Fashion-MNIST CNN

A deep learning project implementing a Convolutional Neural Network (CNN) to classify Fashion-MNIST dataset images using PyTorch.

Overview

This project implements a CNN model to classify fashion items from the Fashion-MNIST dataset. The model architecture includes three convolutional layers with batch normalization and dropout, followed by two fully connected layers.

Features

• Custom CNN architecture optimized for Fashion-MNIST classification
• Training with early stopping mechanism
• Real-time training progress visualization
• Model checkpointing
• Prediction visualization with comparison to ground truth
• Comprehensive training history plots

Project Structure

.
├── main.py             # Main execution script
├── src/
│   ├── dataset.py      # Data loading and preprocessing
│   ├── model.py        # CNN model architecture
│   ├── train.py        # Training loop implementation
│   ├── utils.py        # Utility functions
│   └── visualize.py    # Visualization functions
├── models/             # Saved model checkpoints
└── results/            # Training plots and visualizations

Installation

1. Clone the repository:

   git clone https://github.com/xixu-me/Fashion-MNIST-CNN.git
   cd Fashion-MNIST-CNN
   

2. Install dependencies:

   pip install torch torchvision tqdm matplotlib
   

Usage

Run the training script:

python main.py

The script will:

1. Download and prepare the Fashion-MNIST dataset
2. Train the CNN model
3. Generate training history plots
4. Save model checkpoints
5. Visualize predictions

Model Architecture

The CNN architecture consists of:

• 3 Convolutional layers with increasing filter sizes (32, 64, 128)
• Batch Normalization after each convolution
• Dropout for regularization
• MaxPooling layers for dimensionality reduction
• 2 Fully connected layers (512 units, 10 output classes)

Dataset

The Fashion-MNIST dataset includes 10 categories:

class_labels = [
    "T-shirt/top",
    "Trouser",
    "Pullover",
    "Dress",
    "Coat",
    "Sandal",
    "Shirt",
    "Sneaker",
    "Bag",
    "Ankle boot",
]

Results

Training results will be saved in the results/ directory:

• training_history.png: Loss and accuracy curves
• predictions.png: Sample predictions with ground truth comparison

The following are the example results:

License

Copyright © Xi Xu

Licensed under the GPL-3.0 license.