CON(e)VOLUTION - From LeNet to Vision Transformers

Convolutional Neural Networks Architectures for Image Classification Maria Magdalena Balos

This repository accompanies the Medium article “CON(e)VOLUTION – The evolution of deep learning vision architectures”.

The article shows my journey in implementating and experimenting with classical and modern vision architectures — from early CNNs like LeNet and AlexNet to Inception, ResNet, DenseNet, and Vision Transformers (ViT) — applied to an image classification task.

The purpose of this project is educational and exploratory: understand how vision architectures evolved, what problems they were designed to solve, and how their design choices affect learning behavior — not to propose new models or chase state-of-the-art results.

The architectures explored within this project are:

• BaselineCNN
• LeNet
• LeNetMod
• LeNetModNorm
• LeNetModNorm2
• AlexNet
• GoogLeNet
• ResNet
• DenseNet
• Vit

Read the full article on Medium: CON(e)VOLUTION.

Notes you might want to know about this project

• Models are implemented in PyTorch.
• Experiments use the Neu Surface Defects Dataset.
• AI tools were used for proofreading; all code and content are original.

Getting Started

Setup for training a model

1. use git clone to clone the repository in your local directory.
2. next, use the command cd image_classification to go inside the project's directory.
3. use the command uv sync --upgrade to ensure all project dependencies are installed and up-to-date with the lockfile.
4. activate the environment with the source .venv/bin/activate command.
5. train a model wirh the following example:

python main.py \
    --model_name lenet \
    --epocs 1000 \
    --learning_rate 0.0001 \
    --warmup_period 2000 \

Thank you for taking the time to read the Medium story and this repository. As always, feedback, corrections, or suggestions are deeply appreciated.