NUTS initialization

Hi there,
I’m wondering whether it is useful to run pm.init_nuts(’…’ ) before pm.sample(init=’…’) and what is the actual difference between the init in these two calls?
Cheers,
Vian

You can think of pm.sample() as a wrapper around several other API calls, including initialization, tuning, and sampling itself. In certain scenarios you may wish to make those calls individually, but in the general case you probably don’t.

Hi cluhmann,
I am using pm.sample() with init=“advi+adapt_diag”, and I would like to use a specific obj_optimizer in the advi initialization. Is there a way to do this?
Alternatively, is there a way to specify the starting point of pm.sample() directly, skipping the init method? As I understand it, initvals only overrides the default initial value of the init method, not the starting point of NUTS.
I’m using pymc version 5.1.1.

Looks like you got your question answered here? If not, feel free to ask in that thread.

1 Like