Meta-SGLD

This code repo contains the implementation for the Meta-SGLD algorithm presented in the NeurIPS 2021 paper Generalization Bounds For Meta-Learning: An Information-Theoretic Analysis by Chen, Shui, and Marchand.

Prerequisites

To run the code, make sure that the requirements are installed.

pip install -r requirements.txt

The code was written to run in Python 3.6 or in a more recent version.

Run experiment on Synthetic data

python src/main/toy_exp/toy.py

Run expierement on Omniglot data

python src/main/omniglot_train.py

The program will download omniglot dataset automatically.

References

The code is modified and implemented based on https://github.com/ron-amit/meta-learning-adjusting-priors2 and https://github.com/dragen1860/MAML-Pytorch.