Deep Learning Image Classification Project

Overview

This project implements and compares three deep learning CNN architectures from scratch using PyTorch:

• ResNet
• GoogLeNet
• DenseNet

The models are trained on the MNIST handwritten digit dataset for image classification.

---

Dataset

Dataset used:

• MNIST Handwritten Digits Dataset

Dataset contains:

• 60,000 training images
• 10,000 testing images
• 10 classes (digits 0–9)

---

Technologies Used

• Python
• PyTorch
• Torchvision
• Matplotlib
• Pandas

---

Models Implemented

1. ResNet

Implemented using:

• Residual Blocks
• Skip Connections
• Batch Normalization
• Dropout

2. GoogLeNet

Implemented using:

• Inception Blocks
• Multiple convolution paths
• MaxPooling branches

3. DenseNet

Implemented using:

• Dense Connections
• Feature Concatenation
• Batch Normalization

---

Features

• CNN architectures implemented from scratch
• Model comparison
• Training and testing evaluation
• Accuracy visualization
• Loss visualization
• Final comparison chart
• GPU support using CUDA