tkagg display error

[satra]

@mwaskom: shouldn't this not happen with tkagg?

  File "/srv/software/python/EPD/virtualenvs/7.2/nipype0.5/lib/python2.7/site-packages/nipype/algorithms/rapidart.py", line 357, in _detect_outliers_core
    fig = plt.figure()
  File "/software/python/EPD/7.2/lib/python2.7/site-packages/matplotlib/pyplot.py", line 343, in figure
    **kwargs)
  File "/software/python/EPD/7.2/lib/python2.7/site-packages/matplotlib/backends/backend_tkagg.py", line 80, in new_figure_manager
    window = Tk.Tk()
  File "/software/python/EPD/7.2/lib/python2.7/lib-tk/Tkinter.py", line 1685, in __init__
    self.tk = _tkinter.create(screenName, baseName, className, interactive, wantobjects, useTk, sync, use)
TclError: couldn't connect to display "computor:1.0"

[mwaskom]

The Tk backend still needs a display to draw, I would assume? To draw without a display, use the Agg backend.

[satra]

hmmm ... so that config option is not being executed properly .. let me take a look.

[mwaskom]

If you're running this from e.g. IPython and import pyplot before importing nipype, the nipype config option will not take effect.

[satra]

in some recent changes i took out the matplotlib lines that set agg in the interface code. i'll push a fix shortly.

[djarecka]

@satra : i don't know how to reproduce the issue and it's pretty old, but i'm guessing that matplotlib.use('Agg') or matplotlib.use(config.get('execution', 'matplotlib_backend')) solved problems in most of the cases.
I can see a few places that import matplotlib and don't set backend (e.g. https://github.com/nipy/nipype/blob/master/nipype/algorithms/metrics.py#L141). i can search the code and matplotlib.use(config.get('execution', 'matplotlib_backend'))

[oesteban]

@satra @djarecka What if we clear up all the matplotlib backend options and recommend users to change the matplotlib default? - https://github.com/nipy/nipype/blob/master/Dockerfile#L71

[djarecka]

Not sure what is easier to "maintain". There are only a few files where you import matplotlib, so can review if you decide to keep the matplotlib backend options in the code.

The only thing I'm not completely sure is matplotlib.is_interactive(), should we ensure to be False regardless of backend? The default is False, so it's usually False anyway, but I'm not sure if this can also cause a problem.

[djarecka]

@satra , @oesteban do we still want to do anything about it?