TARIEL, M Clément (2022) PRE - Research Project, ENSTA.

[img]
Preview
PDF
417Kb

Abstract

This document is about presenting tools to help establishing the ecological impact of design decisions in deep learning development, such as the choice of a particular framework. In this paper are presented some tools to make energy profiles of deep learning models, some criteria to characterize them and finally a comparison of these tools based on these criteria. It is mainly focused on image recognition Convolutional Neural Networks built with PyTorch and TensorFlow, the goal being to sort them by eco-friendliness. With this work, a prototype of Application Programming Interface has been developed to simplify energy measurements on PyTorch, TensorFlow and ONNX models. Another prototype of Application Programming Interface has been developed to allow users to use easily some of the tools presented in this document from Python scripts.

Item Type:Thesis (PRE - Research Project)
Uncontrolled Keywords:Deep learning, Energy consumption, Energy profile, Carbon footprint, Ecology, Python
Subjects:Information and Communication Sciences and Technologies
ID Code:9065
Deposited By:Clement Tariel
Deposited On:12 juin 2023 11:33
Dernière modification:12 juin 2023 11:33

Repository Staff Only: item control page