24 October 2018

Deploying Neural Networks with

GR-Wavelearnerinfo@deepwavedigital.com

24 October 2018

Outline

• Introduction to Deep Learning in RF•GR-Wavelearner•Application Example

24 October 2018

Deep Learning and Radio Frequency (RF) SystemsDeep Learning is Emerging

• Intrusion Detection• Threat classification• Facial recognition• Imagery analysis

• Tumor Detection• Medical data analysis• Diagnosis• Drug discovery

• Pedestrian / obstacle detection

• Navigation• Street sign reading• Speech recognition

• Image classification• Speech recognition• Language translation• Document / database

searching

Cyber Medicine Autonomy Internet

Deep learning technology enabled and accelerated by GPU processors - Has yet to impact design and applications in wireless and radio frequency systems

Gam

ma

X-Ray UV

Visi

ble

Infrared Radio Frequencies

1 pm

10 p

m

10 n

m

400

nm

700

nm

1 m

m

100

km

RadarSatellite

CommunicationsElectronic Warfare

Telecommunications

MilitaryCommunications

Navigation

UAV Wireless ControlWireless Networking

Internet of ThingsRF Ablation(Medical)

Radio Frequency Technology is Pervasive

Enabled by low-cost, highly capable general purpose graphics processing units (GPUs)

3

24 October 2018

• Spectrum monitoring (threats)• Intelligent spectrum usage• Electronic protection (anti-jam)• Cognitive system control

Spectrum / Network Centric Applications

• Advanced modulation techniques• Adaptive waveforms• Encryption and security

Device / BasestationCentric Applications

• Voice / image recognition• Multi-sensor fusion• Decision making and data reduction

User ApCentric Applications

Where to Use Deep Learning in RF Systems

4

Modulate1 0 1 1 0 0 1

User Aps

Transmit Frequency

Convert

User Aps

Receive Frequency

Convert

1 0 1 1 0 0 1Demodulate

24 October 2018

Inference at the Edge with GR-Wavelearner

5

Deploy Application

TensorRTOptimizer Runtime

Optimize Neural NetworkTrain Neural Network

24 October 2018

Outline

• Introduction to Deep Learning in RF•GR-Wavelearner•Application Example

24 October 2018

Deepwave’s Solution and Platform

7

Artificial Intelligence Radio Transceiver (AIR-T) GR-Wavelearner Software

7

AIR-T Hardware (2x2 MIMO)

Ubuntu Operating System

GNU Radio DSP Software

GR-Wavelearner

Neural Network

RF Signal In

RF Signal Out

Data I/O

Seamless Deployment of Deep Learning in RF Applications

The AIR-T arrives pre-installed with all software tools necessary for deployment

24 October 2018

Neural Network Deployment in Three Steps

• Step 1 – Train• Using TensorFlow, MATLAB, Keras, PyTorch, etc.

• Step 2 – Optimize• Using NVIDIA’s TensorRT

• Step 3 – Deploy • Using Deepwave’s GR-Wavelearner and GNU Radio

8

TrainedNeural

Network

TensorRTOptimizer

RuntimePLAN File

UntrainedNeural

Network

Training TrainedNeural

Network

RuntimePLAN File

WavelearnerSDK

SoftwareDefinedRadio

24 October 2018

GR-Wavelearner Software

9

• Goal is to help the open source community easily deploy deep learning within signal processing applications

• Well documented README with dependency installation instructions to get started quickly

• Ubuntu 16.04 recommended, Windows 10 supported• NVIDA Docker Container 18.08*

• Signal classifier example provided:• GNU Radio Flowgraph• Python source code• PLAN files that are executable on the AIR-T and Maxwell• Signal data file example for testing

• Support for TensorRT 5.0• Available at: deepwavedigital.com/wavelearner

https://docs.nvidia.com/deeplearning/sdk/tensorrt-container-release-notes/rel_18.08.html

24 October 2018

Outline

• Introduction to Deep Learning in RF•GR-Wavelearner•Application Example

24 October 2018

Deepwave Workflow Example

11

For complete details on this deep learning example, see the Deepwave talk at GTC 2018:Video: http://on-demand.gputechconf.com/gtcdc/2018/video/DC8103/video/video.mp4Slides: http://www.deepwavedigital.net/slides/GTC_Washington_DC_October_2018.pdf

24 October 2018

Step 1: Train Neural Network

• Workflow utilizes TensorRT for deployment• Allows for training on wide array of deep learning

frameworks

12

Train

Native support for TensorRT TensorRT Support Via ONNX**

24 October 2018

Step 1: Train Neural Network

13

Signal Feature Extraction Signal Classification

Max

Poo

l

Flat

ten

Baseband Signal In

Conv

olut

ion

QI

•••

QIQIQISignal Classification Out

Model Definition

24 October 2018

• 1000 training segments per SNR• 55 different SNR values

• Training on low SNR values increase detection sensitivity

• 100% accuracy not expected due to training at extremely low SNR values

• Softmax cross entropy• Adam Optimizer

14

Deep Learning Classifier Training

Step 1: Train Neural NetworkTraining Process

24 October 2018

Step 1: Train Neural Network

15

Prediction

Trut

h

SNR (dB)

Prob

. Of C

orre

ct C

lass

ifica

tion

Model Performance Testing

Confusion MatrixClassification Performance

24 October 2018

Step 2: Optimize Trained Neural Network

16

Optimize

• Optimize the trained neural network using NVIDIA’s TensorRT

• TensorRT produces a PLAN file that is a deployment optimized version of the neural network

24 October 2018

Step 2: Optimize Trained Neural Network

TensorFlow ExampleStep 1: Freeze graph (make variables constants)

Step 2: Convert DNN model to UFF File

Step 3: Convert UFF File to PLAN File

Caveats• Not all layers are supported,

but most common ones are• PLAN file must be created

on deployment architecture• Python conversion not

available on ARM (Jetson)• Limited transferability of

PLAN files

17

* Deep Neural Network

Optimizing a TensorFlow Graph

24 October 2018

Step 2: Optimize Trained Neural Network

Convert TensorFlow Graph to UFF Convert UFF to Optimized PLAN File

18

Optimizing a TensorFlow Graph

Example Python Code Provided Executable Binary Provided

24 October 2018

Step 3: Deploy Optimized Neural Network

19

Deploy

* Deep Neural Network

AIR-T Hardware Platform (2x2 MIMO)

Ubuntu Operating System

GNU Radio DSP Software

GR-Wavelearner

RF Signal In

RF Signal Out

Data I/O

PLAN File

24 October 2018

Step 3: Deploy Optimized Neural Network

20

Deploy

* Deep Neural Network

GNU Radio Companion GUI Python API(Real time DNN* RF system in 35 lines of code!)

24 October 2018

GNU Radio – Software Defined Radio (SDR) Framework

• Popular open source software defined radio (SDR) toolkit:

• RF Hardware optional• Can run full software simulations

• Python API• C++ under the hood

• Easily create DSP algorithms • Custom user blocks

• Primarily uses CPU• Advanced parallel instructions• Recent development: RFNoC for FPGA

processing• Deepwave is integrating GPU support

for both DSP and ML

21

24 October 2018

GR-Wavelearner

• Out of tree (OOT) module for GNU Radio

• Allows users to easily incorporate deep learning into signal processing

• C++ and Python API• Open source GPLv3 license• Two blocks currently:

• Inference – TensorRT wrapper for GNU Radio

• Terminal Sink – Python module for displaying classifier output

22

24 October 2018

Artificial Intelligence Radio Transceiver (AIR-T)

23

Adaptive Digital Wireless Transceiver

GPU for Edge computation of AI

*Patent Pending

FPGA for Real Time DSP