Travis tests failing

[jsalvatier]

The tests are not working and haven't for a while. Lets fix that.
At least one of the errors is:

ERROR: pymc.tests.test_plots.test_plots
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/travis/anaconda/envs/testenv/lib/python2.7/site-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/home/travis/build/pymc-devs/pymc/pymc/tests/test_plots.py", line 20, in test_plots
    forestplot(trace)
  File "/home/travis/build/pymc-devs/pymc/pymc/plots.py", line 325, in forestplot
    interval_plot = subplot(gs[0])
  File "/home/travis/anaconda/envs/testenv/lib/python2.7/site-packages/matplotlib/pyplot.py", line 896, in subplot
    fig = gcf()
  File "/home/travis/anaconda/envs/testenv/lib/python2.7/site-packages/matplotlib/pyplot.py", line 450, in gcf
    return figure()
  File "/home/travis/anaconda/envs/testenv/lib/python2.7/site-packages/matplotlib/pyplot.py", line 423, in figure
    **kwargs)
  File "/home/travis/anaconda/envs/testenv/lib/python2.7/site-packages/matplotlib/backends/backend_qt4agg.py", line 31, in new_figure_manager
    return new_figure_manager_given_figure(num, thisFig)
  File "/home/travis/anaconda/envs/testenv/lib/python2.7/site-packages/matplotlib/backends/backend_qt4agg.py", line 38, in new_figure_manager_given_figure
    canvas = FigureCanvasQTAgg(figure)
  File "/home/travis/anaconda/envs/testenv/lib/python2.7/site-packages/matplotlib/backends/backend_qt4agg.py", line 70, in __init__
    FigureCanvasQT.__init__( self, figure )
  File "/home/travis/anaconda/envs/testenv/lib/python2.7/site-packages/matplotlib/backends/backend_qt4.py", line 207, in __init__
    _create_qApp()
  File "/home/travis/anaconda/envs/testenv/lib/python2.7/site-packages/matplotlib/backends/backend_qt4.py", line 62, in _create_qApp
    raise RuntimeError('Invalid DISPLAY variable')
RuntimeError: Invalid DISPLAY variable
-------------------- >> begin captured stdout << ---------------------
[                  0%                  ] 2 of 3000 complete in 0.5 sec
[-----------------79%----------        ] 2382 of 3000 complete in 1.0 sec
[-----------------100%-----------------] 3000 of 3000 complete in 1.1 sec

--------------------- >> end captured stdout << ----------------------

Any idea what's going on?

[fonnesbeck]

I've been working on this this in the fix_travis branch. It has to do with Matplotlib trying to use a backend, but since this a remote task there is no display (hence the lack of a DISPLAY variable). I have poked around and tried various suggestions on StackOverflow, such as using the Agg backend in our tests, but nothing seems to work.

I'm pretty fed up with MPL, to be honest. I'm half tempted to explore other plotting options, such as D3, that we could simply bake into PyMC and drop this troublesome dependency.

[jsalvatier]

Can we mark the MPL stuff with known failure for now? Also, was this a
recent change? I thought these tests were passing before. Can we just
specify the version?

Have we talked to the MPL user/dev list? Since we're another package trying
to rely on them, maybe they will give us some help.

On Fri, Jan 24, 2014 at 7:10 AM, Chris Fonnesbeck
notifications@github.comwrote:

I've been working on this this in the fix_travis branch. It has to do
with Matplotlib trying to use a backend, but since this a remote task there
is no display (hence the lack of a DISPLAY variable). I have poked around
and tried various suggestions on StackOverflow, such as using the Aggbackend in our tests, but nothing seems to work.

I'm pretty fed up with MPL, to be honest. I'm half tempted to explore
other plotting options, such as D3, that we could simply bake into PyMC and
drop this troublesome dependency.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/474#issuecomment-33230104
.

[kyleam]

I'm surprised that using Agg (as done on the fix_travis branch) didn't
fix it. As long as it comes before any pyplot/pylab imports, I think it
should do the trick. I'm not having any luck finding the plot errors in
the Travis log files. Do all the jobs from the fix_travis branch have
them? I thought that test_plots.py runs during the
-e test_glm -e test_examples -e test_distributions command,
but I'm not seeing the plot failure.

[fonnesbeck]

The errors can be found pretty reliably on runs of the fix_travis branch. Here is one. I have gone through (in that branch) and added `matplotlib.use('Agg') at the beginning of all the tests that use plotting, so I am also surprised at what's going on.

[fonnesbeck]

@jsalvatier I'm sure we will track this down, but MPL is always such a moving target that issues like this come up a fair bit (going back to PyMC 2). Particularly when there are far more pressing development tasks, its a source of frustration for me, at least.

[kyleam]

OK, I may have found the issue. I think matplotlib.use('Agg') is not
getting called before pyplot is imported because when plots and glm
are imported in pymc/init.py, they import pyplot. Making the
following changes on top of master seems to fixes the test_plots.py run
for me.

diff --git i/pymc/__init__.py w/pymc/__init__.py
index 44760fd..5c5ef91 100644
--- i/pymc/__init__.py
+++ w/pymc/__init__.py
@@ -13,9 +13,9 @@ from .debug import *

 from .diagnostics import *

-from .plots import *
+#from .plots import *

 from .tests import test

-from . import glm
-from .data import * 
+# from . import glm
+from .data import *
diff --git i/pymc/tests/test_plots.py w/pymc/tests/test_plots.py
index f31b9a5..5d6f14f 100644
--- i/pymc/tests/test_plots.py
+++ w/pymc/tests/test_plots.py
@@ -1,4 +1,7 @@
 #from ..plots import *
+import matplotlib
+matplotlib.use('Agg')
+
 from pymc.plots import *
 from pymc import psample, Slice, Metropolis, find_hessian, sample

[kyleam]

Sorry for the formatting issues. I can't seem to fix them by editing the comment.

[jsalvatier]

If the import order in init is important let's put a comment mentioning why
On Jan 24, 2014 12:20 PM, "Kyle Meyer" notifications@github.com wrote:

Sorry for the formatting issues. I can't seem to fix them by editing the
comment.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/474#issuecomment-33257173
.

[kyleam]

I'm having all sorts of trouble commenting today. I'm just copying and pasting an email reply that didn't seem to make it through.

---

On 01/24/14 15:33, John Salvatier wrote:

If the import order in init is important let's put a comment mentioning why

It's not the order in pymc/__init__.py, but that pymc is loaded up (I
think by nose) before test_plots.py, where matplotlib.use('Agg') is
called, meaning that an interactive backend has already been selected.

The above patch commenting out the imports was just meant to show that
this was the issue. I think the best way to fix it should be discussed
because removing the imports would remove the plotting functions from
the top-level namespace.

[kyleam]

It seems that the recent Travis runs with the above changes passed (with respect to the MPL backend error).

I've pushed a branch with a proposed fix that still allows the top-level import of the plotting functions.

[jsalvatier]

Thank you :)

On Fri, Jan 24, 2014 at 8:47 PM, Kyle Meyer notifications@github.comwrote:

It seems that the recent Travis runs with the above changes passed (with
respect to the MPL backend error).

I've pushed a branchhttps://github.com/kyleam/pymc/tree/fix-mpl-travis-function-importswith a proposed fix that still allows the top-level import of the plotting
functions.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/474#issuecomment-33281177
.

[twiecki]

👍

[aflaxman]

This is so great, thanks! Nosetests now run to completion for me, although there are still some test failures, about data paths and theano functions:

Ran 152 tests in 3093.846s

FAILED (SKIP=2, errors=5)

[fonnesbeck]

I will test this and merge it, if that's okay. I'm experimenting with using conda for the Travis builds, as the install procedures tend to be faster.

[fonnesbeck]

OK, this works for me too. Errors now seem to be due to inplace_increment in Theano, but MPL-related issues are gone. Will merge.