CNN Architecture Optimization
Project Overview
This project focuses on the iterative improvement of Convolutional Neural Network (CNN) architectures, specifically variants of the LeNet-5 model, for image classification using the FashionMNIST dataset.
Key Objectives
- Implement the original LeNet-5 architecture from scratch using PyTorch
- Develop and evaluate successive variants of LeNet-5, each building upon the previous version
- Optimize model performance for the FashionMNIST classification task
- Gain insights into the impact of architectural modifications on CNN performance
Methodology
- Base Implementation:
- Recreate the LeNet-5 architecture using PyTorch
- Iterative Improvement:
- Design and implement variants of LeNet-5
- Each variant aims to enhance performance over its predecessor
- Training and Validation:
- Utilize the FashionMNIST dataset for model training and evaluation
- Performance Tracking:
- Employ Weights & Biases (wandb) for comprehensive performance monitoring and analysis
- Architectural Optimization:
- Systematically modify network architecture to improve classification accuracy
Tools and Technologies
- PyTorch: Deep learning framework
- FashionMNIST: Modern benchmark dataset for image classification
- Weights & Biases: Experiment tracking and visualization
Resources
- Code Repository: GitHub - Convolutional Neural Networks
- Course: Computer Vision at Utrecht University