Added GeoPandas data interface

[philippjfr]

Adds a GeoPandas interface allowing Path and Polygon types to wrap around geopandas dataframes directly. Here is a basic example:

https://anaconda.org/philippjfr/geopandas/notebook

Requires holoviews master and will likely be released at the same time as HoloViews 1.9.

[jbednar]

Excellent! Would it be difficult to make it work with Matt Rocklin's new cythonized geopandas equivalent as well, which appears to have vastly better performance than bare geopandas?

[philippjfr]

Would it be difficult to make it work with Matt Rocklin's new cythonized geopandas equivalent as well, which appears to have vastly better performance than bare geopandas?

I assume that will eventually be folded into geopandas proper? As long as the API is similar/the same it should be trivial either way. That said if you really need the optimization then your data is likely larger than bokeh can comfortably render.

[mrocklin]

No particular thoughts on this yet. We've been talking about getting coordinates out of GeoPandas a bit to speed up operations like this. I would not be surprised if we have a different API for this in the future, but I wouldn't wait on geopandas development.

I would be curious to know if you run into performance issues on larger datasets.

[philippjfr]

Probably won't have time to add tests here, I'll do that at the same time as moving the iris interface from HoloViews. Otherwise this is ready to review @jbednar. The PR also removes a lot of plotting code relying instead on an approach that simply projects the element and then uses the HoloViews version of the plot to actually plot the element. This means that in most cases plotting classes consist solely of a declaration of the projection operation. There are now projection operations for most elements and I'll finish off better Image and RGB handling in #87.

[philippjfr]

Can close #53 once merged.

[jbednar]

Looks good, apart from the failing test. It would be nice to add a sentence or two saying why to use GeoPandas, i.e. what you gain and tradeoffs there are if any.

[philippjfr]

Very confused by the latest error, dask seems to be raising an error about np.cbrt not being defined:

  File "/home/travis/miniconda/envs/test-environment/lib/python2.7/site-packages/dask/array/ufunc.py", line 138, in <module>
    cbrt = ufunc(np.cbrt)
AttributeError: 'module' object has no attribute 'cbrt'

[philippjfr]

Looks like something in our travis setup is downgrading numpy to numpy 1.9.3-py27h7e35acb_3

[jbednar]

That's odd; surely all Numpy versions have cbrt. Something's fishy.

[philippjfr]

No, np.cbrt is a new feature in NumPy 1.13 so it makes sense.

[jbednar]

Looks like cbrt was added in Numpy 1.10.0.

[jbednar]

So dask should be specifying numpy>=1.10, I guess?

[philippjfr]

Ah, sorry, I just searched the 1.13 changelog and didn't realize it listed all the changelogs.

[philippjfr]

So dask should be specifying numpy>=1.10, I guess?

I suppose so, I think we may have a clash with some other dependency though, we'll see in a minute.

[philippjfr]

In the .travis.yml you mean? I prefer pinning specific versions there, but it probably doesn't make much of a difference.

[jbednar]

It's a tradeoff, I guess, between mistakenly thinking that a change we made caused a problem, when really it was a new release of numpy that did it, versus not realizing that users will have problems from new numpy releases when they do happen.

[philippjfr]

Will go ahead and merge this so I can move on to #87.