Iterations and warm ups in Stan

General
1

Hi all,
I am currently trying to Implement Hidden Markov Models using PyStan. I understand that Stan has a model fitting parameter as n_iter and warmups.
However, I am not able to find the meaning behind the warmups and iterations. I have trained the model and now I am getting the results for 4000 iterations.
Does this mean that last iteration produces the best result or I am missing something?

2

Hi,

See https://mc-stan.org/docs/2_19/reference-manual/hmc-algorithm-parameters.html

Warmup --> learn hmc parameters

Iter-warmup --> posterior draws

iter = 2000 per chain
Warmup = 1000 per chain
Draws = 1000 per chain
4 chains --> 4000 draws

1 Like
3

Thanks for your response.
Is there any way in Stan wherein I can check the parameters after the warmup. I mean if I want to skip the iteration step and just look at the parameters that were learned from warm up.

4

With current develop branch you have this option

Older pystan have textual output in fit.get_adaption_info()

5

Hi, thanks for the help.
However, this method still required the iteration step (even though number of iteration steps=1).
However, I am looking for the below solution:

  • No Iteration.
  • Get the trained parameters as soon as warm-up finishes.
6

There is a small “bug” in PyStan, so it can not be done currently. You could use CmdStan / CmdStanPy and extract metric, stepsize and last location from the generated csv / StanFit.

7

Are you absolutely sure you know what you’re asking for here? More expert folks could chime in to correct me, but I don’t think the final “sample” from the warmup is anything special; it’s not necessarily a sample from the typical set and certainly isn’t any sort of minimum/maximum.

8

For anyone else confused by this, @ahartikainen means: “iter minus warmup”. It’s clear in the numbers, but I was looking for a parameter named iter-warmup!

1 Like