DL_mini_projects

This repository contains two mini-projects for the EPFL deep learning course (EE-559), Spring 2021.

The directories named Proj1 and Proj2 contain the python sources files including the main executable test.py to call without arguments as well as the report corresponding respectively to the first and second project described below.

Project 1: Digit Image classification task with weight sharing and auxiliary losses

The first project (the report can be found here) aims at testing different architectures tocompare two digits visible in a two-channel image. Moreprecisely the purpose is to asses the performance improve-ment that can be achieved with the use of weight sharing orthe use of an auxiliary losses. The goal is to get a series of 2×14×14 tensors, correspondingto pairs of 14 × 14 gray-scale images and with the deepnetworks that we will implement, predict for each pair if first digit is lesser or equal to the second.

Project 2: Mini "Deep Learning Framework"

The second project (the report can be found here) consistend on implementing a DL framework from scratchì using only Pytorch’s tensor operations and the standard math library hence in particular not being allowed to leverage autograd or the neural-network modules.

Group Members

• Germini Lorenzo
• Chrysanthou Melina

Professors

• François Fleuret