Each example has a separate requirements file to avoid unnecessary extra installs for the smaller problems (MNIST,NLP).
The directories mnist nlp and dlrm contain MAP implementations (no sampling for variational inference) of our method and instructions to run the examples.
The directory tensor_layers has the python package for the tensorized layers.
This implementation is a MAP implementation of our paper on Bayesian tensor rank determination for neural networks in which many posterior samples are drawn. However the results are very similar and this is computationally more efficient.
Feel free to contact colepshawkins@gmail.com with any questions, or raise an issue.