AirPack is an introductory framework for creating deep learning models to classify radio frequency (RF) signals and deploying the model to Deepwave Digital’s Artificial Intelligence Radio Transceiver (AIR-T). It provides tools for training a neural network on RF signal data (including source code for a convolutional neural network classifier (CNN) model), tools for developing your own models, bundled training datasets, and examples of deploying a trained model on the AIR-T for inference. By using AirPack, the engineering development and network training schedules can be shortened, leading to a reduction in labor costs and faster time to market.
The AirPack tools are accessible through a Python API
that allows the user to call the functions as needed, either from one of our provided
examples in the
airpack_scripts package, or by directly importing the
airpack package from a Python console or Jupyter notebook. In order to
simplify the hassle of installing all the complicated drivers, files, and
toolboxes to build a machine learning environment, AirPack also provides a custom
Docker container for use on a training PC.
It’s easy to get started: after running one of the provided training examples using the bundled dataset, the software will produce an RF signal classifier neural network that is deployable on the AIR-T software defined radio. In order to train the model within a reasonable period of time, you’ll need to install AirPack’s Docker container for training on a separate PC with a NVidia GPU.
After you are up and running, you can tweak parameters of the existing classifier models, re-train them using your own signal data, or even modify the structure of the classifier to suit your needs.
The following instructions walk though the process of installing the AirPack software on the training computer with a Docker container, training a CNN within the Docker container, installing the trained model on the AIR-T, and then performing inference with the trained model on the AIR-T.
You may refer to the Table of Contents menu for instructions on each step below.
Step 1: Install AirPack on the training computer.
Step 2: Train the model. This will produce a
.onnxfile (recommended) or a
.ufffile (legacy) of the trained model. You may also test the trained model on the training computer.
Step 3: Install AirPack on the AIR-T and copy the trained model (
.onnxfile) to the AIR-T.
Step 4: Run inference on the AIR-T using the provided code which will automatically:
.onnxfile using TensorRT into a
Setup the radio.
Receive RF samples from the radio and send them to the DNN for inference.
Print the resulting classification output.
Step 5: Now you can modify and tune AirPack for your specific applications. This may include:
Copyright (C) 2021 Deepwave Digital, Inc. All Rights Reserved.
You can browse AirPack’s documentation by clicking on one of the topics below:
- AirPack Overview
- AirPack Installation
- Training the Model
- Perform Inference on AIR-T
- Working with Data Sets
- Next Steps
- Frequently Asked Questions
- API Reference