A TypeScript implementation of an autograd engine for educational purposes.
MicroGrad implements backpropagation (reverse-mode autodiff) over a dynamically built Directed Acyclic Graph (DAG). This project demonstrates how to implement automatic differentiation principles in TypeScript.
- Value Class: Core autodiff functionality with gradient computation
- Neural Network Primitives: Simple Neuron, Layer, and MLP implementations
- Graph Visualization: Tools to visualize computation graphs
- API Design: Both instance and static methods for operations compared to instance-only methods
- Higher Order Gradients: Support for computing higher-order derivatives
- Extended Math: Additional operations including log, exp, tanh, and sigmoid
- Gradient Tools: Methods for gradient health checks and gradient clipping
- Performance: Iterative stack-based topological sort for better efficiency
import { Value } from '@2bad/micrograd';
// Create computation graph
const a = new Value(-4.0, 'a')
const b = new Value(2.0, 'b')
let c = Value.add(a, b, 'c') // a + b
let d = Value.add(Value.mul(a, b), Value.pow(b, 3), 'd') // a * b + b**3
// c += c + 1
c = Value.add(c, Value.add(c, new Value(1.0)))
// c += 1 + c + (-a)
c = Value.add(c, Value.add(Value.add(new Value(1.0), c), Value.negate(a)))
// d += d * 2 + (b + a).relu()
const bPlusA = Value.add(b, a)
d = Value.add(d, Value.add(Value.mul(d, 2), Value.relu(bPlusA)))
// d += 3 * d + (b - a).relu()
const bMinusA = Value.sub(b, a)
d = Value.add(d, Value.add(Value.mul(3, d), Value.relu(bMinusA)))
// e = c - d
const e = Value.sub(c, d, 'e')
// f = e**2
const f = Value.pow(e, 2, 'f')
// g = f / 2.0
let g = Value.div(f, 2.0, 'g')
// g += 10.0 / f
g = Value.add(g, Value.div(10.0, f))
// Forward pass
console.log(g.data); // Value of the computation
// Backward pass (compute gradients)
g.backward();
// Access gradients
console.log(a.grad); // dg/da
console.log(b.grad); // dg/db
# Install dependencies
npm install
# Build the project
npm run build
# Run tests
npm test
This project is inspired by micrograd by Andrej Karpathy. The TypeScript implementation extends the core concepts with additional features and type safety.