Bijectors and initial values to NUTS

[adamhaber]

Hi,

I have a question about the initial values passed to a NUTS sampler with constraining bijectors. As far as I understand, the initial values should be in the constrained space, right? If so - perhaps it would be more "friendly" to accept initial values from the unconstrained space? It took me a while to figure out why NUTS is returning nans...

I guess it's not too complicated (for the user, that is) to add something like

current_state=[bijector.forward(init) for bijector, init in zip(bijectors_list, initial_states)],

when calling the sample_mcmc function (this seems to work, at least in a simple example); Just raising the point to hear other people's thoughts on this.

[junpenglao]

To me it is more intuitive to accept value in the constrained space, as you can do something like init = target_joint_distribution.sample([...]), the assumption here is that you should be able to do target_log_prob(init), so accepting initial values from the unconstrained space will not make sense as well.

Maybe we should make clear in the docstring in sample_chain to current_state: Tensoror PythonlistofTensors representing the current state(s) of the Markov chain(s). ***you should be able to do kernel.target_log_prob(current_state)*** - however, I think transform kernel does not necessary have target_log_prob property, so maybe the wording should be more precise here.