LeNet-5 From Scratch

This repository contains the implementation of LeNet-5 Convolutional Neural Network (CNN) Architecture from scratch using numpy. LeNet-5 is a pioneering architecture in the field of deep learning, particularly for handwritten digit classification on the MNIST dataset, developed in the 90s by Yann LeCun.

Architecture

Table of Content

1. models/ - A directory storing Pickle files for pre-trained models for inference
2. modules/ - A directory comprises of the implementation of the layers, loss functions, optimizers and all other modules of the model
3. LeNet5_RBF_model.ipynb - The implementation of LeNet-5 with the original RBF output layer
4. LeNet5_softmax_model.ipynb - The implementation of LeNet-5 with softmax output for higher prediction accuracy
5. app.py - The implementation of the handwritten digit recognition app with LeNet-5 as the model using tkinter

Installation

Clone the Repository

git clone https://github.com/T0mLam/LeNet-5-from-scratch.git
cd LeNet-5-from-scratch

Install dependencies

pip install -r requirements.txt

Usage

Launch the app

python -m app

Import and use the modules

Create a new file experiment.py

# Use the format 'from modules.{filename} import {module}'

# e.g. import the Adam optimizer
from modules.optimizer import Adam
...
optimizer = Adam(model, lr=0.001)

Run the script

python -m experiment

Acknowledgements

• Yann Lecun for the LeNet paper