Seems useful for teaching, but way to slow for actual research. Wonder if anyone has managed anything useful with it.
Anyone used Stan for Bayesian modeling?

Seems useful for teaching, but way to slow for actual research. Wonder if anyone has managed anything useful with it.
It isn’t slow. You are doing it wrong
Either you aren’t saving the result of then symbolic differtiation and are having Stan recompute every time you call it, or you don’t realise that Stan generates posterior samples which are much less auto correlated than other approaches and so you need to generate far fewer samples in total

OP is so obviously a troll who knows absolutely nothing about it.
A sap just had to take the bait and feed the troll.
It isn’t slow. You are doing it wrong
Either you aren’t saving the result of then symbolic differtiation and are having Stan recompute every time you call it, or you don’t realise that Stan generates posterior samples which are much less auto correlated than other approaches and so you need to generate far fewer samples in total 
It isn’t slow. You are doing it wrong
Either you aren’t saving the result of then symbolic differtiation and are having Stan recompute every time you call itHmm, I didn't know Stan could do this. I don't even know how this can work since different datasets (for the same model) will have different gradients. Any link to documentation for this?
or you don’t realise that Stan generates posterior samples which are much less auto correlated than other approaches and so you need to generate far fewer samples in total
I do know this, though sometimes I wish Stan had the option for just MetropolisHastings or plain HMC instead of NUTS.

Cant remember offhand. But basically everything stan does before it starts generating samples (ie the stuff that takes about 5 minutes) is preprocessing, and doing the automatic differentiation. It returns this stuff back to you, so you can just store it in a variable and pass it back in when you next call the Stan method, so it doesnt need to do it again.It isn’t slow. You are doing it wrong
Either you aren’t saving the result of then symbolic differtiation and are having Stan recompute every time you call itHmm, I didn't know Stan could do this. I don't even know how this can work since different datasets (for the same model) will have different gradients. Any link to documentation for this?.

Cant remember offhand. But basically everything stan does before it starts generating samples (ie the stuff that takes about 5 minutes) is preprocessing, and doing the automatic differentiation. It returns this stuff back to you, so you can just store it in a variable and pass it back in when you next call the Stan method, so it doesnt need to do it again.
Thanks, I'll look into this.

In my experience with some fairly small models, it was great. For example, I used it instead of an R package for a stochastic vol model and found Stan substantially faster. I found it fairly simple to learn the basics, but my sense is that learning to make it really fast takes some experience. Installation can also be a hassle.

In my experience with some fairly small models, it was great. For example, I used it instead of an R package for a stochastic vol model and found Stan substantially faster. I found it fairly simple to learn the basics, but my sense is that learning to make it really fast takes some experience. Installation can also be a hassle.
I can find hardly any documentation about speeding up Stan. Something like that would be really useful.