MNIST Interactive Model Analyzer

An interactive tool for experimenting with MNIST handwriting recognition models using a custom-built 28x28 drawing grid. This application allows users to draw digits and get real-time predictions from a trained MNIST model.

Features

• Interactive 28x28 drawing grid that matches MNIST input dimensions
• Real-time prediction updates as you draw
• Blur effect to simulate natural handwriting
• Clear canvas functionality
• Grid overlay for precise pixel manipulation
• Support for custom MNIST-compatible models
• Allows for custom conversion and output functions to allow for increased compatability with different model types
• Display neuron activation values in Tkinter window given the drawing grid values

Installation

1. Clone the repository:

git clone https://github.com/LucaLow/MNIST-Interactive-Model-Analyzer.git
cd MNIST-Interactive-Model-Analyzer

Or install through pip:

pip install -i https://test.pypi.org/simple/ mnist-interactive

2. Install the required dependencies:

pip install numpy
pip install tensorflow

Requirements

• Python >= 3.8
• TensorFlow >= 2.0
• NumPy
• Tkinter (usually comes with Python)

Usage

1. Basic usage with a pre-trained model:

import mnist_interactive.numberCreatorWindow as mi
import tensorflow as tf
import tkinter as tk
from tkinter import ttk


# Load your pre-trained MNIST model
model = tf.keras.models.load_model('sampleModel.keras')

# create the tkinter window
root = tk.Tk()


# Initialize the interactive grid
grid = mi.NumberCreatorWindow(root, model=model, blur=0.3)

root.mainloop()

2. Customize the blur effect:

app = NumberCreatorWindow(root, model=model, blur=0.15)  # Adjust blur intensity

Controls

• Left Click: Draw
• Right Click: Erase
• Clear Button: Reset the canvas
• Real-time predictions displayed above the canvas

Project Structure

• numberCreatorWindow.py: Main application window and drawing logic
• utils.py: Utility functions for predictions and grid operations
• example.py: Example implementation
• requirements.txt: Required Python packages
• setup.py: Package installation configuration

src/mnist_interactive/
├── backends/
│   ├── __init__.py
│   ├── tensorflow_backend.py
│   ├── pytorch_backend.py
│   └── base_backend.py
├── ui/
│   ├── __init__.py
│   ├── number_creator_window.py
│   ├── canvas.py
│   └── visualization.py
├── utils/
│   ├── __init__.py
│   ├── grid.py
│   └── blur.py
└── examples/
    ├── tensorflow_example.py
    └── pytorch_example.py

API Reference

NumberCreatorWindow

NumberCreatorWindow(root, model, blur=0.1, conversion_function, output_function)

Parameters:

• root: Tkinter root window
• model: TensorFlow model for digit prediction
• blur: Blur intensity for drawing (default: 0.1)
• conversion_function: Function receiving a 28×28 NumPy array and returning data in the shape/format the model accepts
• output function: Function receiving the model’s raw output (e.g. probability array) and returning a tuple of (prediction int, confidence float)

Methods

• clear_canvas(): Clear the drawing grid
• predict(): Get prediction for current drawing

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Author

Luca Lowndes (Luca@Lowndes.net)

Acknowledgments

• Built using TensorFlow and MNIST dataset
• Inspired by the need for interactive MNIST model testing tools based upon the models you create