How karpathy/micrograd works
micrograd is a small scalar reverse-mode autodiff library with a PyTorch-like neural-network layer. A user builds a computation from scalar Values, calls backward to obtain gradients, and can use the neural-network API in a documented classifier-training demo. A separate notebook demonstrates graph visualization. The core engine and neural-network code were inspected; demo and visualization behavior is represented only to the extent described in the README.
Main entry points: demo.ipynb Classifier Demo · nn.py Network Modules · engine.py Value Engine · trace_graph.ipynb Graph Visualization
Components 6
Scalar Autodiff
- Value Engine
micrograd/engine.py - Graph Visualization
trace_graph.ipynb
Neural Networks
- Network Modules
micrograd/nn.py - Classifier Demo
demo.ipynb - Decision Boundary
demo.ipynb
Other components
- Library User
Connections 8
- Library User → Value Engine: builds values
README.md - Library User → Value Engine: runs backward
README.md - Network Modules → Value Engine: uses values
micrograd/nn.py - Library User → Classifier Demo: trains with
README.md - Classifier Demo → Network Modules: initializes network
README.md - Classifier Demo → Decision Boundary: produces
README.md - Library User → Graph Visualization: visualizes graph
README.md - Graph Visualization → Network Modules: demonstrates neuron
README.md