search

pyronear / pyro-mlops

3 stars 4 forks source link

Taking into account the CO2 emissions of models training #9

Open fe51 opened 1 year ago

fe51 commented 1 year ago

Hi there,

Given that the Pyronear project has a low-resource approach, as we try to minimise the environmental impact of our products, it seems relevant to have an idea of the CO2 emissions associated with our training sessions.

There are various solutions for this, including codecarbon developed by some D4G volunteers.

Consequently, it would be interesting to include some related metrics in our ml ops workflow.

In this way, we could compare the performance of the training sessions in terms of estimated CO2 emissions and sum it to have estimate CO2 emission of Pyronear model training

What do you think ?

Happy to discuss it :)

gaetanbrison commented 1 year ago

Hi @fe51,

I think it is a great idea:) I have used codecarbon before. I think it would be nice as a 1st step to monitor code execution thanks to codecarbon and / or lineprofiler. In a 2nd step improve run time and co2 emissions. It could even be part of the ML Flow process. What do you think about it @MateoLostanlen ?

MateoLostanlen commented 1 year ago

@gaetanbrison I love the idea !

frgfm commented 1 year ago

We already have it on pyrovision actually :) https://github.com/pyronear/pyro-vision/pull/140