Guided Program Synthesis for MNIST Classification

This project demonstrates a simple guided program synthesis approach applied to the MNIST dataset. A model (Oracle) is trained to synthesize a binary program that classifies handwritten digits. The synthesized program is saved and used for inference.

Features

• Oracle Model: A neural network that generates binary program rules.
• ProgramState: Applies the learned rules to classify MNIST digits.
• Training: Train the Oracle to find optimal program rules.
• Inference: Use the saved program to classify MNIST test data.

Files

• main.py: Training script for generating the binary program.
• inference.py: Inference script to test the binary program on MNIST test data.

Usage

1. Train the Oracle:

   python main.py

   The trained binary program will be saved as binary_program.pth.

2. Run Inference:

   python inference.py

   The script will load the binary program and calculate test accuracy.

Output

• Training: Displays loss, training accuracy, and active rules at each step.
• Inference: Outputs the final test accuracy.

Explanation

This project trains an Oracle model to generate binary program rules for digit classification. The ProgramState module applies these rules to one-hot-encoded MNIST data. The process combines machine learning and programmatic reasoning.