Meta-Autoencoder

This repository contains code for "Meta-Learning to Communicate: Fast End-to-End Training for Fading Channels" - Sangwoo Park, Osvaldo Simeone, and Joonhyuk Kang.

Dependencies

This program is written in python 3.7 and uses PyTorch 1.2.0 and scipy. Tensorboard for pytorch is used for visualization (e.g., https://pytorch.org/docs/stable/tensorboard.html).

• pip install tensorboard and pip install scipy might be useful.

Basic Usage

• Train and test a model:

  To train the autoencoder with default settings, execute

  python main.py
  

  For the default settings and other argument options, see top of main.py

  Once training is done, test will be started automatically based on the trained model.

Toy Example

• In the 'run/toy' folder, meta-learning, joint training, and fixed initialization schemes can be tested based on the pretrained two autoencoder architectures (vanilla autoencoder, autoencoder with RTN)

  In order to train from scratch, remove '--path_for_meta_trained_net ' part.

  In the 'saved_data/toy/nets' folder, trained models used to generate Fig. 2 can be found (proper paths are given in the shell script).

  In order to regenerate Fig. 3, 'toy_visualization.ipynb' may be useful.

A More Realistic Scenario

• In the 'run/realistic' folder, meta-learning, joint training, and fixed initialization schemes can be tested based on the pretrained two autoencoder architectures (vanilla autoencoder, autoencoder with RTN)

  In order to train from scratch, remove '--path_for_meta_trained_net ' part.

  In order to train and/or test with new channels, remove '--path_for_meta_training_channels' and/or '--path_for_test_channels' part.

  In the 'saved_data/realistic/nets' folder, trained models used to generate Fig. 4 can be found (proper paths are given in the shell script).