pythagore-mod-reco

Modulation recognition AI algorithms benchmark.
This project contains a Jupyter Notebook for the interactive benchmark, deep learning networks and a few utility functions gathered into a package.

Package to run modulation recognition on raw I/Q radio samples

The acompagning paper is: "A light neural network for modulation detection under impairments, T. Courtat, H. du Mas des Bourboux, 2021" presented at the "2021 International Symposium on Networks, Computers and Communications (ISNCC'21)" (http://www.isncc-conf.org/).

The dataset, and a notebook to reproduce the results can be found on Kaggle: https://www.kaggle.com/hdumasde/pythagoremodreco

Example: evolution of the error rate with the number of epochs on the AugMod dataset

Setup

The package requires Python >=3.7.

It has been tested on Nvidia GPU with cuda 10.2.

Install package and dependencies

The package installation is as simple as

cd <project_folder>
pip3 install -U .

If you only need to install dependencies you can go with

python -m pip install -r requirements.txt

Get data

The trainning and testing of algorithms can be performed on several datasets:

The pythagore-mod-reco AugMod dataset as presented in our article:
- can be found on Kaggle: https://www.kaggle.com/hdumasde/pythagoremodreco
- can be downloaded with wget

wget https://augmod.blob.core.windows.net/augmod/augmod.zip
unzip augmod.zip

RML 2016 datasets from DeepSig:

wget https://opendata.deepsig.io/datasets/2016.04/2016.04C.multisnr.tar.bz2?__hstc=233546881.9c91e0549f9b6bfce6708a49c211c1c9.1614872457734.1614872457734.1614872457734.1&__hssc=233546881.1.1614872457735&__hsfp=1843090487
wget https://opendata.deepsig.io/datasets/2016.10/RML2016.10b.tar.bz2?__hstc=233546881.9c91e0549f9b6bfce6708a49c211c1c9.1614872457734.1614872457734.1614872457734.1&__hssc=233546881.1.1614872457735&__hsfp=1843090487
wget https://opendata.deepsig.io/datasets/2016.10/RML2016.10a.tar.bz2?__hstc=233546881.9c91e0549f9b6bfce6708a49c211c1c9.1614872457734.1614872457734.1614872457734.1&__hssc=233546881.1.1614872457735&__hsfp=1843090487

RML 2018 dataset from DeepSig:

To get RadioML2018.01A you should connect to https://www.deepsig.ai/datasets

Please update the following paths: data_path and log_path in jupyter/train-test-modulationreco.ipynb, to where data are located and to where you want to store log files.

Run

Train and test the different networks on each datasets running jupyter/train-test-modulationreco.ipynb.

Source

The implementation of RML-ConvNet is found at https://github.com/radioML/examples/blob/master/modulation_recognition/RML2016.10a_VTCNN2_example.ipynb
The implementation of RML-CNN/VGG is found at https://github.com/leena201818/radioml/blob/master/rmlmodels/VGGLikeModel.py
The implementation of RML-ResNet is found at https://github.com/liuzhejun/ResNet-for-Radio-Recognition/blob/master/ResNet_Model.ipynb

Citing

Please cite the following paper if you are using the AugMod dataset or Mod-LCNN or Mod-LRCNN networks

@INPROCEEDINGS{9615851,
       author = {{Courtat}, Thomas and {du Mas des Bourboux}, H{\'e}lion},
    booktitle = {2021 International Symposium on Networks, Computers and Communications (ISNCC)},
        title = {A light neural network for modulation detection under impairments},
         year = {2021},
       volume = {},
       number = {},
        pages = {1-7},
          doi = {10.1109/ISNCC52172.2021.9615851},
archivePrefix = {arXiv},
       eprint = {2003.12260},
 primaryClass = {cs.LG},
}

Please visit https://www.deepsig.ai/datasets to see how to cite RadioML datasets and networks

License

This repository is licensed under the terms of the MIT License (see the file LICENSE).

eos21 / pythagore-mod-reco