RStan/PyStan ADVI status + documentation

[Alp Kucukelbir]

hi ben, allen, &co.

friendly ping: anything i can help with ADVI RStan/PyStan integration?

i should provide documentation/instructions. pointer would be welcome.

thanks so much guys.

cheers
alp

[Ben Goodrich]

I'm going to work on it over the break. I think it should not be too difficult to re-enable it since it was enabled for a few weeks. If you know offhand what in the C++ is now incompatible with how the adapt_eta branch stood on September 3

https://github.com/stan-dev/rstan/compare/feature/adapt-eta

let me know. Otherwise, I'm sure I'll discover it.

Ben

[Alp Kucukelbir]

hi ben,

(can i make a pull request out of that branch so that i can provide
some line comments?)

a few things are simpler now. (eta is just a double, no strings nonsense).

the new parameters structure mirrors that of NUTS:

adapt
Eta Adaptation for Variational Inference
Valid subarguments: engaged, iter

engaged=<boolean>
Adaptation engaged?
Valid values: [0, 1]
Defaults to 1

iter=<int>
Number of iterations for eta adaptation
Valid values: 0 < iter
Defaults to 50

if the timing is bad for you, i can also take a shot at pushing some
of these changes to that branch myself.

in any case, i'll try to update the stan/cmdstan manuals today to make
everyone's life easier.

thanks so much ben.

cheers
alp

> --
> You received this message because you are subscribed to a topic in the
> Google Groups "stan development mailing list" group.
> To unsubscribe from this topic, visit
> https://groups.google.com/d/topic/stan-dev/H9WaggHH4Hk/unsubscribe.
> To unsubscribe from this group and all its topics, send an email to
> stan-dev+u...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Ben Goodrich]

On Wednesday, November 25, 2015 at 11:55:18 AM UTC-5, Alp Kucukelbir wrote:

(can i make a pull request out of that branch so that i can provide
some line comments?)

Let me give it a shot first. That branch is out of date due to non-ADVI changes in rstan and Lib / Math since then that I'll probably have to manage.

Ben

[Alp Kucukelbir]

roger. do let me know how i can help!

@allen: how about PyStan? can i help there?

cheers
alp

[Daniel Lee]

Ben, I should be able to set that up. 

I'm still struggling with the python build, but hopefully we can resolve that at some point soon. 

--
You received this message because you are subscribed to the Google Groups "stan development mailing list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.

[Alp Kucukelbir]

documentation for cmdStan + Stan: done.

[Bob Carpenter]

Thanks --- that looks good and I'm just waiting on
whatever's up with Travis to merge it.

- Bob

> On Nov 25, 2015, at 5:33 PM, Alp Kucukelbir <akucu...@gmail.com> wrote:
>
> documentation for cmdStan + Stan: done.
>

[Allen B. Riddell]

Hi Alp,

It's probably easiest if I just mirror the implementation from RStan
once that's finished. (I'd welcome a reminder whenever ADVI is working in
RStan.)

Best,

Allen

> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.

> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.

[Alp Kucukelbir]

hi allen,

okay sounds great! let me know if i can help. i'll definitely ping you
and keep you updated.

cheers
alp

[Rob J. Goedman]

Hi Alp & Allen,

For Julia (Stan.jl) I was thinking to do the same, if possible. It is still too early to attempt a pure C++ interface, maybe once Julia 0.5 comes around.

For Julia 0.4.x I just read the output files. A few folks seem ok with that approach.

Regards,
Rob