Re: [matplotlib-devel] Empty scatter plot

[Eric F. <ef...@ha...>]

On 02/15/2011 08:50 AM, Benjamin Root wrote:
> On Tue, Feb 15, 2011 at 12:19 PM, Benjamin Root <ben...@ou...
> <mailto:ben...@ou...>> wrote:
>
>     On Tue, Feb 15, 2011 at 11:54 AM, Eric Firing <ef...@ha...
>     <mailto:ef...@ha...>> wrote:
>
>         On 02/15/2011 07:40 AM, Benjamin Root wrote:
>          > I have come across a little inconsistency that was unexpected
>         in the
>          > matplotlib API.  The following is perfectly valid:
>          >
>          > import matplotlib.pyplot as plt
>          > plt.plot([], [])
>          > plt.show()
>          >
>          >
>          > However, this is not valid:
>          >
>          > import matplotlib.pyplot as plt
>          > plt.scatter([], [])
>          > plt.show()
>          >
>          >
>          > The immediate issue that comes up in scatter is that it
>         attempts to find
>          > min/max of the input data for the purpose of autoscaling
>         (this can
>          > probably be done better by just using set_xmargin(0.05) and
>          > set_ymargin(0.05)).  This can easily be bypassed with an if
>         statement.
>          > However, then we discover that polygon collection do not like
>         having
>          > empty offsets, which leads to a failure in the affine
>         transformation.
>          >
>          > So, the question is, is this a bug or a feature?  I
>         personally believe
>          > that empty data is a perfectly valid scenario and given that
>         other
>          > matplotlib functions handle it gracefully, we should make the
>          > collections object more friendly to empty data.
>
>         I agree; a call with empty data should simply not plot anything.
>
>         Eric
>
>
>     Digging further, it appears that the problem is in _path.cpp for
>     _path_module::affine_transform() which explicitly checks for an
>     empty vertices array and throws an exception if it is empty.
>
>     So, do we want to make _path.cpp empty-friendly or should we just
>     make empty collections objects just avoid doing anything that
>     requires doing an affine transform?
>
>     Ben Root
>
>
>
> Ok, some more digging deepens the mystery.  While an empty-friendly
> _path.cpp would be nice, it appears that the collections and axes
> objects are already doing all it can to avoid doing transforms for empty
> collections.
>
> However, it appears that the supposedly empty collection object from
> scatter([], []) is not really empty.  Its self._paths member contains a
> list of unit_circle() from Path.  This is also the case for
> EllipseCollection.  Meanwhile, LineCollection and PatchCollection
> initializes their self._paths in accordance to their given data.

One way to solve the problem would be to start each draw() method with a 
short-circuit return in case there is nothing to draw.  It would be 
needed only for classes for which empty self._paths is not a valid test. 
  So for CircleCollection it would be:

     @allow_rasterization
     def draw(self, renderer):
         # sizes is the area of the circle circumscribing the polygon
         # in points^2
         if len(self._sizes) == 0:
             return
         self._transforms = [
             transforms.Affine2D().scale(
                 (np.sqrt(x) * self.figure.dpi / 72.0) / np.sqrt(np.pi))
             for x in self._sizes]
         return Collection.draw(self, renderer)

(Collection.draw returns nothing, so there is no inconsistency in the 
return value.)

Alternatively, it looks like maybe an empty self._transforms could be 
used in a short-circuit test at the start of Collection.draw.

Eric


>
> Ben Root
>