Koppara is a framework that enables the implementation of fully functional neural networks, machine learning models, and more—all natively within Godot.
Although some applications may be challenging, Koppara holds the potential for implementing:
- Language Models
- Computer Vision Models
- Voice Recognition Models
The Koppara demo includes:
- Self-Driving Simple Car: An autonomous driving simulation using basic neural networks.
- RNN-Based Language Model: A simple Recurrent Neural Network for basic language generation.
- Clone or Copy this repository to
res://in your Godot project. - Use the Demos as a reference to implement and train your own models.
Koppara utilizes a sequential, layer-by-layer, neuron-by-neuron architecture to build neural networks from the ground up, providing flexibility for custom implementations.
Koppara introduces a unique optimization approach, distinct from traditional algorithms:
- Random Variation: The optimizer applies random variations to all weights.
- Gradient Selection: If the model improves, this gradient is saved for the next iteration.
- Negated Variation: If no improvement occurs, a negated variation is attempted in the next iteration, and previous weights are restored.
- Revert & Retry: If the negated variation also fails to improve the model, a new random gradient is generated, and weights are restored.
- Repeat: This cycle continues, iterating until optimal performance is reached.
Koppara is in early development, with known and undiscovered bugs. It’s not production-ready, but is suitable for experimentation, learning, and testing in the Godot environment.
We welcome contributions to help expand Koppara’s potential! Join us in pushing Koppara to new heights.
For support and updates, join our Discord: Koppara Discord