Fix the pyro samplers

[michaeldeistler]

Fix for #286

Ok, I'm still pretty confused why this happens. It seems that, after the mcmc chain is run, .cleanup is called here which then calls ._reset(), which sets the init_params=None, see here.

However, after all this, transforms are performed on the samples. If transforms=None was specified, a transformation is automatically inferred, as explained here. However, this again sets up the kernel here, which requires init_params. It breaks because the init_params had been set to None.

A fix is to set transforms={}. With this, transforms are not automatically inferred and everything works fine.

--- Update ---
This is the exact commit that broke our code. After this change, if kernel is None, it still tries to run self.kernel.setup, which will break.

Should we report this as an issue to pyro?

[alvorithm]

I am not familiar with this part of the code and it is unlikely I can look at it before my holidays - sorry!

[michaeldeistler]

See pyro-ppl/pyro#2585

[michaeldeistler]

I'm closing this. Pyro has fixed the bug: pyro-ppl/pyro#2585

As soon as they release v1.4.1, we can lift our ==1.3.1 pin and also close the issue #286