[Matplotlib-devel] MEP 29 Axes Refactor

[Joe Kington]

My main concern is that this proposal would change the role of Axes.
Currently, it's primarily a container.

If I understand you correctly you don't like the idea of forcing Artists to
> use the transform.
>

Actually, my key problem is having the Axes responsible for going between
"data" coordinates and "screen"/"display" coordinates.  Right now, the Axes
is not at all responsible for that.  If I'm understanding this proposal
correctly, it would shift even more into Axes.

As far as I see it, all coordinates supplied to an Artist will come in the
> form of Axes coordinates, i.e. Axes space, and thus we need to transform
> those coordinates to screen coordinates...
>

The problem here is that most Artists don't work in data coordinates.  Only
a very small number, such as those produced by user-focused plotting
methods, are actually in data coordinates.  Most of the artists that belong
to the Axes actually don't have `ax.transData` as their transform.  For
example, consider the Axes patch, the ticks, tick labels, spines, titles
and labels, etc.  Even artists produced by user-facing functions often
aren't in "data" space. Consider `annotate`, for example.


>  however on a spherical geometry, these "straight" lines do not conform to
> the Euclid definition of straight, we need to draw them curved.  Because of
> this the transform will need to come very late in the drawing process.


For what it's worth, that's exactly the type of thing that I don't think it
makes sense to have the Axes responsible for.  Right now, Transform handles
this.  Specifically the `tranform_path_non_affine` method. This allows
these sort of transformations to happen at draw time.

I think the solution here comes from using a dual approach.  With the
> functions above I used the term base, as in the ``axes_to_base`` and
> ``base_to_axes`` functions.  Here ``base`` defines the projected
> coordinates.  We can then leave it up to the user to decide whether to plot
> in axes coordinates or base coordinates.  We can start of leaving the axes
> side unimplemented, and perhaps we will never implement the axes side for
> some Axes classes... what do you think?
>

For what it's worth, that's what we currently have, in my opinion.

Each artist has a Transform.  Its only role is to convert points and paths
(to handle non-euclidian geometries) from whatever the data for the artist
is in to display coordinates.  The Transform that the artist has defines
what coordinates system that artist's data is in.

This makes it easy to do things like:

import matplotlib.pyplot as plt

fig, ax = plt.subplots()

ax.text(0.1, 0.9, 'This will stay in-place', transform=ax.transAxes)

ax.plot(range(10), color='royalblue', lw=3)
ax.set(title='Be sure to pan and zoom')

plt.show()


Cartopy is another fantastic example.  Have a look at how Cartopy handles
the difference between projected coordinates and geodetic coordinates:
http://scitools.org.uk/cartopy/docs/latest/matplotlib/intro.html#adding-data-to-the-map
Which transform you specify defines which coordinate system is being used
for the data.

The key here is that the Axes isn't doing the transformation at all.
Making the Axes responsible for the transformation between "data" and
"display" coordinates makes it much harder to implement things such as
this.  There are a lot of artists that are logically a part of the Axes but
aren't in the "data" coordinate system.

Hopefully that helps clarify my viewpoint, at any rate.
Cheers!
-Joe

On Sat, Sep 5, 2015 at 3:10 PM, OceanWolf <juichenieder-nabb at yahoo.co.uk>
wrote:

>
>
>
>
> Not too sure what you meant by ````.
>
>
>
> On 05/09/15 19:25, Joe Kington wrote:
>
>
> First off, I don't intend this to come across as overly critical!  I think
> this is a very good discussion to have.
> Also, I tend to have a bad "knee-jerk" reaction to change and tend to come
> around over time, so keep that in mind too. :)
>
> No worries, I experience the same, and yes I wanted to open this up for
> just this kind of interrogation, especially as we have a lot of axes
> related code, and I have only touched a fraction of it to date.
>
> To explain where I come from, I should say that I like to work bottom-up.
> I find designing good code starts with asking probing questions about what
> you want to model, in this case we have an ``_AxesBase`` class and so by
> definition it should model an abstract Axes,  because of this in the
> "Detailed Description" of this MEP I begin by asking the question probing
> the definition of an Axes.  I believe that if we model the world
> intutitively as we see it, everything else will fall into place.  I find
> the most direct route in code usually contains lots of inflexibility, like
> building a road through the mountain, you might go the direct route, but it
> becomes very difficult to maintain and expand upon.  Hence the focus lies
> in the journey.
>
>
> However, while I agree that `Axes` is quite a beast, I'm not sure this
> proposal simplifies things.  From my perspective, it adds complexity. If
> I'm understanding correctly, this would effectively tie the Transform stack
> to the Axes, instead of having the Axes generate a Transform object that
> may or may not be used by the artists in the Axes.
>
> If I understand you correctly you don't like the idea of forcing Artists
> to use the transform.  I don't see this as a problem (at the moment).  As
> far as I see it, all coordinates supplied to an Artist will come in the
> form of Axes coordinates, i.e. Axes space, and thus we need to transform
> those coordinates to screen coordinates... at least at some point, probably
> when it comes to drawing... I especially think of drawing a triangle onto a
> spherical geometry, see https://en.wikipedia.org/wiki/Spherical_geometry.
> We create a Polygon patch and supply the three vertices that define our
> Triangle... however on a spherical geometry, these "straight" lines do not
> conform to the Euclid definition of straight, we need to draw them curved.
> Because of this the transform will need to come very late in the drawing
> process.
>
> First we define our coordinate transformation functions:
>> axes_to_base(self, *q) base_to_axes(self, x, y)
>>
>> The term base could get replaced with screen but for now we will keep it
>> simple to reflect another transformation from base coords to screen coords,
>> e.g. perhaps to differentiate between window and screen coords.
>>
>
> This is my main concern.  We have a (i.m.o.) very flexible and actually
> quite clean Transform system to handle this.  Why shift away from it?
> `ax.transData` may be non-PEP8 naming, but it's a good way to do this.  The
> concept of having Transform objects that handle this but are separate from
> the Axes gives a lot of flexibility.  In my opinion, the core concept of
> having this transformation handled by a Transform object that's separate
> from the Axes is one of the best things about matplotlib's design.
>
> Or am I misunderstanding, and this is just a refactoring of
> `_get_core_transform` and `_get_affine_transform` into one method?
>
> As far as I know, I want to keep the transform system.  I think I do just
> mean refactoring that into one method.  I say think as I still don't feel
> fully understand how it all works, the Transform system, brilliant, but
> very mind-boggling.  I had to delve into it to find a bug reported by a
> user on github, and went through around 50 (perhaps more) Transform
> operations before I got to the problem.  If you need to debug part of it,
> like i had to, it becomes a tangled mess, luckily for most people they
> don't have to, and the usage works quite simply.  When I tracked down the
> bug I also spent quite some time trying to figure out the Transform
> classes, prior to the bug I only knew of Rotation, Shear and Reflection
> Transforms.  Anyway my point here comes that while great, it can become
> quite the head-ache for the average user developer, especially for those
> who know even less then I do about transforms, and so I want to blackbox
> the transforms in the Axes with simple names such as
> axes_to_***_coords(self, *q), and ***_to_axes_coords(self, x, y).
>
> So I want to make it easy for people to write their own axes with their
> own transform methods without having to worry about how the rest of the
> Artist code and plot methods work (unless it works really bizarrely); and I
> want people to work on Artist code, and creating their own tools and user
> interaction stuff without having to worry about learning about transforms
> (they just need to know that these two methods will do the conversion for
> them from data coordinates, which they understand, to the location on the
> screen or whatever, which they will also understand, start talking about
> AffineTransforms and I think we will scare people off).
>
> ---------------------
>
> My other main concern centers on map projections. The MEP currently
> mentions:
>
> an anticipated structure of a base mapping class with a coordinate system
>> in lat/lon coordinates, but with different mapping projections available
>> for the conversion between the Axes coordinate system and the screen.
>>
>
> However, this is a bad approach for cartographic data. Geographic is not
> the base for a projected coordinate system. There are several reasons for
> that.
>
> 1. Map data is usually _in the projected coordinate system_.  Lat, long
> data is actually not terribly common unless you're working with global
> datasets.
> 2. Raster data (i.e. anything displayed with imshow) is typically going to
> be gridded on a regular grid in the projected coordinate system.  Forcing a
> transformation back to a non-uniform grid in lat, long space then back onto
> a different uniform grid than the original in display space is
> unnecessarily expensive.
>
> One of the great things about Cartopy is that it leaves the fundamental
> Cartesian projected space unchanged, and let's you specify the transform if
> you want to use geographic coordinates.  Basemap handles it a bit
> differently but has the same core concept.  Latitudes and longitudes aren't
> the data coordinate system.  The projected coordinate system is.
>
> There's a reason for that approach.  Forcing people to convert their data
> into a geographic coordinate system before plotting it is a bad idea. It's
> good to have plotting methods that allow geographic coordinates, but bad to
> require that transformation.  (I'll skip the very important datum part for
> the moment.  Just be aware that a lat, long only gets you to within ~1km of
> a location without more information.)
>
> Hmm, when I have used Basemap, the data files I work with I always get in
> lat/lon format.  One of my biggest annoyances with Basemap comes from
> having to work projection coordinates.  I move the mouse over the map and
> statusbar shows me useless projection coordinate information; I want to
> rotate the globe (in 'ortho' projection), but I can't, it becomes very
> difficult to use from a user interface point of view.
>
> I think the solution here comes from using a dual approach.  With the
> functions above I used the term base, as in the ``axes_to_base`` and
> ``base_to_axes`` functions.  Here ``base`` defines the projected
> coordinates.  We can then leave it up to the user to decide whether to plot
> in axes coordinates or base coordinates.  We can start of leaving the axes
> side unimplemented, and perhaps we will never implement the axes side for
> some Axes classes... what do you think?
>
> Best,
> OceanWolf
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/matplotlib-devel/attachments/20150906/ef0d46fb/attachment-0001.html>