README.md

Micrograd++ is a minimalistic wrapper around NumPy which adds support for automatic differentiation. It also provides various composable classes ("layers") and other tools to simplify building neural networks.

Micrograd++ draws inspiration from Andrej Karpathy's awesome micrograd library, prioritizing simplicity and readability over speed. Unlike micrograd, which tackles scalar inputs, Micrograd++ supports tensor inputs (specifically, NumPy arrays). This makes it possible to train larger networks.

Usage

Micrograd++ is not yet pip-able. Therefore, you will have to clone the Micrograd++ repository to your home directory and include it in any script or notebook you want to use it in by first executing the snippet below:

import sys
sys.path.insert(0, os.path.expanduser("~/micrograd-pp/python"))

Examples

• Train a simple feedforward neural network on MNIST to classify handwritten digits
• Learn an n-gram model to generate text
• Train a decoder-only transformer to generate text

Features

• Core
  • ☒ Reverse-mode automatic differentiation (.backward)
  • ☒ GPU support
• Layers
  • ☒ BatchNorm1d
  • ☒ Dropout
  • ☒ Embedding
  • ☒ LayerNorm
  • ☒ Linear
  • ☒ MultiheadAttention
  • ☒ ReLU
  • ☒ Sequential
• Optimizers
  • ☐ Adam
  • ☒ Stochastic Gradient Descent (SGD)