I just installed pymc3 from conda-forge on a different computer and don’t see progress bar during init. Is this intentional?
There is not progressbar for the current defautl initiualization jitter+adapt_diag
Uhm, I was using ‘advi+adapt_diag’… … double checking…
OK, confirmed. In fact init_nuts is never entered. I’m investigating if the fact that I initialize the step before calling sample could be the cause
step = pm.NUTS()
trace = pm.sample(draws=conf.block_size,
chains=1,
njobs=njobs,
discard_tuned_samples=conf.discard_tuned_sample,
tune=tune,
init ='advi+adapt_diag',#conf.init,
step=step)
OK, seems to be by design. If I want to use NUTS this way probably have to init it manually using a call to init_nuts.
Yeah if you are passing step
to pm.sample
it turns off the initialization. If you have specific kwarg wants to pass to pm.NUTS()
you should try nuts_kwarg=dict(......)
.
Oh, that was not to supply kwargs but for my ad-hoc save-to-disk and peek-while-you-sample framework. I need to have a name for step instance in the scope of my module in order to pickle it Maybe I’ll add a PR later today to have pm.sample optionally spit out step instance along with the trace. That would enable peeking during sampling for compound steps as well.
I see. Yeah that part of the code is not very exposed. When you say peek while sample do you mean peeking the parameters in the sampler? Like kernel size etc? Maybe it is easier to cherry pick the code from pm.sample and call pm.sampling._iter_sample directly.