MNIST Neural Network Classifier with Flask API

This project implements a simple neural network classifier for the MNIST dataset using Python. The model is trained using the MLPClassifier from scikit-learn, and a Flask API is provided to make predictions on handwritten digit images.

---

Features

• Dataset: MNIST handwritten digits dataset.
• Model: Neural network with one hidden layer, implemented using MLPClassifier.
• Data Preprocessing:
  • Normalizes pixel values to improve model performance.
  • Splits the dataset into training and testing sets.
• Evaluation:
  • Includes cross-validation to assess model generalization.
  • Monitors training loss during training.
• Deployment:
  • Flask API to accept digit images as input and return predictions.

---

Installation

1. Clone the repository:

   git clone https://github.com/Shahad-irl/mnist-classifier.git
   cd mnist-classifier

2. Install the required Python packages:

   pip install numpy scikit-learn flask matplotlib

---

Description

1. Dataset Loading:

   • The code attempts to load the MNIST dataset using three different sources:
     • fetch_openml from scikit-learn.
     • load_digits from scikit-learn as a fallback.
     • Keras MNIST dataset if the first two fail.
   • The dataset is loaded with error handling to ensure the program continues by falling back to alternative datasets when needed.

2. Data Preprocessing:

   • The pixel values are normalized using StandardScaler from scikit-learn.
   • The dataset is split into training and testing sets using train_test_split.

3. Model Training:

   • A neural network model (MLPClassifier) is defined with one hidden layer of 128 neurons, ReLU activation, and the Adam optimizer.
   • The model is trained using 5-fold cross-validation and the training data.

4. Model Evaluation:

   • The model is evaluated on the test set to calculate accuracy.
   • The training loss curve is plotted to visualize the model's performance during training.

5. Flask API:

   • A simple Flask API is set up with a /predict endpoint that accepts a POST request with a JSON object containing the input data (flattened 28x28 image).
   • The API returns the predicted digit based on the trained model.

---

Usage

Training the Model

1. Run the script to train the model:

   python mnist-classifier.py

2. During training, the script will:

   • Perform 5-fold cross-validation.
   • Train the model on the MNIST dataset.
   • Output training loss and accuracy.

Running the Flask API

1. Start the Flask server:

   python mnist-classifier.py

2. Use an API testing tool (e.g., Postman) or a Python script to send POST requests to the /predict endpoint.

   • Example input:

     {
       "data": [0, 0, 0, ..., 255]  // Flattened 28x28 array of pixel values
     }

   • Example response:

     {
       "prediction": 5
     }

---

File Structure

mnist-classifier/
├── mnist_classifier.py  # Main script
├── README.md            # Project documentation

---

Results

• Cross-Validation Accuracy: Displayed during training.

• Test Accuracy: Achieved after training, evaluated on a separate test set.

• Training Loss Curve: Plots the loss during training to monitor performance.

  • Expected output:

    Cross-validation scores: [0.84375    0.84722222 0.80836237 0.84320557 0.82926829]
    Average cross-validation score: 0.834361691831204
    Test Accuracy: 89.17%
    

    

---

Limitations

• The model is designed for educational purposes and might not handle real-world images without preprocessing.
• Flask API expects input as a flattened 28x28 array of pixel values, normalized between 0 and 1.

---

Future Improvements

• Enhance the neural network by adding more layers or using a deep learning framework (e.g., TensorFlow or PyTorch).
• Integrate image preprocessing for real-world handwritten digit recognition.
• Dockerize the Flask API for easier deployment.

---

Author

Shahad-irl
A computer engineering graduate passionate about artificial intelligence and machine learning.