Stan and Math v2.7.1
Daniel Lee
Math:
* sparse multiply (Krzysztof)
Bob Carpenter
We're currently blocked
* Krzysztof's tests to be included (and pass) for sparse multiply
* advi-adapt-eta unit tests failing
If you feel strongly about either of these being in 2.7.1, speak up now!
No rush on my part and it doesn't matter to me if either of the
above are in 2.7.1. But I do think they shouldn't go in without
tests passing.
- Bob
> 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.
> For more options, visit https://groups.google.com/d/optout.
Ben Goodrich
I need to do a PR for Stan to doc inv_Phi. I would say that ADVI in its current form is not really widely usable, so if we bump advi-adapt-eta, then we need to unexpose vb() in rstan and eliminate the option to call vb() from rstanarm. That is a pretty big price to pay, so if Dustin / Alp think it is feasible to merge advi-adapt-eta in the next couple of days, then it is worth waiting for.
Ben
Krzysztof Sakrejda
> I just talked to Daniel for clarification.
>
> We're currently blocked
>
> * Krzysztof's tests to be included (and pass) for sparse multiply
The specific tests I included (not sure why they got dropped before) do pass, though we're still waiting on the rest of the unrelated test suite. I also need to add tests/doc to stan-dev/stan. I'd love to see it go in for 2.7.1 since it's so close to being done and the rest is just grunt work. I should be able to get it done this week.
Krzysztof
>
> * advi-adapt-eta unit tests failing
>
> If you feel strongly about either of these being in 2.7.1, speak up now!
>
> No rush on my part and it doesn't matter to me if either of the
> above are in 2.7.1. But I do think they shouldn't go in without
> tests passing.
>
> - Bob
>
Bob Carpenter
research projects! We appreciate it.
- Bob
Sebastian Weber
As far as I understood, you plan to advertise with 2.7.1 the autodiff library, so maybe this is of interest in this context: Currently the issue
https://github.com/stan-dev/math/issues/133
is still open. It describes test failures of g++ 4.4.7 & Intel compilers in stan-math. The g++ 4.4.7 is part of RedHat Linux Enterprise 6.5 which is installed in many production systems.
Moreover: Are those tests which fail critical? Can I not trust stan if those tests above fail with my compiler? Is there somewhere some information where users can get such information (how severe a failing stan-math test is with their compiler)?
Best,
Sebastian
On Tuesday, September 1, 2015 at 9:29:27 PM UTC+2, Bob Carpenter wrote:
> It's that grunt work that separates working code from
> research projects! We appreciate it.
>
> - Bob
>
>
> > On Aug 31, 2015, at 9:47 PM, Krzysztof Sakrejda
> >
Bob Carpenter
MKL's non-compliant arithmetic like the other bugs we've tracked
down.
But even were we to relax the constraints in the tests,
users will have to decide for themselves how much the Intel
compiler's lowered arithmetic precision is worth in terms of
speed or accuracy --- we don't know the answer ourselves.
- Bob
Sebastian Weber
The MKL+Intel is more sensitive than other configurations, yes. But note that I reported also results for Intel without MKL and for the g++ 4.4.7 without the MKL. So there are issues which affect RHEL6.5 installations which use g++ 4.4.7 as it was reported in the issue.
BTW, I have disabled MKL in all tests here run with stan-math.
Sebastian
Bob Carpenter
- Bob