Sine AI

A neural network that learns to approximate the sine function using PyTorch.

Features

• Modular Architecture: Clean separation of concerns with dedicated modules for model, data, training, and inference
• Model Persistence: Save and load trained model weights
• Three Operation Modes:
  • Train: Train a new model and save weights
  • Evaluate: Load trained model and visualize performance
  • Predict: Interactive mode for real-time predictions

Installation

# Clone the repository
git clone <repository-url>
cd sine-ai

# Install dependencies (using uv)
uv sync

Usage

Interactive Menu

Run without arguments to get an interactive menu:

uv run python main.py

This will display:

=== Sine Approximation with PyTorch ===

Select mode:
1. Train - Train a new model and save weights
2. Evaluate - Load and evaluate trained model
3. Predict - Interactive prediction mode
4. Exit

Enter choice (1-4):

Command-Line Interface

Use command-line arguments for direct mode selection:

# Train a new model
uv run python main.py train

# Evaluate existing model
uv run python main.py evaluate

# Interactive prediction mode
uv run python main.py predict

Alternatively, activate the virtual environment:

source .venv/bin/activate
python main.py [train|evaluate|predict]

Interactive Prediction Mode

In predict mode, you can continuously enter values and get predictions:

=== Interactive Sine Prediction ===
Enter a number to get the sine approximation.
Type 'quit' or 'exit' to stop.

Enter x value: 1.5
  Predicted: 0.997495
  Actual:    0.997495
  Error:     0.000000

Enter x value: 3.14159
  Predicted: 0.000026
  Actual:    0.000000
  Error:     0.000026

Enter x value: quit
Exiting prediction mode.

Project Structure

sine-ai/
├── main.py                     # Entry point with CLI
├── sine_ai/                    # Main package
│   ├── __init__.py             # Package initialization
│   ├── config.py               # Configuration settings
│   ├── model.py                # SineNet architecture
│   ├── data.py                 # Data generation
│   ├── train.py                # Training and evaluation
│   └── inference.py            # Model loading and prediction
├── models/                     # Saved model weights (created automatically)
│   └── sine_model.pth
├── pyproject.toml              # Project dependencies
└── README.md                   # This file

Configuration

Default hyperparameters can be modified in sine_ai/config.py:

• Model Architecture: Hidden layer sizes (64, 64, 32)
• Training: 5000 epochs, learning rate 0.001
• Data: 1000 training samples, 200 test samples
• Model Path: models/sine_model.pth

Model Architecture

The SineNet model uses a simple feedforward architecture:

• Input: 1 neuron (x value)
• Hidden Layer 1: 64 neurons + ReLU
• Hidden Layer 2: 64 neurons + ReLU
• Hidden Layer 3: 32 neurons + ReLU
• Output: 1 neuron (sin(x) approximation)

Dependencies

• Python >= 3.14
• PyTorch >= 2.9.1
• NumPy >= 2.3.5
• Matplotlib >= 3.10.7
• Jupyter >= 1.1.1 (for notebook experiments)

Development

Keeping Notebooks Clean in Git

This project uses nbstripout to prevent notebook outputs and metadata from cluttering git commits. It's already configured in the dev dependencies.

To set it up for your local repository:

# nbstripout is included in dev dependencies via uv sync
uv run nbstripout --install

This adds a git filter that automatically strips:

• Cell outputs
• Execution counts
• Metadata

The notebook (sine_approximation.ipynb) will remain functional with its outputs when you work locally, but commits will only include the code and markdown cells.

Manual setup (if not using the dev dependencies):

pip install nbstripout
nbstripout --install --attributes .gitattributes

Verify it's working:

git diff --cached sine_approximation.ipynb

After making changes to the notebook, the diff should only show code/markdown changes, not execution metadata.

License

This project is open source and available under the MIT License.