PyTorch eco-design rules


After participating to a hackathon about developing an eco-design plugin for SonarQube, I wanted to apply the same concept for deep learning (PyTorch first, since it’s the main framework I use) by listing eco-design rules that can be detected using Python AST (aka. static parsing). The idea is to provide a list of rules that can reduce Co2 emission.

I already implemented 6 basic rules here, for example avoid using bias in convolutional layers before batch norms since it wastes memory and computational power.

TLDR; If people have ideas about eco-design rules in Pytorch, feel free to reach me, respond to this thread, or send a carrier pigeon :slight_smile:

1 Like