Model-based Deep Learning for Joint RIS Phase Shift Compression and WMMSE Beamforming

Source code for the research paper: Model-based Deep Learning for Joint RIS Phase Shift Compression and WMMSE Beamforming, link to preprint: arxiv.

Required

Run the following commands in the terminal to install Python libraries to the local Python environment.

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install numpy
pip install pandas
pip install tabulate
pip install tqdm
pip install matplotlib

To install MATLAB CVX see https://cvxr.com/cvx/doc/install.html for instructions (not required but may be useful).

Description

Descriptions of the Python projects, directories, and content included in this repository.

MATLAB - system model and RIS beamforming/phaseshift optimization

MATLAB (version 2022a) to generate a dataset for a simulated RIS communication system model.

• generateHDRISData.m
  • run this MATLAB script to generate data for a Half-Duplex RIS model
  • generates the channels / channel errors, solves the beamformer / phase shifts with WMMSE-PI algorithm, then saves the generated data as .csv files located in the datasets/ directory
• src/
  • directory containing MATLAB scripts and functions to generate RIS data
  • to change the parameters of the system model, change the contents of: src/systemModelParameters.m
• datasets/
  • directory to store the generated datasets
  • for Python scripts use the .csv files from this directory to load the datasets into a numpy array

Python - deep learning architecture and benchmarks

Python (version 3.11) to run deep learning experiments.

• mumiso_autoencoder_quantization.py
  • written in a single python file for the convenience of locating all models and relevant source code (i.e. proposed AQE-WMMSE, benchmarks, individual deep learning modules, etc.)
  • the script will not run unless modified appropriately based on how the dataset was generated

Instructions to run project

1. Generate dataset of channels and optimal RIS phase shifts / beamforming
   1. Choose the number of dataset samples by changing the "mc_runs" variable in src/systemModelParameters.m, along with any other system model parameters.
   2. (Optional) It is recommended to generate the dataset samples in parallel with multiple compute nodes / supercomputer to reduce the computation time of large dataset sizes.
   3. run generateHDRISData.m in MATLAB and the data to be used for the deep learning experiments will be generated and saved to datasets/.
2. Run deep learning experiments
   1. In mumiso_autoencoder_quantization.py, set the dataset/results directory and load the number of dataset directories correspondingly with the "num_dirs" variable (i.e. num_dirs > 1 if the data samples were generated from parallel processes).
   2. Run mumiso_autoencoder_quantization.py in Python and the log files of the experiments will be saved to the results directory specified in the script.