| title | tags | authors | affiliations | date | bibliography | |||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
TidyTensor: Utilities for multidimensional arrays as named hierarchical structures |
|
|
|
13 July 2021 |
paper.bib |
Deep learning applications commonly employ the use of tensors (multidimensional arrays) for data storage and manipulation. In addition to being space efficient tensors provide a common framework for, the engine of deep neural network training [@baydin2018automatic]. TidyTensor is an R package for inspecting and manipulating tensors. It provides an improved print() function for summarizing structure, named tensors, conversion to data frames, and high-level manipulation functions. Designed to complement keras package (or other packages that utilize native R arrays), functionality is layered on top of base R types.
While R supports multidimensional arrays natively, there are important differences between these types and tensors as typically used in deep learning applications. First, R assumes array data are organized in a column-major order as opposed to the row-major order typically used in Python, rendering the default print() and other inspection methods far less useful. TidyTensor additionally provides support for "named" tensors, allowing researchers to use semantically-relevant names for working with and manipulating tensors via a familiar tidy-friendly interface. While other packages exist for named arrays in R, TidyTensor leverages a hierarchical interpretation of tensor data that makes it easy to manipulate and investigate tensor data of various kinds. Early versions of TidyTensor supported a 3-week researcher-oriented workshop "Deep Learning for Life Scientists" at Oregon State University's Center for Genome Research and Biocomputing, and these features supported graduate students and faculty new to the field in applying deep learning techniques to datasets from their own work.
Other packages implement named tensors, though without the "tidy" features of TidyTensor or the simplification of asserting nesting/set-of-sets semantics. In Python, the xarray package provides named tensors and a wide variety of manipulation and related features [@hoyer2017xarray], while in R the garray package provides features similar to TidyTensor [@garray], including implementing names in the same way allowing for easy conversion to and from TidyTensors. The stars R package also provides named arrays with a focus on spatiotemporal data [@stars]. TidyTensor utilizes abind internally for array manipulation [@abind].
TidyTensor was designed primarily as a companion to the keras and tensorflow ports for R which utilize native R types [@rkeras; @rtensorflow]. By contrast, the recently released torch port for R does not use native R types for tensors [@rtorch]; using TidyTensor with this package would require costly conversion to-and-from Torch-native types.