Gadfly is an implementation of a "grammar of graphics" style statistical graphics system for Julia. This tutorial will outline general usage patterns and will give you a feel for the overall system.
To begin, we need some data. Gadfly works best when the data is supplied in a DataFrame. In this tutorial, we'll pick and choose some examples from the RDatasets package.
Let us use Fisher's iris dataset as a starting point.
using Gadfly using RDatasets iris = dataset("datasets", "iris")
plot function in Gadfly is of the form:
plot(data::DataFrame, mapping::Dict, elements::Element...)
The first argument is the data to be plotted, the second is a dictionary mapping "aesthetics" to columns in the data frame, and this is followed by some number of elements, which are the nouns and verbs, so to speak, that form the grammar.
Let's get to it.
p = plot(iris, x=:SepalLength, y=:SepalWidth, Geom.point);
This produces a
Plot object. It can be saved to a file by drawing to one or more backends using
img = SVG("iris_plot.svg", 6inch, 4inch) draw(img, p)
Now we have the following charming little SVG image.
If you are working at the REPL, a quicker way to see the image is to omit the semi-colon trailing
plot. This automatically renders the image to your default multimedia display, typically an internet browser. No need to capture the output argument in this case.
plot(iris, x=:SepalLength, y=:SepalWidth, Geom.point)
Alternatively one can manually call
display on a
Plot object. This workflow is necessary when
display would not otherwise be called automatically.
function get_to_it(d) ppoint = plot(d, x=:SepalLength, y=:SepalWidth, Geom.point) pline = plot(d, x=:SepalLength, y=:SepalWidth, Geom.line) ppoint, pline end ps = get_to_it(iris) map(display, ps)
For the rest of the demonstrations, we'll simply omit the trailing semi-colon for brevity.
In this plot we've mapped the x aesthetic to the
SepalLength column and the y aesthetic to the
SepalWidth. The last argument, Geom.point, is a geometry element which takes bound aesthetics and render delightful figures. Adding other geometries produces layers, which may or may not result in a coherent plot.
plot(iris, x=:SepalLength, y=:SepalWidth, Geom.point, Geom.line)
This is the grammar of graphics equivalent of "colorless green ideas sleep furiously". It is valid grammar, but not particularly meaningful.
Let's do add something meaningful by mapping the color aesthetic.
plot(iris, x=:SepalLength, y=:SepalWidth, color=:Species, Geom.point)
Ah, a scientific discovery: Setosa has short but wide sepals!
Color scales in Gadfly by default are produced from perceptually uniform colorspaces (LUV/LCHuv or LAB/LCHab), though it supports RGB, HSV, HLS, XYZ, and converts arbitrarily between these. Of course, CSS/X11 named colors work too: "old lace", anyone?
Scale transforms also work as expected. Let's look at some data where this is useful.
mammals = dataset("MASS", "mammals") plot(mammals, x=:Body, y=:Brain, label=:Mammal, Geom.point, Geom.label)
This is no good, the large animals are ruining things for us. Putting both axis on a log-scale clears things up.
plot(mammals, x=:Body, y=:Brain, label=:Mammal, Geom.point, Geom.label, Scale.x_log10, Scale.y_log10)
Since all continuous analysis is just degenerate discrete analysis, let's take a crack at the latter using some fuel efficiency data.
gasoline = dataset("Ecdat", "Gasoline") plot(gasoline, x=:Year, y=:LGasPCar, color=:Country, Geom.point, Geom.line)
We could have added Scale.x_discrete explicitly, but this is detected and the right default is chosen. This is the case with most of elements in the grammar: we've omitted Scale.x_continuous and Scale.y_continuous in the previous plots, as well as Coord.cartesian, and guide elements such as Guide.xticks, Guide.xlabel, and so on. As much as possible the system tries to fill in the gaps with reasonable defaults.
Gadfly uses a custom graphics library called Compose, which is an attempt at a more elegant, purely functional take on the R
grid package. It allows mixing of absolute and relative units and complex coordinate transforms. The primary backend is a native SVG generator (almost native: it uses pango to precompute text extents), though there is also a Cairo backend. See Backends for more details.
Building graphics declaratively let's you do some fun things. Like stick two plots together:
fig1a = plot(iris, x="SepalLength", y="SepalWidth", Geom.point) fig1b = plot(iris, x="SepalWidth", Geom.bar) fig1 = hstack(fig1a, fig1b)
Ultimately this will make more complex visualizations easier to build. For example, facets, plots within plots, and so on. See Layers and Stacks for more details.
Though not a replacement for full-fledged custom interactive visualizations of the sort produced by d3, this sort of mild interactivity can improve a lot of standard plots. The fuel efficiency plot is made more clear by toggling off some of the countries, for example.