Getting Started

Installation

pip install leanpass

That's it. No CUDA, no C++ extensions, no 500MB download. Just NumPy.

Your First Forward Pass

import numpy as np
from leanpass import Tensor

# Create some inputs
x = Tensor(np.array([1.0, 2.0, 3.0]), requires_grad=True)
w = Tensor(np.array([0.5, -0.2, 0.1]), requires_grad=True)
b = Tensor(np.array([0.0]), requires_grad=True)

# Forward pass
y = (x * w).sum() + b

print(y)  # Tensor(shape=(), requires_grad=True)

Your First Backward Pass

# Backward pass — trace gradients through the graph
y.backward()

print(w.grad)  # [1. 2. 3.]
print(x.grad)  # [ 0.5 -0.2  0.1]
print(b.grad)  # [1.]

LeanPass builds a computation graph as you write expressions, then walks it backward when you call .backward() — applying the chain rule at each node.

Your First Model

from leanpass.nn import MLP, mse_loss
from leanpass.optim import SGD

# 2-layer network: 4 → 16 → 2
model = MLP([4, 16, 2])
optimizer = SGD(model.parameters(), lr=0.01)

# Dummy data
x = Tensor(np.random.randn(32, 4))
target = Tensor(np.random.randn(32, 2))

for step in range(100):
    pred = model(x)
    loss = mse_loss(pred, target)

    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

    if step % 20 == 0:
        print(f"step {step:3d}  loss {loss.data:.6f}")

That's the core loop — you'll see this pattern throughout the demos.