Re: [Matplotlib-users] Combining pcolormesh and contour

[Andreas H. <li...@hi...>]

Am Di 28 Feb 2012 19:23:14 CET schrieb Eric Firing:
> On 02/28/2012 08:08 AM, Andreas H. wrote:
>> Am 28.02.2012 18:56, schrieb Eric Firing:
>>> On 02/28/2012 06:28 AM, Andreas H. wrote:
>>>>>> On Tuesday, February 28, 2012, Andreas H. wrote:
>>>>>>
>>>>>>> Good morning,
>>>>>>>
>>>>>>> I'm creating the attached plot using pcolormesh(). What I would like to
>>>>>>> do now is draw contour lines at +/- 2.5%, which follow the grid edges.
>>>>>>>
>>>>>>> The problem is that when I use contour(), the lines drawn do not follow
>>>>>>> the grid edges but seem to be interpolated somehow.
>>>>>>>
>>>>>>> Do you have an idea how to draw the contour lines following the grid
>>>>>>> edges?
>>>>>>>
>>>>>>> Your insight is very much appreciated :)
>>>>>>>
>>>>>>> Cheers,
>>>>>>> Andreas.
>>>>>>>
>>>>>>
>>>>>> This is because of a subtle difference in how pcolor-like functions and
>>>>>> contour-like functions work.  I always forget which is which, but one
>>>>>> assumes that the z value lies on the vertices of the grid while the
>>>>>> other
>>>>>> assumes that it lies in the middle of each grid point.  This is why you
>>>>>> see
>>>>>> them slightly offset from each other.
>>>>>
>>>>> Thanks, Ben!
>>>>>
>>>>> To `pcolormesh`, I pass the *edges* of the grid:
>>>>>
>>>>>       xbin = linspace(0, 12, nxbin + 1)
>>>>>       ybin = np.linspace(-90, 90, nybin + 1)
>>>>>
>>>>>       pl = spl.pcolormesh(xbin, ybin, pdata.T, cmap=cmap, edgecolors='None',
>>>>>                       vmin=-5, vmax=20)
>>>>>
>>>>> `contour`, however, wants the coordinates themselves. So I do
>>>>>
>>>>>       spl.contour((xbin[:-1]+xbin[1:])/2., (ybin[:-1]+ybin[1:])/2, pdata.T,
>>>>> [-2.5, 2.5])
>>>>>
>>>>> Still, the outcome is, well, unexpected to me. Actually, no matter if
>>>>> contour wants centres or edges, the actual behaviour seems strange. There
>>>>> is some interpolation going on, apparently. The input `pdata` has shape
>>>>> (12, 72) (or 72,12), and I definitely wouldn't expect this sub-grid
>>>>> movement in the x-direction.
>>>>>
>>>>> Any ideas?
>>>>
>>>> Okay, after some diving into matplotlib sources, I guess the interpolation
>>>> comes within the function `QuadContourSet._get_allsegs_and_allkinds`. So
>>>> there seems to be no way to accomplish what I actually want with the
>>>> current matplotlib API. Correct?
>>>>
>>>> If I wanted to do something about this, I would need to
>>>>
>>>> * implement a class `GriddedContourSet`, derived from `ContourSet`, where
>>>> I implement the `_get_allsegs_and_allkinds` method appropriately.
>>>> * add an additional keyword argument to `contour()` to make this gridded
>>>> contourset an option when calling `contour()`.
>>>>
>>>> Is this all correct? If yes, I might start working on this if I get the
>>>> time ...
>>>
>>> It is not at all clear to me what you want to do, as compared to what
>>> contour does.  Can you illustrate with an extremely simple example?
>>> Maybe even a scanned sketch, if necessary? Do you want the contour lines
>>> to be stepped, like the rectilinear boundaries of the pcolormesh
>>> cells--that is, composed entirely of horizontal and vertical line segments?
>>
>> Yes, Eric, that's exactly what I want. Since my case was simple enough,
>> I did it completely manually, with loads of calls to `plot` (I'm sure
>> there would've been a simpler solution ... -- which one?). I attached
>> the plot so you get an idea of what I want to do.
>
> Andreas,
>
> I have never seen a contour algorithm with an option to do that, but I 
> understand the motivation.  I don't think it would be easy to implement; 
> contouring algorithms generally are not.
>
> You might get an adequate approximation by using nearest-neighbor 
> interpolation to interpolate your data to a very fine grid, and then 
> contour that.

Eric,

thanks, that's a good hint. I took a look into the C source of the 
contour algorithm and decided not to bother right now. But your 
suggestion should do it.

Cheers,
Andreas.