It is slightly awkward to work with path length to then divide it by the step size systematically. It seems like tensorflow-probability uses number of steps instead; investigate to see whether there is a reason to use path length instead. If not, switch to number of steps.
Switched to num_integration_steps; it was surprisingly easy (3 instances of the variable in the codebase), and could be reversed any time should it be a bad idea.
It is slightly awkward to work with path length to then divide it by the step size systematically. It seems like tensorflow-probability uses number of steps instead; investigate to see whether there is a reason to use path length instead. If not, switch to number of steps.