rdegraafwhizzkit / modelmaker

Training and inferencing with a Google Coral Edge TPU coprocessor

Model maker

Info

• https://coral.ai/docs/accelerator/get-started
• https://coral.ai/docs/edgetpu/compiler/
• https://www.tensorflow.org/lite/models/modify/model_maker/image_classification

Prerequisites / known good configuration

• macOS Monterey 12.7.3
• EdgeTPU runtime 20221024
• libusb 1.0.27
• Docker Desktop 24.0.2
• Python 3.9.6
• pip 24.0
• pip-tools 7.4.1

Prepare environment

System stuff

brew install libusb
curl -LO https://github.com/google-coral/libedgetpu/releases/download/release-grouper/edgetpu_runtime_20221024.zip
unzip edgetpu_runtime_20221024.zip
cd edgetpu_runtime
sudo bash install.sh
rm -rf edgetpu_runtime*

Virtual environment

deactivate > /dev/null 2>&1 || :
rm -rf .venv
/usr/bin/python3 -m venv .venv
. .venv/bin/activate
pip install --upgrade pip==24.0
pip install --upgrade pip-tools==7.4.1
pip-compile
pip install --upgrade -r requirements.txt

Train

mkdir -p models
python train.py

Compile to EdgeTPU

Build Docker image

cd docker
docker build -t edgetpu:latest .
cd -

Start Docker container

docker run --rm --interactive --tty --name Edge-TPU-Compiler --mount type=bind,src=./models,dst=/models edgetpu:latest
cd /models
rm -f model_edgetpu.*
edgetpu_compiler model.tflite
exit

Delete Docker container (if still present)

docker rm $(docker ps -a|grep 'Edge-TPU-Compiler'|awk '{print $1}')

Inference

python inference.py

Nuisances

ZScaler issues:

pip-compile --trusted-host pypi.org --trusted-host google-coral.github.io --trusted-host objects.githubusercontent.com

WSL2 issue:

dockerd --iptables=false