MNIST Digit Classification with TensorFlow

A simple machine learning project that trains a neural network to recognize handwritten digits using TensorFlow and the MNIST dataset.

Project Overview

This project demonstrates a complete ML pipeline:

• Data loading and preprocessing
• Neural network model building
• Model training with early stopping
• Model evaluation and predictions
• Visualization of results

Dataset

The MNIST dataset contains 70,000 grayscale images of handwritten digits (0-9):

• 60,000 training images
• 10,000 test images
• Image size: 28x28 pixels

Model Architecture

• Input Layer: 784 features (28x28 flattened)
• Hidden Layer 1: 128 neurons, ReLU activation, 20% dropout
• Hidden Layer 2: 64 neurons, ReLU activation, 20% dropout
• Output Layer: 10 neurons (one per digit), Softmax activation

Project Structure

ml-tensorflow-project/
│
├── data/                      # Data directory (MNIST auto-downloads)
├── models/                    # Saved models and plots
│   ├── mnist_model.keras     # Trained model
│   ├── training_history.png  # Training plots
│   └── predictions_sample.png # Prediction visualizations
│
├── train_model.py            # Training script
├── predict.py                # Prediction script
├── requirements.txt          # Python dependencies
└── README.md                 # This file

Installation

1. Install Python 3.8 or higher

2. Install dependencies:

pip install -r requirements.txt

Usage

1. Train the Model

Run the training script to train the neural network:

python train_model.py

This will:

• Load and preprocess the MNIST dataset
• Build the neural network
• Train for up to 20 epochs (with early stopping)
• Save the trained model to models/mnist_model.keras
• Generate training history plots

Expected accuracy: ~97-98% on test set

2. Make Predictions

Run the prediction script to test the trained model:

python predict.py

This will:

• Load the trained model
• Evaluate performance on the test set
• Show predictions on 10 random samples
• Optionally enter interactive mode to test specific images

Interactive Mode

In interactive mode, you can:

• Enter any test image index (0-9999)
• See the true label, predicted label, and confidence
• View probability distribution for all digits
• Visualize the image

Results

The model typically achieves:

• Test Accuracy: ~97-98%
• Training Time: 2-5 minutes (CPU)
• Model Size: ~400 KB

Key Features

• ✅ Simple and easy to understand
• ✅ Well-commented code
• ✅ Reproducible results (random seeds set)
• ✅ Early stopping to prevent overfitting
• ✅ Dropout layers for regularization
• ✅ Comprehensive evaluation metrics
• ✅ Interactive prediction mode
• ✅ Visualization of training and predictions

Requirements

• Python 3.8+
• TensorFlow 2.15+
• NumPy
• Matplotlib
• scikit-learn

Next Steps

To extend this project, consider:

• Try different model architectures (CNN, deeper networks)
• Experiment with hyperparameters (learning rate, batch size)
• Add data augmentation
• Implement confusion matrix visualization
• Try other datasets (Fashion-MNIST, CIFAR-10)
• Deploy the model as a web application

License

This project is for educational purposes.