Superfluous matplotlib figures shown

[willfurnass]

Various statements in the MuMoTtest.ipynb Notebook result in bare, superfluous Figure objects being displayed before the Figures that those statements should produce. An example statement is:

bookmark1 = model4.bifurcation('s', 'A-B', initialState = {'U': 0.0, 'A': 1.0, 'B': 0.0}, initBifParam = 2.0, params = [('\\mu', 2.0), ('\\Delta', 2.0), ('plotLimits', 1), ('systemSize', 1)], initWigdets = {'\\mu': [3, 1.1, 5, 0.1], '\\Delta': [0.1, 0, 5, 0.1]}, bookmark = False)

Could there be a superfluous call to matplotlib.pyplot.figure() somewhere?

More info:

pip freeze | grep -E 'matplotlib|ipython|notebook'                                                                                                                                   
ipython==6.4.0                                                                                                                                                                               
ipython-genutils==0.2.0                                                                                                                                                                      
matplotlib==2.2.2                                                                                                                                                                            
notebook==5.6.0

Noticed using commit af774f5

[jarmarshall]

This is old behaviour - definitely using MumotEnv?

[willfurnass]

Nope, just the mandatory dependencies inc minimum versions listed in setup.py. Know what fixed this previously? Had a look for pertinent GH Issues but didn't find anything.

NB continuous integration runs currently install MuMoT plus dependencies in a virtualenv using the package info in setup.py; I could change this to use conda but want the dependency info in setup.py to be right, as that defines what's absolutely required to get expected functionality.

[jarmarshall]

@joefresna @tbose1 do you recall?

[jarmarshall]

This is probably to do with the version of matplotlib, ipywidgets, or another of the libraries. @tbose1 is on holiday this week but I can talk to him next week about it when we're at the same conference, I hope.

[willfurnass]

Yes, likely. I tried forcing matplotlib, ipywidgets and sympy versions in setup.py so they matched those in environment.yml but that didn't help.

[jarmarshall]

Running a build from today via MyBinder does not exhibit this problem, as far as I can see...

[willfurnass]

MyBinder uses the conda environment.yml. The problem seems to be with a virtualenv created with the minimal set of essential dependencies defined in setup.py.

[tbose1]

The original double-figure problem was solved with commit 473e975 (committed on 3rd May 2018).

I remember trying a lot of things.
The solution that worked in MumotEnv was to modify the following function _buildFig() (essentially this means turning interactive mode on and off with plt.ion() and plt.ioff() ):

def _buildFig(object, figure = None):
    global figureCounter
    object._figureNum = figureCounter
    if figureCounter==1:
        plt.ion()
    else:
        plt.ioff()
    figureCounter += 1
    with warnings.catch_warnings(): # ignore warnings when plt.hold has been deprecated in installed libraries - still need to try plt.hold(True) in case older libraries in use
        warnings.filterwarnings("ignore",category=MatplotlibDeprecationWarning)
        warnings.filterwarnings("ignore",category=UserWarning)
        plt.hold(True)  
    if figure == None:
        if figureCounter>2:
            plt.ion()
        object._figure = plt.figure(object._figureNum) 
    else:
        object._figure = figure

This function replaced the previous version ( renamed to _buildFigOLD() ):

def _buildFigOLD(object, figure = None):
    global figureCounter
    object._figureNum = figureCounter
    figureCounter += 1

[willfurnass]

@tbose1 Can you add a docstring for _buildFig explaining what the function does.

[tbose1]

Added docstring to buildFig in commit eb90d05.

[jarmarshall]

The environments created via pip and conda differ markedly in the library versions installed - not being expert in pip or conda I am not sure of how to ensure consistency - I have briefly tried downgrading matplotlib, ipython, and ipywidgets to the versions in environment.yml - this failed to solve the problem and created new problems. Is there an elegant way of copying a minimal working environment to requirements.txt @willfurnass ?

[willfurnass]

Here's a maximal working requirements.txt. Not yet had chance to pick apart which package(s)/versions are required to avoid the duplicate figure issue...

[jarmarshall]

I'm a bit confused now - is requirements.txt deprecated, and all requirements should go in setup.py?

[jarmarshall]

I've found the old issue this cropped up in (#108)
I've copied those requirements in full into setup.py in my branch https://github.com/jarmarshall/MuMoT
There are various build errors, and then a notebook error that looks familiar, but I can't find the relevant issue:

TraitError: The 'selected_index' trait of an Accordion instance must be an int, but a value of None <class 'NoneType'> was specified.

@willfurnass it seems we're just chasing our tails now trying to get virtualenv working - you said you could change the continuous integration to work with conda - I think we should do that and only support conda installation at this point.

[jarmarshall]

I've noticed something else - in the virtualenv installation figures are no longer in interactive mode (pannable, etc.) - this may be related - have tried installing ipympl and using the matplotlib widget magic but this hasn't made a difference - want to check if the conda installation is still in interactive mode but the MyBinder site has just spent 10 minutes failing to build...

[jarmarshall]

Having checked the conda installation and MyBinder deployment too, both work in interactive mode for plots correctly - by contrast the virtualenv approach does not, and I think this is the root of this issue. It seems that virtualenvs may not play nicely with matplotlib, e.g. see https://matplotlib.org/faq/virtualenv_faq.html (although not sure when that was written) - I recommend we only support the conda installation route and change the continuous integration over to use this @willfurnass

[willfurnass]

Addressed by #210

[tbose1]

Installation of MuMoT within the mumot-env conda environment following the documentation brings back the superfluous figure problem on my system (macOS), i.e. each time a slider is changed a new figure appears below the previous one. Also the rendering of the figures is affected. This is probably due to pinning matplotlib to a version <3.0 which currently installs matplotlib=2.2.3. Pinning the version to matplotlib=2.1.2 seems to solve this problem. The most recent version is 3.0.2 but this appears to be incompatible with other dependencies.

[tbose1]

Update: pinning matplotlib to another version doesn't help. I'm working on another solution.

[tbose1]

Solved with commit 97e1074. Added the following to the MuMoT code (chooses TkAgg backend if mac-user):

if sys.platform == "darwin": #if operating system is macOS
import matplotlib
matplotlib.use('TkAgg')

[willfurnass]

@tbose1 I suggest also adding comments above those three lines to explain why they're needed.

[tbose1]

Added comments via 6d9520c.