FashionMNIST CNN with PyTorch

This repository contains an implementation of a Convolutional Neural Network (CNN) for the FashionMNIST dataset using PyTorch. FashionMNIST is a dataset of Zalando's article images—consisting of 60,000 training examples and 10,000 test examples. Each example is a 28x28 grayscale image associated with a label from 10 classes.

Project Structure

fashion-mnist-cnn/
│
├── src/
│   ├── dataset.py       # Data loading and preprocessing
│   ├── model.py         # CNN architecture definition
│   ├── train.py         # Training loop and logic
│   ├── evaluate.py      # Model evaluation script
│   └── utils.py         # Utility functions for training and evaluation
│
├── tests/
│   └── test_model.py    # Unit tests for the model
│
├── requirements.txt     # Dependencies for the project
├── README.md            # This file
└── LICENSE              # Project license

CNN Architecture

The CNN model defined in model.py employs the following structure:

• Input Layer: Takes in a 1x28x28 image (grayscale).
• Convolutional Layers:
  • Two convolutional blocks:
    • First Block:
      • Conv2D: 64 filters, kernel size 3x3, padding 1
      • ReLU activation
      • Batch Normalization
      • Conv2D: 64 filters, kernel size 3x3, padding 1
      • ReLU activation
      • Batch Normalization
      • Max Pooling: kernel size 2, stride 2
      • Dropout: p=0.25 for regularization
    • Second Block:
      • Conv2D: 128 filters, kernel size 3x3, padding 1
      • ReLU activation
      • Batch Normalization
      • Conv2D: 128 filters, kernel size 3x3, padding 1
      • ReLU activation
      • Batch Normalization
      • Max Pooling: kernel size 2, stride 2
      • Dropout: p=0.25 for regularization
• Fully Connected Layers:
  • Flatten layer
  • Dense layer with 256 units, ReLU activation, followed by Batch Normalization
  • Dropout with p=0.5
  • Output layer with 10 units (one for each class in FashionMNIST), no explicit softmax as CrossEntropyLoss in PyTorch includes it.

Training

The training script train.py uses:

• Optimizer: Adam with learning rate = 0.01
• Loss Function: Cross Entropy Loss
• Batch Size: 64
• Epochs: Configurable, default set to 20
• Device: Automatic selection between CUDA (if available) or CPU

Implementation Details:

• Data Augmentation: Random rotation up to 10 degrees, random horizontal flips to increase dataset diversity.
• Learning Rate Scheduler: StepLR with step_size=7, gamma=0.1 to decrease learning rate over time.
• Early Stopping: Implemented to prevent overfitting, patience set to 5 epochs.

Evaluation

The evaluate.py script provides:

• Metrics: Accuracy, Confusion Matrix, Classification Report
• Visualization: Examples of misclassified images, ROC curves for multi-class classification if extended.

Setup Instructions

1. Clone this repository:

   git clone https://github.com/mathiasmendozav/DeepCNN-Pytorch-FashionMNIST
   cd DeepCNN-Pytorch-FashionMNIST

2. Install dependencies:

   pip install -r requirements.txt

3. Download the FashionMNIST dataset: This is automatically handled by src/dataset.py on the first run.

4. Train the model:

   python src/train.py

5. Evaluate the model:

   python src/evaluate.py

Advanced Features

• TensorBoard Integration: For real-time tracking of training metrics.
• Model Checkpointing: Saving the best model based on validation loss.
• Hyperparameter Tuning: Script included for tuning using grid search or random search over defined parameters.

Contributing

Contributions are welcome. Please fork the repository, create a feature branch, and submit a pull request. Ensure your code passes the existing unit tests and adds new tests if introducing new functionality.

License

This project is licensed under the MIT License.