Use arviz for plotting

[ColCarroll]

This starts moving pymc3 plotting over to arviz. Other things I'd like to include in this PR:

1. Update setup.py to include arviz as optional (pip install pymc3[all] or pip install pymc3[plots]?)
2. Rerun at least a few notebooks to make sure the graphs are actually well supported.

This is still WIP because of arviz-devs/arviz#529

I am also making an issue for plot_posterior_predictive_glm, which is heavily used in the GLM notebooks, but not yet supported by arviz. I think there's interest in doing so, though!

Here's an example running under this PR (and the arviz PR). Note that as a side effect of this change, traceplot is more flexible.

[twiecki]

[eigenfoo]

Deleted code is debugged code!!

[aloctavodia]

The documentation will be updated as part of this PR, or a separated one?

[canyon289]

Wellll, the code has just shifted :)

But it has become more general purpose which is still a win!

[ColCarroll]

• Fixed f-string
• The API docs for plots will now include a link to the ArviZ documentation
• pip install pymc3[plots] will include arviz, otherwise it remains optional

[fonnesbeck]

Its tempting to reconcile the names, but I suppose we should just leave things be.

[ColCarroll]

from running through the first notebook, varnames --> var_names is the worst right now. I am a little tempted to add a decorator just for that case:

def map_args(func):
    @functools.wraps(func)
    def wrapped(*args, **kwargs):
        if 'varnames' in kwargs and 'var_names' not in kwargs:
            warnings.warn('Keyword argument `varnames` renamed to `var_names`, and will be removed in pymc3 3.8.0')
            kwargs['var_names'] = kwargs.pop('varnames')
        return func(*args, **kwargs)
    return wrapped

Thoughts on that?

[twiecki]

I think that's a good idea.

[twiecki]

As the plan is to eventually remove this, we should probably also raise deprecation warnings. Ideally with the new call if that's not too cumbersome to do.

[ColCarroll]

Wait... is the plan to eventually remove this? I guess in favor of just using arviz directly?

[fonnesbeck]

I think we should have pm.plot_trace (for example) that is pure Arviz and then have a mapping to pm.traceplot that converts the arguments as well, so that it behaves just as traceplot currently does.

I don't think we should deprecate traceplot, it should just remain available, but the documentation will describe plot_trace going forward. This will allow us to make the changes we want without breaking old code (or littering output with deprecations).

[ColCarroll]

ah, neat -- I think I can just dump az.plots.__all__ into pm.plot.__all__ to get the "pure arviz" part of things. Then we can do custom stuff to pm.traceplot and the like.

[ColCarroll]

Sorry for the long pause. I addressed these two issues:

• both original names and arviz names work now. if we want to customize a plot, we can override the pymc3 name
• I put a small deprecation warning function in (the same one from this discussion)

Here's a traceplot using both pm.plot_trace and pm.traceplot(..., var_names=...):

[fonnesbeck]

Is this still a WIP?

[ColCarroll]

Oops, changed the title - I think it is ready. Tab completion and docs come from arviz, and look ok to me.

I will mess with the tests tomorrow to get them passing. I haven't played around too much without arviz installed, which might be causing CI problems.

[twiecki]

+++pip install -e .
Obtaining file:///home/travis/build/pymc-devs/pymc3
    Complete output from command python setup.py egg_info:
    Traceback (most recent call last):
      File "<string>", line 1, in <module>
      File "/home/travis/build/pymc-devs/pymc3/setup.py", line 66
        tests_require=test_reqs,
                    ^
    SyntaxError: invalid syntax
    
    ----------------------------------------

[twiecki]

Looks great. Needs a prominent line in the release notes.

We should probably also fix the docs, but certainly can do that in a separate PR.

[junpenglao]

We should have a docathon week for update all docs ;-)

[ColCarroll]

Looks like there was a surprising test failure because we are now picking up numpy 1.16.

[ColCarroll]

This looks ready to merge again. There was a travis failure from trying to send data to coveralls.

[ericmjl]

I'm hitting the big green button!

[ericmjl]

Thanks @ColCarroll!

[twiecki]

Niceeeee!