This example demonstrates how to run deploy and run a simple pretrained tensorflow model, in this case the inception V3 model, on a stream of video data from the Raspberry Pi camera to enable realtime object recognition. It makes use of the Movidius Neural Compute Stick to accelerate the recognition.
- Create an account and an Application at resin.io, if you don't know how to do this, just check out the Getting Started Guide
- Add the following to your Fleet Configuration in your newly created Application.
| Configuration Name | Value |
|---|---|
| RESIN_HOST_CONFIG_start_x | 1 |
| RESIN_HOST_CONFIG_gpu_mem | 192 |
| RESIN_HOST_CONFIG_max_usb_current | 1 |
| RESIN_HOST_CONFIG_safe_mode_gpio | 4 |
- Now provision your RPI devices into the Application, making sure to connect up the RPI camera module and the Movidius Compute Stick.
git push resin masterthis repository to the git remote for the application you created.
In order to deploy various models we first need to be able to compile our tensorflow or caffe models into a graph format that the movidius NCS stick understands. To do this we need to use the NCS toolkit (not installed on the RPI because its uncessary bloat), for this we can just pull down and run the full ubuntu image with the NCSDK toolkit installed. In our example below use the ncappzoo repo and mount it into the toolkit repo which will allow us to compile and test all the examples in repo.
git clone -b ncsdk2 https://github.com/movidius/ncappzoo.git
cd ncappzoo
docker run -v `pwd`:/ncappzoo -net=host --privileged -v /dev:/dev --name ncsdk -it shaunmulligan/movidius-x86-toolkit /bin/bash
Once we have the docker containeou need tor running, we can just runs make all at the top of the /ncappzoo folder and it will go away and compile all the models into graphs the movidius is capable of running. NOTE you need to have the movidius usb stick plugged into the machine you are running this container on.
Once all the models are compiled you can just navigate to the ncappzoo directory you cloned and you should be able to grab the various graphs from any of the examples, in the same why we did for the basic inception v3 model.
If you want to check that you Pi is looking at the right things, you can put the application into preview mode by adding an Environment Variable called PREVIEW set with an value. This will cause the application to snap a still photo and the serve that photo on <YOUR_DEVICE_IP>:8080/preview.jpg
To return to recognition mode simply delete the PREVIEW variable on your dashboard.
If you want to rebuild the base image or even just check the things that are included in it, you can find the repo here: https://github.com/resin-io-playground/movidius-rpi3-baseimage . It should be noted that currently this dockerfile needs to be built on an emulated machine or locally on an RPI3. To build it on the resin build servers you will need to do git push resin master:resin-emulated