tntorch: Tensor Network Learning with PyTorch

In many machine learning and data analysis tasks one is faced with multi-dimensional data arrays. Tensors are a powerful tool to represent and handle such data, but often constitute a bottleneck in terms of storage and computation. Tensor decompositions expand a tensor into a set of separable terms. If the tensor has low rank (i.e., there are much fewer degrees of freedom than tensor elements), then such a decomposition can dramatically reduce the representation size (Kolda and Bader, 2009; Cichocki et al., 2016; Khrulkov et al., 2019).