Skip to content

README.md

Latest commit

 

History

History
194 lines (144 loc) · 4.96 KB

README.md

File metadata and controls

194 lines (144 loc) · 4.96 KB

Fashion-MNIST CNN Classification

A PyTorch implementation of a Convolutional Neural Network (CNN) for classifying Fashion-MNIST dataset. This project demonstrates image classification using deep learning techniques with batch normalization and proper training/validation procedures.

Overview

The Fashion-MNIST dataset consists of 70,000 grayscale images of 10 different clothing categories:

  • T-shirt/top
  • Trouser
  • Pullover
  • Dress
  • Coat
  • Sandal
  • Shirt
  • Sneaker
  • Bag
  • Ankle boot

Each image is 28x28 pixels, and in this implementation, we resize them to 16x16 for faster training.

Model Architecture

The CNN model includes:

  • First Conv Block: Conv2d(1→16) + BatchNorm + ReLU + MaxPool
  • Second Conv Block: Conv2d(16→32) + BatchNorm + ReLU + MaxPool
  • Fully Connected: Linear(512→10) + BatchNorm
  • Output: 10 classes (Fashion-MNIST categories)

Key Features:

  • Batch normalization for better training stability
  • ReLU activation functions
  • Max pooling for dimensionality reduction
  • Cross-entropy loss for multi-class classification
  • SGD optimizer with learning rate 0.1

Requirements

  • Python 3.7+
  • PyTorch 2.0+
  • torchvision
  • matplotlib
  • numpy
  • pillow

Installation

  1. Clone this repository or download the files
  2. Install the required dependencies:
pip install -r requirements.txt

Usage

Training the Model

Run the main script to start training:

python main.py

The script will:

  1. Download the Fashion-MNIST dataset automatically
  2. Display sample images from the dataset
  3. Initialize the CNN model with batch normalization
  4. Train for 5 epochs with validation
  5. Plot training loss and validation accuracy
  6. Save the trained model

Training Parameters

  • Image Size: 16x16 pixels
  • Batch Size: 100
  • Learning Rate: 0.1
  • Epochs: 5
  • Optimizer: SGD
  • Loss Function: Cross-Entropy

Output Files

The script generates:

  • sample_data.png: Sample images from the dataset
  • training_results.png: Training loss and validation accuracy plot
  • fashion_mnist_cnn.pth: Trained model checkpoint

Model Performance

The model typically achieves:

  • Training convergence within 5 epochs
  • Validation accuracy of ~85-90%
  • Fast training due to reduced image size (16x16)

Project Structure

.
├── main.py              # Main training script
├── requirements.txt     # Python dependencies
├── README.md           # Project documentation
├── fashion/            # Dataset directory (created automatically)
├── sample_data.png     # Generated sample images
├── training_results.png # Generated training plots
└── fashion_mnist_cnn.pth # Saved model checkpoint

Key Features

Data Preprocessing

  • Automatic dataset download
  • Image resizing to 16x16 pixels
  • Tensor conversion with proper normalization

Model Architecture

  • Convolutional layers with batch normalization
  • Proper forward pass implementation
  • GPU support (if available)

Training Pipeline

  • Proper train/validation split
  • Batch processing with DataLoader
  • Real-time training progress monitoring
  • Model checkpointing

Visualization

  • Sample data visualization
  • Training metrics plotting
  • Results visualization

Customization

You can modify the following parameters in main.py:

IMAGE_SIZE = 16        # Image dimensions
BATCH_SIZE = 100       # Batch size for training
LEARNING_RATE = 0.1    # Learning rate for SGD
NUM_EPOCHS = 5         # Number of training epochs

GPU Support

The script automatically detects and uses GPU if available:

  • CUDA-enabled GPU will be used automatically
  • Falls back to CPU if GPU is not available

Dataset Information

Fashion-MNIST is a dataset of Zalando's article images consisting of:

  • 60,000 training examples
  • 10,000 test examples
  • 10 classes
  • 28x28 grayscale images

Original dataset: https://github.com/zalandoresearch/fashion-mnist

License

This project is based on educational material and is intended for learning purposes.

Acknowledgments

  • Original Fashion-MNIST dataset by Zalando Research
  • PyTorch community for excellent deep learning framework
  • Educational content adapted from IBM's deep learning course materials

Troubleshooting

Common Issues

  1. CUDA out of memory: Reduce batch size or use CPU
  2. Slow training: Enable GPU or reduce image size
  3. Dependencies issues: Make sure all packages are installed correctly

Performance Tips

  • Use GPU for faster training
  • Increase batch size if you have more memory
  • Adjust learning rate based on convergence behavior
  • Monitor validation accuracy to avoid overfitting

Results Interpretation

The training script provides:

  • Real-time loss and accuracy monitoring
  • Final model performance metrics
  • Visual plots of training progress
  • Saved model for future use

Expected results:

  • Training loss should decrease over epochs
  • Validation accuracy should improve and stabilize
  • Model should achieve reasonable classification performance