If you are reading this on GitHub be sure to check out the full documentation on GitHub pages.

TidyTensor is an R package for inspecting and manipulating tensors (multidimensional arrays). It provides an improved print() for summarizing structure, named tensors, conversion to data frames, and high-level manipulation functions. Designed to complement the keras package for deep learning in R, functionality is layered on top of base R types.

TidyTensor was inspired by a workshop I taught in deep learning with R, and a desire to explain and explore tensors in a more intuitive way.

Installation

A simple devtools::install_github("oneilsh/tidytensor") will do it. If you don’t have devtools, first grab it with install.packages("devtools").

If you use TidyTensor, let us know in an issue!

Examples and Usage

Here we provide just two basic examples of how TidyTensor can help illuminate data used for deep learning. See the Getting Started vignette for more examples of visualizing tensor structure, filtering and data augmentation, producing train/test splits, and other handy features.

Consider the CIFAR10 dataset distributed with the keras library:

## library(keras)
cifar10_raw <- dataset_cifar10()$train$x 

TidyTensor can be used to plot the contained image data with the help of other tidyverse packages:

## library(tidytensor)
## library(ggplot2)
## library(tidyr)

cifar10_raw %>%
  as.tidytensor() %>%
  set_ranknames(image, row, col, channel) %>%
  set_dimnames_for_rank(channel, R, G, B) %>%
  subset(image = 1:4) %>%
  as.data.frame() %>%
  spread(channel, value) %>%
  ggplot() +
    geom_tile(aes(x = col, y = row, fill = rgb(R, G, B, maxColorValue = 255))) +
    facet_wrap( ~ image) +
    coord_equal() +
    scale_y_reverse() +
    scale_fill_identity()

For a second example, we can start with a keras model and generate a function mapping inputs to internal feature map representations:

vgg_model <- application_vgg16(include_top = FALSE, input_shape = c(32, 32, 3))

input <- vgg_model$input
output <- get_layer(vgg_model, name = "block1_conv2")$output

# input shape (N, 32, 32, 3)
# output shape (N, 32, 32, 64) tensor, where last rank are feature maps
compute_featuremaps <- k_function(input, output)

TidyTensor can then be used to feed a given set of input tensors through the function and visualize the resulting convolutional feature maps:

(cifar10_raw / 255) %>%
  as.tidytensor() %>%
  set_ranknames(image, row, col, channel) %>%
  set_dimnames_for_rank(channel, R, G, B) %>%
  subset(image = 1:4) %>%
  compute_featuremaps() %>% 
  as.tidytensor() %>%
  set_ranknames(image, row, col, featuremap) %>%
  subset(featuremap = 1:6) %>%
  as.data.frame() %>%
  ggplot() +
    geom_tile(aes(x = col, y = row, fill = value)) +
    facet_grid(image ~ featuremap) +
    coord_equal() +
    scale_y_reverse() 

Contributing

Pull requests welcome! Please see the CONTRIBUTING.md file for details. We are especially interested in additional methods for visualizing or summarizing the structure and content of TidyTensors.

Changelog

• v1.0.0: Minor documentation improvements, version 1.0 minted for JOSS publication
• v0.9.1: preparation for JOSS submission, many documentation improvements, removing allow_huge option from as.data.frame.tidytensor()
• v0.9.0: refactor bottom paramter of print() to base_rank
• v0.8.2: minor bugfixes, new combine_ranks() function
• v0.8.1: add [] and [] <- functionality
• v0.8: first version on github