MEP10: Modernizing the documentation

Michael Droettboom <mdroe@...552...> writes:

> Michael Droettboom <mdroe@...552...> writes:
>> Working with the documentation this past week has me a little
>> frustrated with the state of it. Enough to write a MEP.
>> In particular, it would be nice to compile a list of concerns about
>> the docstrings and documentation layout so that we can address as
>> much as possible in a single pass. Also, let me know if there are
>> any relevant PRs and Issues.
>> In particular, I think PR #1032, as it is a large structural
>> reorganization, my dovetail well with the proposed reorganization of
>> the docs.
>> Mike
> The proposal looks great. I would like to comment on one issue that it


> and which I found very uncomfortable to work with as a newcomer. I think


> matplotlib style of using *args and **kwargs for delegation of arguments is


> rather bad practice, which is hard to solve by just updating documentation.


> breaks many rules of pep 20: it is implicit, since it is not allowing
> introspection, it is nested, since it always involves nested calls, it


> for alternative ways to do things, and I also don't think it's anyhow


> Most of the things passed with *args, **kwargs can be done with an added
> function call, like:
> points = ax.scatter(data)
> points.update(*args, **kwargs)
> What would be the disadvantage of abolishing this practice?
I understand the comments about the difficulty of introspection. The
reason it works the way it does is so that additional parameters can be
added to the artist layer without needing to update every single
plotting function. A real world example of this is when hatching was
added -- that feature only had to be added in one place and most artists
were able to use it. In that sense, I think this approach is very
beautiful in terms of code maintainability and extensibility.

First of all, to be clear, I am not trying to be negative about the current
situation, I rather search for ways to improve on it. I also realize that any
changes like this would be very serious and require a lot of thinking and of
course work. Let me first point some drawbacks of the current kwargs usage
before discussing the solution. Let me know if you agree.

1. Currenly the aims of extensibility and maintainability are not completely
reached, as seen from this part of Axes.scatter docstring:

        Optional kwargs control the
        :class:`~matplotlib.collections.Collection` properties; in

            The string 'none' to plot faces with no outlines

            The string 'none' to plot unfilled outlines

So the explicit descriptions of useful **kwargs are provided. If a new useful
property of collections appears, it stays unknown.

2. Yet another problem that **kwargs cause is that it is sometimes completely
unclear where an argument ends up. `Axes` methods have a reasonably well-
documented kwargs, while, e.g. axis module often fails to mention what happens
to kwargs. Another example of misuse is figure.add_axes, where the docstring

"""...kwargs are legal Axes kwargs plus projection which sets the projection
type of the axes. (For backward compatibility, polar=True may also be provided,
which is equivalent to projection='polar'). Valid values for projection are:
[‘aitoff’, ‘hammer’, ‘lambert’, ‘mollweide’, ‘polar’, ‘rectilinear’]. Some of
these projections support additional kwargs, which may be provided to

Despite this is an extensible interface, I still wouldn't call it user-friendly;
kwargs are passed to at least two places: Axes and projection.

3. About the hatching: I think this is solved in a rather elegant fashion with
symbols for scatter plot: the interpretation of the format is deferred to
`mmarkers.MarkerStyle(marker)`. This means that `marker` is in fact already
effectively a style object, corresponding to the style object format that you
suggested. One can also easily refer to the MarkerStyle docstring in order to
make the description complete.

4. The idea behind kwargs, if I understand it correctly, is to let high-level
functions, such as all the plot functions handle low-level objects in full
generality. This aim cannot be achieved anyway, because high-level functions
often generate more than one low-level object. This is why something like this:

axis.label.set_text(r"$\theta = 60^{\circ}$")

Is still provided as a separate command, and cannot be done via kwargs to
scatter. So the user already doesn't have a situation where a single function
call is sufficient.


On 08/26/2012 05:33 AM, Anton Akhmerov wrote:


These altogether seem serious enough to consider changing the behavior in one or
another way. They are often also directly related to documentation quality.