Make use of DVC's pipeline feature

This would allow us running e.g.

dvc run -d data/train/video -o data/model/out.p creator/train.py

The following steps need to be done first

[ ] Create creator/train.py and copy all training related code from Jupyter notebook, while moving shared code to separate file
[ ] Adjust model and checkpoint export in Jupyter notebook, such that those are always stored in data/train/model

deep-dance / core