becomes 0, and every transition is an always-accepted no-op: the chain never moves and nothing warns about it.

Clamp num_steps to at least 1. Results are unchanged whenever the ceiling was already at least 1.

1 of 1 new or added line in 1 file covered. (100.0%)

15946 of 17154 relevant lines covered (92.96%)

1.7 hits per line