GI-PVI

Repository for MPhil Thesis: Global Inducing Point Variational Approximations for Federated BNNs using PVI.

The repository is structured as follows. gi/ contains the implementation of GI-PVI as well as MFVI-PVI. experiments/ contains everything that is related to training the GI-PVI but independent of the method, including datasets (dgp.py), prior specification (priors.py).

Classification experiment

Arguments are specified via command-line. Precision initialization of the weights is done inside classification.py and is currently set to 1e3 - D_in.

• Command: python classification.py --q {GI,MFVI} -d {A,B} --prior {neal,std} --server {SYNC,SEQ} --split {A,B} --lr 0.001 --local_iters 20000 --global_iters 10 --batch 256 --num_clients=1

Regression experiment

• Specify arguments in config/ober.py.
• Command: python pvi_regression.py or python mfvi_regression.py.