A retrospect on 2021 and looking ahead to 2022

An overview of the new version of TensorLy and the many new features and improvements it brings.

In version 0.2.0, TensorLy was refactored to support backends. As proof of concept I put together a PyTorch backend. It makes it trivial to combine pytorch code with tensor methods. In this post we demonstrate this by performing Tucker tensor decomposition using autograd and gradient descent.

From version 0.2.0, TensorLy has an MXNet Backend, in addition to the NumPy backend. This allows to perform tensor operations on multi-machines, on CPU and GPU seemlessly, as well as to integrate with Deep Neural Networks. This posts goes over this new version and how to install it.

A look at tensor unfolding and its different definitions. We go through their mathematical properties, and python implementation with NumPy.

A demonstration on how to use TensorLy to perform robust tensor Principal Component Analysis, with an application on images and videos.