nndiagram_oversize: Producing 'LaTeX' Code for Drawing Over-sized Neural Network Diagrams

Description

The nndiagram_oversize command is used to produce 'LaTeX' code for drawing well-formatted neural network diagrams, some layers of which have excess neurons that users hope to leave out. To make the code work in a 'LaTeX' editor, users need to install and import the 'TeX' package TikZ in the setting of 'TeX' file. Syntax of importing this package is included in the output of function.

Usage

nndiagram_oversize(
  input,
  hidden,
  size.cutoff = 7,
  title = NULL,
  color = "black",
  alpha = 1,
  layer.sep = 2.5,
  layer.label = NULL,
  input.label = NULL,
  output.label = NULL,
  suppress = FALSE
)

Value

nndiagram_oversize uses cat() to print out 'LaTeX' code on console, if not suppressed. Also, nndiagram_oversize saves the same output as a character vector invisibly, so users could use cat() to print it out later at their demand, as shown in Examples. The nndiagram_oversize 'LaTeX' output can be directly copied and pasted to produce neural network diagram in any 'LaTeX' editor.

Arguments

input: a positive integer that specifies the number of input neurons.
hidden: a positive integer vector that specifies the number of neurons on each hidden layer. For example, c(6,6,6) specifies that there are 3 hidden layers and 6 neurons on each hidden layer. Non-positive and non-integer numbers are not allowed.
size.cutoff: an optional numeric value that specifies the cutoff number of neurons. If the number of neurons on a certain layer is greater than this cutoff value, it is considered as an over-sized layer, which will be drawn with some neurons left out. Otherwise, all neurons on the layer will be drawn. In nndiagram_oversize, the cutoff number is defined to be no less than 5. Default is 7.
title: an optional character that specifies the main title of diagram. Default is NULL.
color: an optional character that specifies the color of lines. Default is "black".
alpha: an optional numeric value between 0 and 1 that specifies the opacity of lines. 1 indicates lines to be opaque, and 0 indicates lines to be transparent. Default is 1.
layer.sep: an optional positive numeric value that specifies the distance between layers of a neural network. Default is 2.5.
layer.label: an optional character vector that specifies label for each layer, including input, hidden and output layers.
input.label: an optional character vector that specifies label for each input neuron.
output.label: an optional character that specifies label for output neuron.
suppress: an optional logical value that specifies whether nndiagram_oversize should suppress the output of 'LaTeX' code to be directly printed on console. Default is FALSE.

Author

Chencheng Fang, Bonn Graduate School of Economics, University of Bonn. Email: ccfang@uni-bonn.de

Examples

Run this code

# A neural network with 3 neurons on input layer, 6 neurons on each of 3 hidden layers, and 1 neuron
# on output layer. All other arguments are default, so no layer is considered to be over-sized.
nndiagram_oversize(input=3, hidden=c(6,6,6))

# Same as the first example but cutoff value of size is designed to be 5, so all hidden layers are
# considered to be over-sized.
nndiagram_oversize(input=3, hidden=c(6,6,6), size.cutoff=5)

# Same as the second example but labels of input neurons are designed to be letters from a to c.
nndiagram_oversize(input=3, hidden=c(6,6,6), size.cutoff=5, input.label=letters[1:3])

# Suppress the output of 'LaTeX' code to be directly printed on the console and save the output
# to an object, which can be printed later at demand.
nnd_oversize <- nndiagram_oversize(input=3, hidden=c(6,6,6), size.cutoff=5, suppress=TRUE)
cat(paste(nnd_oversize,"\n"))