A simple and understandable deep learning framework built from scratch.

Ragnarok is a lightweight educational deep learning framework designed to help you understand the inner workings of neural networks — from variable(tensor) operations and automatic differentiation to layers, optimisers, and training loops.

Original concept and design are based on Dezero.

Most modern frameworks (like PyTorch or TensorFlow) abstract away the fundamentals. Ragnarok is built to demystify how deep learning actually works under the hood — providing readable, minimal code without black-box magic.

• Pure Python implementation with numpy — minimal dependencies
• Variable(Tensor) class with basic operations and gradients
• Automatic differentiation engine (autograd)
• Layers and activations built from first principles
• Optimisers (SGD, Adam, etc.)
• Training loop utilities
• Readable, modular codebase for learning and extension