This is definitely not production code, but it runs on my machine(tm).
It requires the latest nightly build of Docker for the Pi.
I originally started this project to create a simple way for Pi owners to experiment with TensorFlow.
At the time, neither Docker nor TensorFlow were officially supported on the Pi.
More recently, this project has been dormant, because
- TensorFlow and Docker have officially supported Raspbian for a while, and
- While TensorFlow will run in a Docker container with Jupyter, the limited memory of the Pi 3B+ meant that you couldn't do much with it.
This has changed with the 4 GB model of the Pi 4B.
I'm now attempting an experimental build for Raspbian buster that runs on a Raspberry Pi 4.
I am testing the build using Raspbian/buster and a 16 GB SD card on a Raspberry Pi model 4B with 4 GB of RAM.
The script builds a container based on Katsuya Hyodo's build of TensorFlow 1.14 for the Raspberry Pi, with TensorFlow Lite enabled.
The docker image contains TensorFlow, Jupyter and a TensorFlow notebook copied from the official Google docker TensorFlow build.
The original build relied heavily on the work of resin.io, the Docker team, Sam Abrahams and the Google TensorFlow team. Sources are listed below.
This is not an official TensorFlow port, so don't ask for or expect support from the TensorFlow team.
- Install Docker nighly build on your Raspberry Pi.
curl -fsSL get.docker.com | CHANNEL=nightly shsudo usermod -aG docker pi- log out, then log back in again for the change to take effect
sudo systemctl start docker- Clone this repository into a directory of your choice
git clone https://github.com/romilly/rpi-docker-tensorflow.git- Build the image
cd rpi-docker-tensorflow/build-tensor-pi/docker build -t='yourName/rpi-docker-tensorflow' .
This run instruction expects a directory called myNotebooks within your home directory.
If you save an IPython notebook to the myNotebooks sub-directory
while running your container, it will get saved to the myNotebooks
directory on your Pi.
Notebooks saved to that directory will be persistent - in other words, they will still be there when the container is stopped and restarted.
docker run -it -p 8888:8888 -v ~/myNotebooks:/notebooks/myNotebooks yourName/rpi-docker-tensorflow
You can probably ignore the warnings about the insecurity of the IPython server configuration so long as you do not store any sensitive data or code in your notebooks.
Open a browser on http://raspberrypi:8888 where raspberrypi is the
hostname of the Pi on which the docker image is running, or on
http://localhost:8888 on the Pi itself.
You should see a screen like this:
aaand you're away to the races!
I'll blog more info about this shortly, and will commit the image to docker central once it's working.
In the terminal session where you ran the container, type Ctr-C twice in quick succession. The container will stop.