Manually defined partially conjugate update step

Specific Domains Probabilistic programming
1

I have a model where I can apply Turing to the full parameter vector, that seems to work. However, I can partition the parameter such that a Gibbs sampler can be devised where one parameter is updated in a partially conjugate step (the remainder of the parameters can be dealt with for example with HMC). Is there a way to pass the partially conjugate step to the model?

I know about syntax like

Gibbs(HMC(0.2, 3, :p), PG(20, :z))

So rather than PG I would like to to specify this step manually.

1 Like
2

There is a PR for this but I’m not sure if it is released or fully documented yet but there’s an example.

1 Like
3

Thanks, the example is very helpful.