How to use TensorFlow Lite on VOXL
We have developed an OpenCL GPU-accelerated TensorFlow Lite for ARM Linux and made it available for VOXL.
A video demonstration of TensorFlow Lite running in real-time on the VOXL platform follows the Table of Contents
Table of contents
hellotflite Example
Running on VOXL
- (PC)
adb shell
to get into voxl - (VOXL) Run:
hello-voxl-tflite-gpu -c "HIRES-CAMERA_ID" -W 1024 -H 768
(e.g. hello-voxl-tflite-gpu -c 0 -W 1024 -H 768)- For help with command line, run:
hello-voxl-tflite-gpu -h
- For help with command line, run:
- (PC) open a new terminal window, and
adb shell
to get into voxl - (VOXL) run
voxl-perfmon
to view GPU usage - (PC) open a third terminal window, and navigate back to the apps-proc-examples/hellotflitegpu/ directory
- (PC) Run:
sudo apt-get install python-qt4
to install dependencies - (PC) Run:
python image_viewer_rgb.py --i IP-ADDRESS-OF-VOXL
- get ip address by running
ifconfig
in a voxl shell
- get ip address by running
Please see example code details here
voxl-mpa-tflite-server Example
The Modal Pipe Architecture allows for modular access and sharing of low-level data around the VOXL platform. The voxl-mpa-tflite-server takes advantage of this architecture to enable a broad set of camera inputs and streaming outputs.
In this example image data from voxl-camera-server is processed by voxl-mpa-tflite-server and then streamed via RTSP using voxl-streamer
Setup
adb shell
bash
vi /etc/modalai/voxl-streamer.conf
- change the “mpa-camera” field under “hires-mpa” from “hires/preview” to “tflite/image/”
Running on VOXL
- (VOXL-1)
adb shell
- (VOXL-1)
bash
- (VOXL-1)
voxl-camera-server -c /etc/modalai/voxl-camera-server.conf
- (VOXL-2)
adb shell
- (VOXL-2)
bash
- (VOXL-2)
voxl-mpa-tflite-server -m mobilenet
- (VOXL-3)
adb shell
- (VOXL-3)
bash
- (VOXL-3)
voxl-streamer
To view the output rtsp stream, open VLC media player. Select media, open network stream, and the URL will be rtsp://YOUR-VOXL-IP-ADDRESS:8900/live