An investigation of sparse tensor formats for tensor libraries

Abstract

Tensors provide a generalized structure to store arbitrary indexable data, which is applicable in fields such as chemometrics, physics simulations, signal processing and lies at the heart of machine learning. Many naturally occurring tensors are considered sparse as they contain mostly zero values. As with sparse matrices, various techniques can be employed to more efficiently store and compute on these sparse tensors. This work explores several sparse tensor formats while ultimately evaluating two implementations; one based on explicitly storing coordinates and one that compresses these coordinates. The two formats, Coordinate and CSF2, were evaluated by comparing their execution time of tensor-matrix products and the MTTKRP operation on several datasets. We find that the Coordinate format is superior for uniformly distributed sparse tensors or when used in computation that emits a sparse tensor via a mode dependent operation. In all other considered cases for large sparse tensors, the storage savings of the compressed format provide the best results.