ipython notebook kernel dies with njobs=4

[jsalvatier]

I'm trying to add multithreading to the [UQ book chapter](https://github.com/pymc-devs/uq_chapter/blob/master/reference/UQ Handbook Chapter.ipynb) by adding njobs=4 to one of the calls to sample, but it kills my kernel. Have you seen this before?

[twiecki]

Is this with theano master? They had some parallel fixes recently.

[PaulSorenson]

Using pymc3 '3.0' and theano 0.7.0 and so far haven't found a setting of njobs that doesn't crash my iPython kernel. I pulled the latest theano from github and same deal.

[kyleam]

I tested the first sample call in the chapter with njobs set to 2 and 4, and my kernel survives. Does yours seem to crash right away or mid-sampling?

---

Python 3.4.3
IPython 3.1.0
Theano 0.7.0
Linux

[jsalvatier]

Interesting

Python 2.7.3
Ipython 3.1
Theano 0.7.0
Ubuntu

I'll try Python 3

On Mon, May 11, 2015 at 8:22 AM, Kyle Meyer notifications@github.com
wrote:

I tested the first sample call in the chapter with njobs set to 2 and 4,

and my kernel survives. Does yours seem to crash right away or mid-sampling?

Python 3.4.3
IPython 3.1.0
Theano 0.7.0
Linux

—
Reply to this email directly or view it on GitHub
#707 (comment).

[twiecki]

This might in fact be a python 3 issue: stan-dev/pystan#142

[PaulSorenson]

And my python is 3.4.3 on Fedora 21. Thomas that link seems to indicate it is a Python 2 issue (contrary to my experience).

[PaulSorenson]

It worked without incident just now with progressbar=False and njobs=3. With progressbar=True and no other changes it crashes, right around where the progress bar would have appeared.

[twiecki]

Interesting, can anyone confirm that it works without the progressbar?

[PaulSorenson]

I just ran latest github theano and pymc3 and updated ipython with "pip3.4 install --update ipython[notebook]".

Notebook doesn't crash with the progress bar when njobs > 1 although the progress bar doesn't show until it reaches 100% (except when njobs=1 which works as before).

The version of the notebook server is 3.1.0-cbccb68 and is running on:
Python 3.4.3 (default, Mar 22 2015, 11:46:05)
[GCC 4.8.3 20140911 (Red Hat 4.8.3-7)]

[twiecki]

Great, thanks. We should just disable the progress bar if njobs != 1.

[PaulSorenson]

I think this could do the trick since there is already an if njobs > 1 block:

diff --git a/pymc3/sampling.py b/pymc3/sampling.py
index 7f66987..6c2c63f 100644
--- a/pymc3/sampling.py
+++ b/pymc3/sampling.py
@@ -68,7 +68,7 @@ def sample(draws, step, start=None, trace=None, chain=0, njobs

         chains = list(range(chain, chain + njobs))

-        pbars = [progressbar] + [False] * (njobs - 1)
+        pbars = [False] * njobs

         argset = zip([draws] * njobs,
                      [step] * njobs,

I can make a pull request if people are happy with that.

[kyleam]

I think we should work toward a solution for displaying progress
information for multiple chains. For many analyses, people are going
to use multiple chains, and sampling will take a substantial chunk of
time, so there should be some way to monitor the status.

The current hack is based on the idea that having information about
one chain is better than no information at all, but of course the
chains can get into different states and one chain is not always very
representative of the other ones. Although this hack is now pretty
useless in the notebook (since the complete bar just pops up at the
end), it still works for programs that aren't run in a notebook (to
the extent that information from just one chain is useful).

It'd be great if someone works out a way display progress for all jobs
in the notebook. Outside of the notebook, this could be as simple as
each chain spitting out "N / M" every so many iterations (similar to
CmdStan).

Perhaps we could have a few different methods for displaying sampling
progress and choose based on the whether the code's being run in a notebook (is this
easy to know?).

[twiecki]

@kyleam I agree, it would definitely be a nice feature.
I think a way to check whether in an IPython NB is:

try:
   get_ipython()
except NameError:
   # not in IPython