plot.mnist: Nice plot of the MNIST digits with dimensionality reduction
In xrobin/mnist: MNIST dataset

Description Usage Arguments Details Value Author(s) See Also Examples

View source: R/nice_digits.R

Plots MNIST digits (left), projections on a 2-D space (prediction, center) and their reconstructions (right).

plot.mnist(x = mnist$test$x, label = mnist$test$y + 1, model = prcomp(mnist$train$x),
           predictions = predict(model, x)[,1:2], reconstructions = tcrossprod(predictions, model$rotation[,1:2]),
		   show.reconstructions = TRUE, highlight.digits = c(11, 3, 2, 91, 20, 188, 92, 1, 111, 13),
		   digits.col = c("#FF0000FF", "#0000FFFF", "#964B00FF", "#FF00FFFF", "#00AAAAFF", "#00EE00FF", "#000000FF", "#000000FF", "#AAAAAAFF", "#FF9900FF"),
		   digits.alphas.reverse = c(FALSE, TRUE)[c(1, 1, 1, 1, 1, 1, 2, 1, 1, 1)],
		   pch = 21,
		   pch.col = c("black", "white")[c(1, 2, 1, 1, 1, 1, 1, 2, 1, 1)],
		   pch.bg = c("#FF0000FF", "#0000FFFF", "#964B00FF", "#FF00FFFF", "#00AAAAFF", "#00EE00FF", "#FFFFFFFF", "#000000FF", "#AAAAAAFF", "#FF9900FF"),
		   cex = 1.5,
		   cex.axis = .5,
		   cex.lab = .5,
		   cex.highlight = 2.5,
		   xlab = "Node 1",
		   ylab = "Node 2",
		   xlim = NULL,
		   ylim = NULL,
		   ncol = 10,
		   ...)

`x`	A Nx784 matrix of data from which to draw the MNIST digits.
`label`	A vector of length N with labels corresponding to `data`.
`model`	An optional model to compute predictions and reconstructions. If supplied, you probably need to supply `reconstructions` as well. Ignored when both `predictions` and `reconstructions` are provided
`predictions`	A Nx2 matrix with 2 columns containing predictions.
`reconstructions`	A Nx784 matrix with the reconstructions of the digits.
`show.reconstructions`	Whether to plot a column of reconstructions to the right.
`highlight.digits`	The indices of the P digits to draw in the margin, ideally 10 of them.
`digits.col`	A vector of P colors for the `highlight.digits`.
`digits.alphas.reverse`	A logical vector of length P, `TRUE` when the digits has to be plotted white-on-color, `FALSE` when the digit has to be plotted color-on-white.
`pch, pch.col, pch.bg`	Plotting character and associated `col` and `bg` properties for the central plot.
`cex,cex.axis, cex.lab, cex.highlight`	Character expansion factor for the center points, the center plot axis, the center plot labels and the highlight digit points, respectively.
`xlab, ylab`	Labels for the x and y axis, see `par`
`xlim, ylim`	Plot limits, see `par`
`col`	a vector of colors to represent the digit. See `image` for more details.
`ncol`	Number of `layout` columns assigned to the central predictions plot. See “Layout” section below.
`...`	further arguments passed to `image`.

plot.mnist plots model predictions surrounded by a selection of the original digits (x) and, if requested (with show.reconstructions = TRUE), their reconstructions. The highlight.digits argument is an numeric index of which digits to highlight, in order.

Layout

plot.mnist uses a graphical layout. The original digits x will always use exactly one column. If requested, the reconstructions will use a second one. The ncol argument controls how many columns are assigned to the predictions themselves and should be adjusted to the width of the graphical device. For instance if ncol = 1 the projections will use 1/2 or 1/3 of the plot.

With a square device and 10 highlighted digits, ncol = 8 will make the digits and reconstructions square. To have the predictions displayed in a square too, you should open a device with a 12:10 ratio. Graphical devices with a 16:10 ratio should use ncol = 14 to have the predictions slightly stretched. If more or fewer than 10 digits are highlighted, ncol should be adjusted too. See the Examples sections for a demo.

None

Xavier Robin

mnist

data(mnist)

# DeepBeliefNet
# library(DeepLearning)
# plot.mnist(prediction = predict(trained.mnist.dbn, mnist$test$x), reconstruction=reconstruct(trained.mnist.dbn, mnist$test$x))

# With PCA
pca <- prcomp(mnist$train$x)
plot.mnist(prediction = predict(pca, mnist$test$x), reconstruction=tcrossprod(predict(pca, mnist$test$x)[,1:2], pca$rotation[,1:2]))

# Reconstruct with all 784 components of the prediction
plot.mnist(prediction = predict(pca, mnist$test$x), reconstruction=tcrossprod(predict(pca, mnist$test$x), pca$rotation))

# Note that it also works with no arguments - pca is performed
plot.mnist()

## Not run: 
# Change the number of columns. Digits will be square in a square plot device
plot.mnist(ncol = 8)
# Digits will be sqare in a 16:10 plot
plot.mnist(ncol = 14)
# With a 16:9 device we'd need ncol = 15.7778, ncol = 16 is a fair approximation
plot.mnist(ncol = 16)

## End(Not run)

# With a different number of digits (here 3: 0, 1 and 2) one should adjust ncol to something reasonable.
# Note that with ncol = 1, projections will be very stretched. The following code reverses the x and y from PCA
# to have more space for the 1st component.
mnist.02 <- mnist
mnist.02$train$x <- mnist$train$x[mnist$train$y < 3,]
mnist.02$train$y <- mnist$train$y[mnist$train$y < 3]
mnist.02$test$x <- mnist$test$x[mnist$test$y < 3,]
mnist.02$test$y <- mnist$test$y[mnist$test$y < 3]
pca <- prcomp(mnist.02$train$x)
plot.mnist(prediction = predict(pca, mnist.02$test$x)[,2:1], reconstruction=tcrossprod(predict(pca, mnist.02$test$x)[,1:2], pca$rotation[,1:2]),
           x = mnist.02$test$x, label = mnist.02$test$y + 1, highlight.digits = c(3, 2, 1),
           ncol = 1)