MEP26: Artist styling

.. contents::

:local:

Status

**Proposed**

Branches and Pull requests

Abstract

This MEP proposes a new stylesheet implementation to allow more

comprehensive and dynamic styling of artists.

The current version of matplotlib (1.4.0) allows stylesheets based on

the rcParams syntax to be applied before creation of a plot. The

methodology below proposes a new syntax, based on CSS, which would

allow styling of individual artists and properties, which can be

applied dynamically to existing objects.

This is related to (and makes steps toward) the overall goal of moving

to a DOM/tree-like architecture.

Detailed description

Currently, the look and appearance of existing artist objects (figure,

axes, Line2D etc...) can only be updated via `set_` and `get_` methods

on the artist object, which is quite laborious, especially if no

reference to the artist(s) has been stored. The new style sheets

introduced in 1.4 allow styling before a plot is created, but do not

offer any means to dynamically update plots or distinguish between

artists of the same type (i.e. to specifiy the `line color` and `line

style` separately for differing `Line2D` objects).

The initial development should concentrate on allowing styling of

artist primitives (those `artists` that do not contain other

`artists`), and further development could expand the CSS syntax rules

and parser to allow more complex styling. See the appendix for a list

of primitives.

The new methodology would require development of a number of steps:

- A new stylesheet syntax (likely based on CSS) to allow selection of

artists by type, class, id etc...

- A mechanism by which to parse a stylesheet into a tree

- A mechanism by which to translate the parse-tree into something

which can be used to update the properties of relevant

artists. Ideally this would implement a method by which to traverse

the artists in a tree-like structure.

- A mechanism by which to generate a stylesheet from existing artist

properties. This would be useful to allow a user to export a

stylesheet from an existing figure (where the appearance may have

been set using the matplotlib API)...

Implementation

It will be easiest to allow a '3rd party' to modify/set the style of

an artist if the 'style' is created as a separate class and store

against the artist as a property. The `GraphicsContext` class already

provides a the basis of a `Style` class and an artists `draw` method can

be refactored to use the `Style` class rather than setting up it's own

`GraphicsContext` and transferring it's style-related properties to

it. A minimal example of how this could be implemented is shown here:

https://github.com/JamesRamm/mpl_experiment

IMO, this will also make the API and code base much neater as

individual get/set methods for artist style properties are now

redundant... Indirectly related would be a general drive to replace

get/set methods with properties. Implementing the style class with

properties would be a big stride toward this...

For initial development, I suggest developing a syntax based on a much

(much much) simplified version of CSS. I am in favour of dubbing this

Artist Style Sheets :+1: :

BNF Grammar

I propose a very simple syntax to implement initially (like a proof of

concept), which can be expanded upon in the future. The BNF form of

the syntax is given below and then explained ::

`ArtistIdent`, `Ident`, `Number` and `Colour` are tokens (the basic

building blocks of the expression) which are defined by regular

expressions.

Syntax

A CSS stylesheet consists of a series of **rule sets** in hierarchical

order (rules are applied from top to bottom). Each rule follows the

syntax ::

Each rule can have any number of `attribute`:`value` pairs, and a

stylesheet can have any number of rules.

The initial syntax is designed only for `artist` primitives. It does

not address the question of how to set properties on `container` types

(whose properties may themselves be `artists` with settable

properties), however, a future solution to this could simply be nested

`RuleSet`s

Selectors

Selectors define the object to which the attribute updates should be

applied. As a starting point, I propose just 2 selectors to use in

initial development:

Artist Type Selector

Select an `artist` by it's type. E.g `Line2D` or `Text`::

The regex for matching the artist type selector (`ArtistIdent` in the BNF grammar) would be::

GID selector

Select an `artist` by its `gid`::

A `gid` can be any string, so the regex could be as follows::

The above selectors roughly correspond to their CSS counterparts

(http://www.w3.org/TR/CSS21/selector.html)

Attributes and values

- `Attributes` are any valid (settable) property for the `artist` in question.

- `Values` are any valid value for the property (Usually a string, or number).

Parsing

Parsing would consist of breaking the stylesheet into tokens (the

python cookbook gives a nice tokenizing recipe on page 66), applying

the syntax rules and constructing a `Tree`. This requires defining the

grammar of the stylesheet (again, we can borrow from CSS) and writing

a parser. Happily, there is a recipe for this in the python cookbook

aswell.

Visitor pattern for matplotlib figure

In order to apply the stylesheet rules to the relevant artists, we

need to 'visit' each artist in a figure and apply the relevant rule.

Here is a visitor class (again, thanks to python cookbook), where each

`node` would be an artist in the figure. A `visit_` method would need

to be implemented for each mpl artist, to handle the different

properties for each ::

An `evaluator` class would then take the stylesheet rules and

implement the visitor on each one of them.

Backward compatibility

Implementing a separate `Style` class would break backward

compatibility as many get/set methods on an artist would become

redundant. While it would be possible to alter these methods to hook

into the `Style` class (stored as a property against the artist), I

would be in favor of simply removing them to both neaten/simplify the

codebase and to provide a simple, uncluttered API...

Alternatives

No alternatives, but some of the ground covered here overlaps with

MEP25, which may assist in this development

Appendix

Matplotlib primitives

This will form the initial selectors which stylesheets can use.

* Line2D

* Text

* AxesImage

* FigureImage

* Patch

MEP26