kclip / bayesian_active_meta_learning

Code for paper https://arxiv.org/abs/2108.00785

bayesian_active_meta_learning

Scope

Code for paper "Bayesian Active Meta-Learning for Reliable and Efficient AI-Based Demodulation" 2022, https://arxiv.org/abs/2108.00785

Files Breaddown

This repository includes two folders, each runs separately, one for each experiment considered.

1. demodulation/

• main_demod.py: the main file, runs frequentist and Bayesian meta-learning.
• baml4demod.py: auxiliary file.

2. equalization/

• main_eq_mtr.py: the main file for meta-training, save to mat files the learnt model parameters.
• main_eq_mte.py: the main file for meta-testing, loads from file system the model and meta-test. This file can be run even if main_eq_mtr.py did not finish, and will use the hyperparameters learned up till the start of its running.
• baml4eq.py: auxiliary file.

In both folders, an empty sub folder named 'run' for files generated while running to be saved should be made.

Implementation

The meta-learning using Hessian-vector-product is used via pytorch autograd's create_graph=True option.