This repository contains a collection of machine learning practical projects completed during a Deep Learning course.
The goal of these projects was to implement neural network architectures from scratch to understand their internal mechanics rather than relying on high-level frameworks.
It features implementations of neural networks, convolutional models, VGG-like architectures, and U-Net for image segmentation.
Note: The focus of this repository is on understanding model internals. Most architectures are implemented from scratch without using high-level APIs.
As indicated above, several kind of models were implemented, simple ones like Fully Connected NN and relatively small CNN. There were implemented with TensorFlow and trained on different datasets like MNIST, CIFAR10 or the Iris dataset. U-Net and VGG16 models were implemented in PyTorch regarding of the following architectures and respectively trained on CIFAR100 and CamVid datasets.
| UNet | VGG16 |
|---|---|
 |
 |
.
├── data/
├── images/ # illustration and output images
├── notebooks/ # fcnn and cnn tensorflow implementation
├── unet/ # pytorch unet implementation
├── vgg/ # pytorch vgg implementation
├── README.md
├── exp_unet.ipynb # experimentation notebooks
├── exp_vgg.ipynb # experimentation notebooks
└── utils.py # plot function
- Clone the repository:
git clone https://github.com/maegonz/deep-learning-architectures-from-scratch.git
- Install dependencies:
pip install -r requirements.txt
- Choose a model and choose :
exp_unet.ipynb
exp_vgg.ipynb
These implementations were developed for educational purposes as part of a university course, in order to explore CNN architectures and learn encoder-decoder and skip connections.
| Img Initial | Mask |
|---|---|
 |
 |
| UNet Performance | VGG Performance |
|---|---|
 |
Project created by Antony Manuel, as part of the FDAA course, under Pr Sebastien Ambellouis supervision.
IMT Nord Europe — 2025–2026