learngrad 0.2.1

A minimal scalar-valued autograd engine with neural network primitives, built for learning purposes.

learngrad

A minimal scalar-valued autograd engine with neural network primitives, built for learning purposes. Inspired by micrograd.

Install

pip install learngrad

What's inside

• Value — scalar with automatic differentiation via backprop
• MLP — multi-layer perceptron built on top of Value
• Optimizers — SGD, SGDMomentum, RMSProp, Adam

Quick example

from learngrad.engine import Value
from learngrad.nn import  MLP
from learngrad.optimizers import Adam

# Autograd
x = Value(2.0)
y = x ** 2 + x * 3
y.backward()
print(x.grad)  # dy/dx = 2x + 3 = 7.0

# Neural net
model = MLP(2, [4, 4, 1])
opt = Adam(model.parameters(), lr=1e-3)

x = [Value(1.0), Value(0.5)]
out = model(x)
out.backward()
opt.step()

Demo

notebooks/demo.ipynb walks through a full training example:

• Binary classification on the make_circles dataset
• MLP with hinge loss and L2 regularization
• Mini-batch training loop with Adam
• Train/val accuracy tracking
• Decision boundary visualization

Reaches ~94% val accuracy in 100 epochs.

Requirements

• Python >= 3.10
• numpy