OOP the hard way: Deep Learning Library

This is the project for the Object Oriented Programming course at the University of Paris-Saclay, M1 GENIOMHE: 2425-m1geniomhe-group-1
Check out this quick demo:

Description

This project is a deep learning library that allows the user to create and train artificial neural networks. The library is implemented in Python and uses the NumPy library for matrix operations. The aim of this project is to perform the task using object-oriented programming principles and design patterns. The user-interface and performance were consequently compared to the PyTorch¹ library. The library relies on gradient calculation and backpropagation through a computational graph. The library supports the following features:

• Fully connected layers
• Activation functions (ReLU, Sigmoid, Softmax)
• Loss functions (MSE, Cross-Entropy, Negative Log-Likelihood)
• Optimizers (SGD)

Note

The implementation of library is based on OOP principles such as encapsulation, inheritance, and abstraction with the use of properties, magic dunder methods and decorators.

For a more detailed description refer to the docs readme.

Installation

To install and test the library, install using pip:

pip install git+https://github.com/deeplearning-oop/2425-m1geniomhe-group-1.git

It will install all dependencies found in requirements.txt, and the library will be available for use through a simple import from the Python environment where it was installed:

import che3le
help(che3le) # this will display a short description of the modules and classes available

The name of the library is che3le (shi‘le), a Lebanese Arabic word meaning "flame"

Project proposal

Take a look at the project description proposed to the students in 2425-project-proposal.pdf.

Report

You can find the description of the deep learning library in the docs folder.

Submission

Submitted a report and a presentation. The code submission is in the src folder. The overall directory structure is as follows (dating 08/01/2025):

2425-m1geniomhe-group-1/
├── 2425-project-proposal.pdf
├── LICENSE.md  # -- MIT License
├── README.md
├── VERSION # -- version file to keep tracks of updates
├── ann/   # -- the python library that is set up by setup.py (pip install)
│   ├── __init__.py
│   ├── extensions/
│   │   ├── __init__.py
│   │   ├── dataloader.py
│   │   ├── dataset.py
│   │   └── transforms.py
│   ├── nn/
│   │   ├── __init__.py
│   │   ├── activation.py
│   │   ├── linear.py
│   │   ├── loss.py
│   │   ├── module.py
│   │   ├── optimizer.py
│   │   └── parameter.py
│   ├── tensor.py
│   └── utils/
│       ├── __init__.py
│       ├── functions.py
│       ├── processing.py
│       ├── simulation.py
│       └── validation.py
├── assets/ # -- images 
│   ├── comparison1.png
│   ├── comparison2.png
│   ├── comparison3.png
│   ├── comparison4.png
│   ├── repo_structure.png
│   └── data_viz.png
├── docs/
│   ├── diagrams/
│   │   ├── class-v1.jpg
│   │   ├── object-v1.jpg
│   │   ├── v0-dark.jpg
│   │   └── v0-light.jpg
│   ├── readme.md
│   └── report.pdf
├── example/
│   └── MNIST_model.py
├── requirements.txt
├── setup.py
├── src/
│   ├── activation.py
│   ├── dataloader.py
│   ├── dataset.py
|   ├── linear.py
│   ├── loss.py
│   ├── module.py
│   ├── optimizer.py
│   ├── parameter.py
│   ├── tensor.py
│   └── transforms.py
└── tests/
|   ├── benchmarks/
|   |   ├── gene_expression.py
|   │   └── MNIST.py
|   ├── unit/  # -- unti tests in classes and comparison with pytorch 
|   |   ├── vision_transforms.ipynb
|   │   └── vision.ipynb
|   ├── mocks/  # -- checking consistency with pytorch 
|   |   ├── simulated_linear.ipynb
|   |   ├── simulated_circular.ipynb
|   |   ├── simulated_checkerboard.ipynb
|   │   └── simulated_spiral.ipynb
|   └── backup/ 
|       ├── test_simulation.py
|       └── test_with_dataloaders.py
└── utils
    ├── changelog.md
    ├── make-global.sh
    ├── readme.tree
    └── update-version.sh

Demo

Kindly find demos in the example folder. If you install the library locally, you can run the following script to train a simple model on the MNIST dataset:

python example/MNIST_model.py

Or directrly from the python notebook, which outputs are saved: example/MNIST_model.ipynb

You can test it as well on Google Colab by clicking on this link and running the code cells.

Authors

• Joelle ASSY (@joelleas)
• Yazid HOBLOS (@yazid-hoblos)
• Rayane ADAM (@raysas)

Footnotes

1. https://pytorch.org/ ↩