Go Neural Network

This project implements a feed-forward neural network in Go, designed to be flexible and adaptable to various datasets. It currently supports training new models and loading pre-trained models for prediction.

Features

• Modular Design: Code is organized into separate packages (cli, data, neuralnetwork, tui) for better maintainability and reusability.
• Dynamic Network Architecture: A feed-forward neural network with a configurable number of hidden layers and neurons per layer.
• Multiple Activation Functions: Supports ReLU, Sigmoid, Tanh, and Linear activation functions for each hidden layer and the output layer.
• Training: Train the neural network using your own CSV data. The data is automatically split into training and testing sets.
• Model Persistence: Save and load trained models to/from model.json files.
• Prediction: Use a loaded model to make predictions on new input data.
• Real-time Visualization: Optional SDL2-based visualization showing network structure, activations, and weight strengths during training.
• He Initialization: Weights are initialized using He initialization.
• Backpropagation: Implements the backpropagation algorithm for training.

Getting Started

To run this application, you need to have Go installed on your system.

1. Clone the repository:

   git clone https://github.com/andrewthecodertx/go-neural-network.git
   cd go-neural-network

2. Run the application:

   The application features a full-screen terminal user interface (TUI).

   go run .

Usage

Upon launching the application, you will be greeted with the main menu. You can navigate through the interface using the arrow keys and press Enter to select an option. Press q or Ctrl+C to quit at any time.

Train New Model

1. Select "Train New Model" from the main menu.
2. The application will automatically find any .csv files in the root directory.
3. Fill out the configuration form:
   • Select CSV File: The number corresponding to the dataset you want to use.
   • Hidden Layers: A comma-separated list of neuron counts for each hidden layer (e.g., 20,20).
   • Hidden Activations: A comma-separated list of activation functions (relu, sigmoid, tanh, linear).
   • Output Activation: The activation function for the output layer.
   • Epochs: The number of training iterations.
   • Learning Rate: The step size for gradient descent.
   • Error Goal: The target error at which training will stop.
   • Enable Visualization: Enter y to enable real-time SDL2 visualization of the network during training.
4. Navigate to the "[ Start Training ]" button and press Enter.
5. If visualization is enabled, an SDL2 window will open showing the network structure with live updates. A progress view will show the current epoch and loss in the terminal.
6. After training, the model will be evaluated on the test set, and the accuracy will be displayed.
7. Once training is complete, you will be prompted to enter a name to save the model. The saved model will be placed in the saved_models/ directory.

Load Model & Predict

1. Select "Load Model & Predict" from the main menu.
2. The application will list all models found in the saved_models/ directory.
3. Fill out the prediction form:
   • Select Model: The number corresponding to the model you want to use.
   • Input Data: A comma-separated list of numerical values for prediction. The number of values must match the model's expected input size.
4. Navigate to the "[ Predict ]" button and press Enter.
5. The calculated prediction will be displayed on the screen.

Datasets

The redwinequality.csv and iris.csv datasets are included as samples. These datasets are from the UCI Machine Learning Repository.

Citation: P. Cortez, A. Cerdeira, F. Almeida, T. Matos and J. Reis. Modeling wine preferences by data mining from physicochemical properties. In Decision Support Systems, Elsevier, 47(4):547-553, 2009.

Future Enhancements

• Command-Line Arguments: Allow all parameters to be passed via command-line arguments instead of interactive prompts.
• Additional Optimizers: Implement other optimization algorithms like Adam or RMSprop.
• Regularization: Add support for L1/L2 regularization to prevent overfitting.
• Testing: Add a comprehensive test suite.

Contributing

Contributions are welcome! If you'd like to improve this project, feel free to open an issue or submit a pull request.

License

This project is licensed under the terms of the LICENSE file.