This is a Pyro implementation of Multi-task Bayesian Neural Networks (BNNs). It allows to train BNNs on a distribution of tasks in order to distill the structure shared by the tasks in a BNN prior. Note that this project is in a very early stage of development.
First install metalearning_benchmarks from here.
Then clone this repository and run
pip install .
from the source directory.
To get familiar with the code, have a look at the example script ./scripts/mtbnn/run_mtbnn.py.
This code is in a very early stage of development and thus not thoroughly tested. Features will change in the future. It was tested only with the packages listed in ./setup.cfg.
This code is licensed under the MIT license and is free to use by anyone without any restrictions.