This repository contains the code needed to reproduce results in the paper by M. Belgiovine, et al. “Deep Learning at the Edge for Channel Estimation in Beyond-5G Massive MIMO”, accepted at IEEE Wireless Communications Magazine (WCM), April 2021. The manuscript is available on IEEE Xplore.
Please cite our work if using this code for your experiments:
@article{Belgiovine2021DeepLA,
title={Deep Learning at the Edge for Channel Estimation in Beyond-5G Massive MIMO},
author={M. Belgiovine and Kunal Sankhe and Carlos Bocanegra and D. Roy and K. Chowdhury},
journal={IEEE Wireless Communications},
year={2021},
volume={28},
pages={19-25}
}
The Training and Testing dataset to reproduce the paper results can be downloaded on Genesys Website. The web page also offer additional insights on the dataset created and the obtained performance.
A walkthrough video explaining the dataset generation and how to run the code is available here.
Data generation requires a lot of resources, a machine with multicore and lots of RAM is recommended. Also, for TensorFlow, a GPU is recommended. Software requirements:
- Python 3.7
- TensorFlow (2.3)
- Matlab (2020b)
- Matlab Phased Antenna Array Toolbox and Hybrid Beamforming Example installed on your machine
First configure simulation parameters in setenv.sh. Here is an example of a simulation configuration:
#!/bin/bash
# main paths
PY=~/anaconda3/envs/py3_tf2.3/bin/python
MATLAB=matlab20
MAT_CODEDIR=packet_generation/phased_arr
MODEL_DIR=magazine_review/BS32_denoise_3k_SNR120
PYDATASET_DIR=datasets_maMIMO
MMIMO_BF_EX_DIR=~/Documents/MATLAB/Examples/R2020a/phased_comm/MassiveMIMOHybridBeamformingExample
# flag to generate or not data
IS_GEN_DATA=0
# simulation parameters
Nt=32
Nr=4
TRAIN_Npkt=3000 # tot. num of transmissions in training set
TEST_Npkt=500 # for each SNR level
SNRLev="-25 -20 -15 -10 -5 0 5 10" # SNR levels considered
Is it possible to define different SNR levels to be evaluated, number of transmitter and receiver antennas, number of training and test samples. Note that increasing the values of simulation parameters will requires more resources.
Run the script full_pipeline_maMIMO_DNNEst.sh to:
- generate the Training and Testing datasets through Matlab
- train the DNN models for channel estimation
- test the trained models on test data
- generate output performance
To run the script, type the following in your terminal from the project root directory:
bash full_pipeline_maMIMO_DNNEst.sh
If you have already downloaded the dataset, you can skip the data generation portion and directly train the DNN models. The provided setenv.sh has a specific flag for this purpose:
...
# flag to generate or not data
IS_GEN_DATA=0
...In case you want to re-generate the data or generate a new dataset with different parameters, you can set this flag to 1. Also, make sure that the PYDATASET_DIR is pointing to the folder where the dataset has been downloaded.
Please reach out to belgiovine.m@northeastern.edu for any question. Disclaimer: this code was part of a larger code base that involved multiple experiments. In the future it will be polished to trim the unnecessary parts of code.