The R Project SVN R

\documentclass[a4paper]{article}
%\VignetteIndexEntry{Writing grid Code}
%\VignettePackage{grid}
\newcommand{\code}[1]{\texttt{#1}}
\newcommand{\pkg}[1]{{\normalfont\fontseries{b}\selectfont #1}}
\newcommand{\grid}{\pkg{grid}}
\newcommand{\grob}{\code{grob}}
\newcommand{\gTree}{\code{gTree}}
\newcommand{\R}{{\sffamily R}}
\setlength{\parindent}{0in}
\setlength{\parskip}{.1in}
\setlength{\textwidth}{140mm}
\setlength{\oddsidemargin}{10mm}

\newcommand{\aside}[1]{\begin{list}{}
                                   {\setlength{\leftmargin}{1in}
                                    \setlength{\rightmargin}{1in}
                                    \setlength{\itemindent}{0in}}
                       \item {\sc Aside:} {\it #1}
                       \end{list}}

\title{Writing \grid{} Code}
\author{Paul Murrell}

\begin{document}
\maketitle

<<echo=FALSE, results=hide>>=
library(grDevices)
library(stats) # for runif()
library(grid)
ps.options(pointsize=12)
options(width=60)

@
The \grid{} system contains a degree of complexity in order
to allow things like editing graphical objects, ``packing'' graphical
objects, and so on.  This means that many of the
predefined Grid graphics functions are
relatively complicated\footnote{Although there are exceptions;  some
functions, such as \code{grid.show.viewport}, are purely for producing
illustrative diagrams and remain simple and procedural.}.

One design aim of \grid{} is to allow users to create simple
graphics simply and not to force them to use  complicated concepts or
write complicated code
 unless they actually need to.  Along similar lines, it is
intended that people should be able to prototype even complex graphics
very simply and then refine the implementation into a more
sophisticated form if necessary.

With the predefined graphics
functions being fully-developed and complicated implementations,
 there is a lack of examples of simple, prototype code.
Furthermore, given that the aim is to allow a range of ways to produce the same
graphical output, there is a need for examples which demonstrate the
various stages, from simple to complex,
that a piece of \grid{} code can go through.

This document describes the
construction of a scatterplot object, like that shown below,
going from the simplest, prototype
implementation to the most complex and sophisticated.
It demonstrates that if you
only want simple graphics output then you can do it pretty
simply and quickly.  It also demonstrates how to write functions
that allow your graphics to be used by other people.  Finally, it
demonstrates how to make your graphics fully interactive (or at least
as interactive as Grid will let you make it).

@
This document should be read {\em after} the \grid{} Users'
Guide.  Here we are assuming that the reader has an understanding
of viewports, layouts, and units.  For the later sections of the
document, it will also be helpful to have an understanding of
\R{}'s \code{S3} object system.

\section*{Procedural \grid{}}

The simplest way to produce graphical output in Grid is just
like producing standard R graphical output.  You simply issue a
series of graphics commands and each command adds more ink to the
plot.
The purpose of the commands is simply to produce
graphics output;  in particular, we are not concerned with any
values returned by the plotting functions.  I will call this
{\it procedural graphics}.

In order to draw a simple scatterplot, we can  issue a series
of commands which draw the various components of the plot.

Here are some random data to plot.

<<>>=
x <- runif(10)
y <- runif(10)

@
\noindent
The first step in creating the plot involves defining a ``data'' region.
This is a region which has sensible scales on the axes for plotting the
data and margins around the outside
for the axes to fit in, with a space for a title at the top.

<<datavp>>=
data.vp <- viewport(x=unit(5, "lines"),
                    y=unit(4, "lines"),
                    width=unit(1, "npc") - unit(7, "lines"),
                    height=unit(1, "npc") - unit(7, "lines"),
                    just=c("left", "bottom"),
                    xscale=range(x) + c(-.05, .05)*diff(range(x)),
                    yscale=range(y) + c(-.05, .05)*diff(range(y)))

@
\noindent
Now we create the data region and
draw the components of the plot relative to it:
points, axes, labels, and a
title.

<<procplot>>=
pushViewport(data.vp)
grid.points(x, y)
grid.rect()
grid.xaxis()
grid.yaxis()
grid.text("x axis", y=unit(-3, "lines"),
          gp=gpar(fontsize=14))
grid.text("y axis", x=unit(-4, "lines"),
          gp=gpar(fontsize=14), rot=90)
grid.text("A Simple Plot",
          y=unit(1, "npc") + unit(1.5, "lines"),
          gp=gpar(fontsize=16))
popViewport()

<<fig=TRUE, echo=FALSE, results=hide>>=
<<procplot>>

@
\section*{Facilitating Annotation}

Issuing a series of commands to produce a plot, like in the previous
section, allows the user to have a great deal of flexibility.
It is always possible to recreate viewports in order to add
further annotations.  For example, the following code
recreates the data region in order to place the date
at the bottom right corner.

<<ann1>>=
pushViewport(data.vp)
grid.text(date(), x=unit(1, "npc"),
  y = 0, just=c("right", "bottom"), gp=gpar(col="grey"))
popViewport()

<<fig=TRUE, echo=FALSE, results=hide>>=
<<procplot>>
<<ann1>>

@

When more complex arrangements of viewports are involved, there may be
a bewildering array of viewports created, which may make it difficult
for other users to revisit a particular region of a plot.  A \code{lattice}
plot is a good example.
In such cases, it will be more cooperative to use \code{upViewport()}
rather than \code{popViewport()} and leave the viewports that were
created during the drawing of the plot.  Other users can then use
\code{vpPath}s to navigate to the desired region.  For example, here
is a slight modification of the original series of commands, where
the original data viewport is given a name and
\code{upViewport()} is used at the end.

<<results=hide>>=
data.vp <- viewport(name="dataregion",
                    x=unit(5, "lines"),
                    y=unit(4, "lines"),
                    width=unit(1, "npc") - unit(7, "lines"),
                    height=unit(1, "npc") - unit(7, "lines"),
                    just=c("left", "bottom"),
                    xscale=range(x) + c(-.05, .05)*diff(range(x)),
                    yscale=range(y) + c(-.05, .05)*diff(range(y)))
pushViewport(data.vp)
grid.points(x, y)
grid.rect()
grid.xaxis()
grid.yaxis()
grid.text("x axis", y=unit(-3, "lines"),
          gp=gpar(fontsize=14))
grid.text("y axis", x=unit(-4, "lines"),
          gp=gpar(fontsize=14), rot=90)
grid.text("A Simple Plot",
          y=unit(1, "npc") + unit(1.5, "lines"),
          gp=gpar(fontsize=16))
upViewport()

@

The date is now added
using \code{downViewport()} to get to the data region.

<<results=hide>>=
downViewport("dataregion")
grid.text(date(), x=unit(1, "npc"),
  y = 0, just=c("right", "bottom"), gp=gpar(col="grey"))
upViewport()

@
\section*{Writing a \grid{} Function}

Here is the scatterplot code wrapped up as a simple function.

<<funcplot>>=
splot <- function(x=runif(10), y=runif(10), title="A Simple Plot") {
data.vp <- viewport(name="dataregion",
                    x=unit(5, "lines"),
                    y=unit(4, "lines"),
                    width=unit(1, "npc") - unit(7, "lines"),
                    height=unit(1, "npc") - unit(7, "lines"),
                    just=c("left", "bottom"),
                    xscale=range(x) + c(-.05, .05)*diff(range(x)),
                    yscale=range(y) + c(-.05, .05)*diff(range(y)))
pushViewport(data.vp)
grid.points(x, y)
grid.rect()
grid.xaxis()
grid.yaxis()
grid.text("y axis", x=unit(-4, "lines"),
          gp=gpar(fontsize=14), rot=90)
grid.text(title,
          y=unit(1, "npc") + unit(1.5, "lines"),
          gp=gpar(fontsize=16))
upViewport()
}

@
There are several advantages to creating a
function:
\begin{enumerate}
\item We get the standard advantages of a function:
we can reuse and maintain the plot code more easily.
\item We can slightly generalise the plot.  In this case, we can use it for
different data and have a different title.  We could
add more arguments to allow different margins, control over
the axis scales, and so on.
\item The plot can be embedded in other graphics output.
\end{enumerate}
Here is an example which uses the \code{splot()} function to
create a slightly modified scatterplot, embedded within
other \grid{} output.

<<embed, fig=TRUE, results=hide>>=
grid.rect(gp=gpar(fill="grey"))
message <- paste("I could draw all sorts",
  "of stuff over here",
  "then create a viewport",
  "over there and stick",
  "a scatterplot in it.", sep="\n")
grid.text(message, x=0.25)
grid.lines(x=unit.c(unit(0.25, "npc") + 0.5*stringWidth(message) +
  unit(2, "mm"),
  unit(0.5, "npc") - unit(2, "mm")),
  y=0.5,
  arrow=arrow(angle=15, type="closed"),
  gp=gpar(lwd=3, fill="black"))
pushViewport(viewport(x=0.5, height=0.5, width=0.45, just="left",
  gp=gpar(cex=0.5)))
grid.rect(gp=gpar(fill="white"))
splot(1:10, 1:10, title="An Embedded Plot")
upViewport()

@

It is still straightforward to annotate the scatterplot as long as
we have enough information about the viewports.  In this case,
a non-strict \code{downViewport()} will still work (though note
that \code{upViewport({\bf 0})} is required to get right back to the
top level).

<<ann2, echo=FALSE, eval=FALSE>>=
downViewport("dataregion")
grid.text(date(), x=unit(1, "npc"),
  y = 0, just=c("right", "bottom"), gp=gpar(col="grey"))
upViewport(0)

<<echo=FALSE, results=hide>>=
<<embed>>
<<ann2>>

@
\section*{Creating \grid{} Graphical Objects}

A \grid{} function like the one in the previous section provides
output which is very flexible and can be annotated in arbitrary ways
and can be embedded within other output.  This is likely
to satisfy most uses.

However, there are some things that cannot be done (or at least would
be extremely hard to do) with such a function.  The output produced by
the function cannot be addressed as a coherent whole.  It is
not possible, for example, to
to change the \code{x}
and \code{y} data used in the plot and have the points and axes update
automatically.  There is no scatterplot object to save;  the individual
components exist, but they are not bound together as a whole.  If/when
these sorts of issues become important, it becomes necessary to
create a \grid{} graphical object (a \grob{}) to represent the plot.

The first step is to write a function which will create a \grob{}
-- a {\it constructor} function.  In most cases, this will involve
creating a special sort of \grob{} called a \gTree{};  this is just
a \grob{} that can have other \grob{}s as children.  Here's an example
for creating an \code{splot} \grob{}.  I have put bits of the
construction into separate functions, for reasons which will become
apparent later.

<<>>=
splot.data.vp <- function(x, y) {
  viewport(name="dataregion",
                    x=unit(5, "lines"),
                    y=unit(4, "lines"),
                    width=unit(1, "npc") - unit(7, "lines"),
                    height=unit(1, "npc") - unit(7, "lines"),
                    just=c("left", "bottom"),
                    xscale=range(x) + c(-.05, .05)*diff(range(x)),
                    yscale=range(y) + c(-.05, .05)*diff(range(y)))
}

splot.title <- function(title) {
      textGrob(title, name="title",
          y=unit(1, "npc") + unit(1.5, "lines"),
          gp=gpar(fontsize=16), vp="dataregion")
}

splot <- function(x, y, title, name=NULL, draw=TRUE, gp=gpar(), vp=NULL) {
  spg <- gTree(
    x=x, y=y, title=title,
    name=name,
    childrenvp = splot.data.vp(x, y),
    children=gList(rectGrob(name="border", vp="dataregion"),
      xaxisGrob(name="xaxis", vp="dataregion"),
      yaxisGrob(name="yaxis", vp="dataregion"),
      pointsGrob(x, y, name="points", vp="dataregion"),
      textGrob("x axis", y=unit(-3, "lines"), name="xlab",
          gp=gpar(fontsize=14), vp="dataregion"),
      textGrob("y axis", x=unit(-4, "lines"), name="ylab",
          gp=gpar(fontsize=14), rot=90, vp="dataregion"),
      splot.title(title)),
    gp=gp, vp=vp,
    cl="splot")
  if (draw)
    grid.draw(spg)
  spg
}

@

There are four important additions to the argument list compared
to the original \code{splot()} function:
\begin{enumerate}
\item The \code{name} argument allows a string identifier to be
associated with the scatterplot object we create.  This is
important for being able to specify the scatterplot when we try to edit
it after drawing it and/or when it is part of a larger \grob{}
(see later examples).
\item The \code{draw} argument
 makes it possible to use the function in a procedural
manner as before:

<<splotgrob, eval=FALSE, echo=FALSE>>=
sg <- splot(1:10, 1:10, "Same as Before", name="splot", draw=FALSE)

<<>>=
splot(1:10, 1:10, "Same as Before", name="splot")
downViewport("dataregion")
grid.text(date(), x=unit(1, "npc"),
  y = 0, just=c("right", "bottom"), gp=gpar(col="grey"))
upViewport(0)

@
\item The \code{gp} argument allows the user to supply \code{gpar()}
settings for the scatterplot as a whole.
\item The \code{vp} argument allows the user to supply a viewport
for the \code{splot} \grob{} to be drawn in.  This is especially
useful for specifying a \code{vpPath} when the \code{splot} is
used as a component of another \grob{} (see scatterplot matrix
example
below).
\end{enumerate}

The important parts of the \gTree{}  definition are:
\begin{enumerate}
\item The \code{children} argument provides a list of \grob{}s which are
part of the scatterplot.  When the scatterplot is drawn, all children
will be drawn.  Notice that instead of the procedural \code{grid.*()} functions
we use \code{*Grob()} functions which just produce \grob{}s and do not
perform any drawing.  Also notice that I have given each of the
children a name;  this will make it possible to access the components
of the scatterplot (see later examples).
\item The \code{childrenvp} argument provides a viewport (or
\code{vpStack}, \code{vpList}, or \code{vpTree}) which will be pushed before
the children are drawn.  The difference between this argument and the
\code{vp} argument common to all \grob{}s is that the \code{vp} is pushed before
drawing the children and then popped after, whereas the \code{childrenvp}
gets pushed {\it and} then a call to \code{upViewport()} is made
before the children are drawn.  This allows the children to simply
specify the viewport they should be drawn in by way of a \code{vpPath}
in their \code{vp} argument.  In this way, viewports remain available
for further annotation such as we have already seen in procedural code.
\item The \code{gp} and \code{vp} arguments are automatically handled by
the \gTree{} drawing methods so that \code{gpar()} settings will be
enforced and the viewport will be pushed when the \code{splot} is drawn.
\item The \code{cl} argument means that the \grob{} created is a special
sort of \grob{} called \code{splot}.  This will allow us to write methods
specifically for our scatterplot (see later examples).
\end{enumerate}


@

Now that we have a \grob{}, there are some more interesting things that
we can do with it.
First of all, the \code{splot} \grob{} provides a container for the
\grob{}s which make up the scatterplot.
If we modify the \code{splot} \grob{}, it affects all of the children.

<<results=hide>>=
splot(1:10, 1:10, "Same as Before", name="splot")
grid.edit("splot", gp=gpar(cex=0.5))

<<fig=TRUE, echo=FALSE, results=hide>>=
<<splotgrob>>
sg <- editGrob(sg, gp=gpar(cex=0.5))
grid.draw(sg)

@

We can access elements of the \code{splot} \grob{} to edit them
individually.

<<results=hide>>=
splot(1:10, 1:10, "Same as Before", name="splot")
grid.edit(gPath("splot", "points"), gp=gpar(col=1:10))

<<fig=TRUE, echo=FALSE, results=hide>>=
<<splotgrob>>
sg <- editGrob(sg, gPath="points", gp=gpar(col=1:10))
grid.draw(sg)

@

With a little more work we can make the scatterplot a bit more dynamic.
The following describes a \code{editDetails()} method for the
\code{splot} \grob{}.  This will be called whenever a scatterplot
is edited and will update the components of the scatterplot.

<<>>=
editDetails.splot <- function(x, specs) {
  if (any(c("x", "y") %in% names(specs))) {
    if (is.null(specs$x))
      xx <- x$x
    else
      xx <- specs$x
    if (is.null(specs$y))
      yy <- x$y
    else
      yy <- specs$y
    x$childrenvp <- splot.data.vp(xx, yy)
    x <- addGrob(x, pointsGrob(xx, yy, name="points", vp="dataregion"))
  }
  x
}
splot(1:10, 1:10, "Same as Before", name="splot")
grid.edit("splot", x=1:100, y=(1:100)^2)

<<fig=TRUE, echo=FALSE, results=hide>>=
<<splotgrob>>
sg <- editGrob(sg, x=1:100, y=(1:100)^2)
grid.draw(sg)

@

The \code{splot} \grob{} can also be used in the construction of
other \grob{}s.  Here's a simple scatterplot matrix
\grob{}\footnote{{\bf Warning:}  As the number of \grob{}s in a \gTree{} gets
larger the construction of the \gTree{} will get slow.  If this happens,
the best solution is to just use a \grid{} function rather than
a \gTree{}, and wait for me to implement some ideas for speeding things
up!}.

<<fig=TRUE>>=
cellname <- function(i, j) {
  paste("cell", i, j, sep="")
}

splom.vpTree <- function(n) {
  vplist <- vector("list", n^2)
  for (i in 1:n)
    for (j in 1:n)
      vplist[[(i - 1)*n + j]] <-
        viewport(layout.pos.row=i, layout.pos.col=j,
          name=cellname(i, j))
  vpTree(viewport(layout=grid.layout(n, n), name="cellgrid"),
    do.call("vpList", vplist))
}

cellpath <- function(i, j) {
  vpPath("cellgrid", cellname(i, j))
}

splom <- function(df, name=NULL, draw=TRUE) {
  n <- dim(df)[2]
  glist <- vector("list", n*n)
  for (i in 1:n)
    for (j in 1:n)
      if (i == j)
        glist[[(i - 1)*n + j]] <- textGrob(paste("diag", i, sep=""),
          gp=gpar(col="grey"), vp=cellpath(i, j))
      else if (j > i)
        glist[[(i - 1)*n + j]] <- textGrob(cellname(i, j),
          name=cellname(i, j),
          gp=gpar(col="grey"), vp=cellpath(i, j))
      else
        glist[[(i - 1)*n + j]] <- splot(df[,j], df[,i], "",
          name=paste("plot", i, j, sep=""), vp=cellpath(i, j),
          gp=gpar(cex=0.5), draw=FALSE)
  smg <- gTree(name=name,
    childrenvp=splom.vpTree(n),
    children=do.call("gList", glist))
  if (draw)
    grid.draw(smg)
  smg
}

df <- data.frame(x=rnorm(10), y=rnorm(10), z=rnorm(10))
splom(df)

@

This \grob{} can be edited as usual:

<<>>=
splom(df)
grid.edit("plot21::xlab", label="", redraw=FALSE)
grid.edit("plot32::ylab", label="", redraw=FALSE)
grid.edit("plot21::xaxis", label=FALSE, redraw=FALSE)
grid.edit("plot32::yaxis", label=FALSE)

<<splomgrob, eval=FALSE, echo=FALSE>>=
smg <- splom(df, draw=FALSE)

<<fig=TRUE, echo=FALSE, results=hide>>=
<<splomgrob>>
smg <- editGrob(smg, gPath="plot21::xaxis", label=FALSE)
smg <- editGrob(smg, gPath="plot21::xlab", label="")
smg <- editGrob(smg, gPath="plot32::yaxis", label=FALSE)
smg <- editGrob(smg, gPath="plot32::ylab", label="")
grid.draw(smg)

@

But of more interest, because this is a \grob{}, is the {\it programmatic}
interface.  With a \grob{} (as opposed to a function) it is possible
to modify the description of what is being drawn via an API (as opposed
to having to edit the original code).  In  the following, we
remove one of the ``spare'' cell labels and put in its place the
current date.

<<>>=
splom(df, name="splom")
grid.remove("cell12")
grid.add("splom", textGrob(date(), name="date", gp=gpar(fontface="italic"),
  vp="cellgrid::cell12"))

<<fig=TRUE, echo=FALSE, results=hide>>=
<<splomgrob>>
smg <- removeGrob(smg, "cell12")
smg <- addGrob(smg, textGrob(date(), name="date", gp=gpar(fontface="italic"),
  vp="cellgrid::cell12"))
grid.draw(smg)

@

With the date added as a component of the scatterplot matrix, it is
saved as part of the matrix.  The next sequence saves the scatterplot
matrix, loads it again, extracts the bottom-left plot and the date
and just draws those two objects together.

<<>>=
splom(df, name="splom")
grid.remove("cell12")
grid.add("splom", textGrob(date(), name="date", gp=gpar(fontface="italic"),
  vp="cellgrid::cell12"))
smg <- grid.get("splom")
save(smg, file="splom.RData")
load("splom.RData")
plot <- getGrob(smg, "plot31")
date <- getGrob(smg, "date")
plot <- editGrob(plot, vp=NULL, gp=gpar(cex=1))
date <- editGrob(date, y=unit(1, "npc") - unit(1, "lines"), vp=NULL)
grid.newpage()
grid.draw(plot)
grid.draw(date)

<<fig=TRUE, echo=FALSE, results=hide>>=
<<splomgrob>>
smg <- removeGrob(smg, "cell12")
smg <- addGrob(smg, textGrob(date(), name="date", gp=gpar(fontface="italic"),
  vp="cellgrid::cell12"))
save(smg, file="splom.RData")
load("splom.RData")
plot <- getGrob(smg, "plot31")
date <- getGrob(smg, "date")
plot <- editGrob(plot, vp=NULL, gp=gpar(cex=1))
date <- editGrob(date, y=unit(1, "npc") - unit(1, "lines"), vp=NULL)
grid.draw(plot)
grid.draw(date)

@

All of this may seem a bit irrelevant to interactive use, but it
does provide a basis for creating an editable plot interface like
M Kondrin's \pkg{Rgrace} package.

% Start a new page
% Not echoed, not evaluated
% ONLY here for checkVignettes so that all output doesn't
% end up on one enormous page

<<eval=FALSE, echo=FALSE>>=
grid.newpage()

@
\end{document}