search

donaldRwilliams / chkptstanr

Checkpoint Stan R
https://donaldrwilliams.github.io/chkptstanr/
Other
9 stars 5 forks source link

"Initial Warmup (Typical Set)" taking a long time #11

Open vascosa opened 2 years ago

vascosa commented 2 years ago

Hello,

I am running a fairly complex model with chkpt_brms, and it gets stuck on the initial warmup/typical set stage.

If I set iter_typical to 1, it starts sampling at an acceptable speed, but with 150 warm-up samples it stopped the model after about 3 days as I was seeing no progress.

I was wondering if there is any explanation you can think for this,m because as far as I can tell these are just additional warmup samples. And the first 150 warm-up samples when iter_typical is 1 are MUCH faster.

I was also wondering what the need for these is in chkptstanr, as with brms I would just run the model without additional warmup samples to find the typical set. What was the rationale for including this stage in chkptstanr?

Cheers, Vasco

donaldRwilliams commented 2 years ago

Hey !

Just getting back from a vacation :-)

Indeed, the "typical" set is required. In testing, we found this was critical for the sampler to converge.

Did you check convergence with 1 for the iter_typical ?

Can you provide a trace plot here, please ?

cpfiffer commented 1 year ago

I'm not clear on why the typical set sampling is required -- shouldn't this just be part of standard warmup and burn-in? Why is this a separate step in the sampling procedure?