Digit Decoder

Digit Decode is a Python-based handwritten digit recognition program that uses a trained ResNet-50V2 This project is a digit recognition application that uses a trained ResNet-50V2 model to predict handwritten digits. Users can draw a digit on a Pygame canvas, and the program will use the trained model to recognize and predict the digit. The predicted digit is printed in the terminal.

---

Project Overview

This project allows users to:

1. Train a ResNet-50V2 model on the MNIST dataset for digit recognition.
2. Interact with the trained model by drawing digits on a Pygame canvas.
3. Obtain predictions from the model, which will be displayed in the terminal.

---

Files in the Repository

1. trainer.py

   • Trains a ResNet-50V2 model on the MNIST dataset.
   • Saves the trained model as resnet50v2_mnist_model.keras.
   • Handles model evaluation on the test dataset, providing test accuracy and loss metrics.

2. canvas.py

   • A Pygame-based interactive canvas where users can draw digits using their mouse/trackpad.
   • Uses the trained ResNet-50V2 model (resnet50v2_mnist_model.keras) to predict digits.
   • The predicted digit is printed directly to the terminal when the user presses the spacebar.
   • Clear the canvas by clicking the C key.

3. resnet50v2_mnist_model.keras

   • Pre-trained ResNet-50V2 model file for digit recognition using the following parameters:
     • Epochs: 10
     • Batch Size: 64
     • Learning Rate: 0.001 (default for Adam optimizer)
     • Loss Function: Sparse Categorical Crossentropy
     • Validation Data: 10,000 test samples from the MNIST dataset
   • Used by canvas.py to make predictions without retraining the model.

---

Future Plans

• [Short Term] Revamping Pygame Canvas to be more visually appealing
• [Short Term] Train model using Google Colab so that more epochs can be run during training
• [Short Term] Plot epoch times and accuracy using matplotlib
• [Long Term] Support for multi-digit numbers (i.e. recognizing '5' and '1' as '51')

---

Acknowledgements

I would like to thank my friend, Parshva, who introduced me to some of the key concepts used to make this project a possibility.