Most of the repo is a mess. If you're just looking for the final product with (arguably) better quality code that kind of should maybe work, check the inside of pushup_lockscreen folder. The normal flow would be: labeling_tool.py > clean_labeled_data.py > landmark_offline.py > training.py > inference.py. If you really want to get this working, open an issue and I'll clean up the repo :)
HPO was a test to optimize the model using ClearML
notebooks can be safely ignored, the good training code is in pushup_lockscreen
Most of the repo uses ClearML as experiment manager and central output manager. Get a free account on app.clear.ml or host your own server (check their github).
pip install clearml
and
clearml-init
Every script assumes you have the OAK-1 attached to the current machine. I was able to run everything on the same machine and then moved only the inference to the raspberry pi!
To use playaudio, install gstreamer bindings:
sudo apt install python3-gst-1.0