signn: Signal Detection and Classification using Neural Networks

Disclaimer

This is a forked repository from https://gitlab.com/librespacefoundation/sdrmakerspace/signn, commit SHA hash acb41e9d.

Requirements

The requirements are listed in the requirements3.txt located in the root directory of this project.

• Tensorflow [2.0.0]
• Keras [2.2.4]
• sklearn
• numpy [1.16.2]
• matplotlib [3.1.0]
• h5py [2.9.0]
• keras-tuner [1.0.0]

In addition, for running some auxilliary scripts some extra dependencies should be met:

• GNURadio [3.7+]

Download & Install

$ git clone https://github.com/srhee0187/signn.git
$ cd signn
$ python3 -m pip install --user -r requirements3.txt
$ python -m pip install --user h5py==2.9.0 numpy==1.16.2

In order to install GNURadio 3.7 please check for the appropriate developement package for your distribution.

Usage

1. Generate the signal dataset:

First download and extract the source material need for the generation of signals:

$ cd utils/dataset/gnuradio_sim
$ wget https://cloud.libre.space/s/rzS3QaXLY6BTN3x/download/source_material.tar.gz
$ tar xvzf source_material.tar.gz

Then, run the dataset generation script:

$ python2 generate_sim_dataset.py

Note: The generate_sim_dataset.py requires python2.7 as it depends on GNURadio 3.7+.

This command will create the SIGNN_2019_01.hdf5 dataset. This step can be skipped in order to use a different signal dataset, as long as it follows the specific scheme described here. Depending on the processing power of your machine, this step can take as long as several hours.

If you desire to skip the data generation step, then you simply replace the outputs: these are the classes.txt and SIGNN_2019_01.hdf5 files used in the rest of this tutorial. Copy it to the utils/dataset/gnuradio_sim folder.

2. Generate the Keras model used for training:

$ python3 signn_modeler.py -i 2 1024 -s model.h5

For more information, please use the help argument:

$ python3 signn_modeler.py --help

3. Train the Keras model exported with the previous command:

$ python3 signn_trainer.py -p utils/dataset/gnuradio_sim -d SIGNN_2019_01_1024.hdf5 -m model.h5 --train --dataset-shape 2 1024 --snr 28 &

For more information, please use the help argument:

$ python3 signn_trainer.py --help

4. Monitor the training progress using Tensorboard:

$ tensorboard --logdir logs/plots

Then navigate to http://0.0.0.0:6006

5. Tune Keras model hyperparameters by using keras-tuner library.

$ python3 signn_tuner.py -p utils/dataset/gnuradio_sim -d SIGNN_2019_01.hdf5 --dataset-shape 2 1024 -s artifacts --test

For more information, please use the help argument:

$ python3 signn_tuner.py --help