A scalable, GPU-accelerated online tensor LDA, built on TensorLy and PyTorch.

As batches of documents arrive online, they are first pre-processed. The resulting document term matrix is centered.

The whitening transformation is updated online and used to whiten \(\matrix{X}\). Finally, the third order moment is updated, directly in factorized form. This learned factorization can be directly unwhiten and uncentered to recover the classic solution to the tensor LDA and recover the topics.