This repository contains code for "Learning to Demodulate from Few Pilots via Offline and Online Meta-Learning" - Sangwoo Park, Hyeryung Jang, Osvaldo Simeone, and Joonhyuk Kang (fully revised and to appear in IEEE Transactions on Signal Processing, arXiv post updated on 4 Dec. 2020)
The code has been updated in order to capture more interesting aspects of meta-learning usages for demodulation (e.g., I/Q imbalance, comparison with conventional communication scheme: MMSE channel estimator + Maximum Likelihood demodulator). Essential part of meta-learning schemes has not been changed. Current version contains full code which may contain deprecated modules for final experiments (i.e., figures for the paper). Most essential experimental set-up has been organized in the paper in detail but if any ambiguity raises, please feel free to contact the authors (e.g., sangwoo.park@kcl.ac.uk).
Run files and data sets for better reproducibility have been added. Minor bugs are fixed.
This program is written in python 3.7 and uses PyTorch 1.1.0 and scipy. Tensorboard for pytorch is used for visualization (e.g., https://pytorch.org/docs/stable/tensorboard.html).
- pip install tb-nightly, pip install future, and pip install scipy might be useful.
-
Train model:
To train the demodulator with default settings, execute
python main_offline.pyFor the default settings and other argument options, please see top of
main_offline.py. -
Test model:
To test the demodulator with default settings, execute
python main_offline.py --path_for_meta_trained_net </path/to/saved-meta-model> --path_for_bm2_net </path/to/saved-joint-trained-model> -
In 'run' folder, basic examples (including experiments for the figures in the paper) can be found.
In 'generated_data' folder, meta-training and meta-test set for the realistic scenario can be found.
-
Run model:
To train and test the demodulator in an online manner with default settings, execute
python main_online.pyFor the default settings and other argument options, see top of
main_online.pyIn 'run' folder, basic examples (including experiments for the figures in the paper) can be found.