numpygrad

A NumPy-only autograd and neural network library with a PyTorch-like API. No C++ extensions, no CUDA — just readable, hackable Python.

Getting Started

• Getting Started

Concepts

• How Autograd Works
  • The computation graph
  • Calling backward()
  • Gradient accumulation
  • Non-differentiable operations
  • Broadcasting and gradients
• Disabling Autograd
  • npg.no_grad
  • Use as a decorator
  • Typical usage

User Guide

• The Training Loop
  • Data
  • Model
  • Optimizer
  • The loop
  • Validation
  • Putting it all together
• Custom Modules
  • Minimal example
  • Composing modules
  • Using Sequential
  • Buffers
  • Inspecting parameters

API Reference

• Array
• Random (npg.random)
• Operators
• Neural Network (nn)
• Optimizers (optim)
• Utilities (utils)

Examples

• 1D Regression
• 2D Classification
• MNIST
• GPT-2 Character Language Model