gradlite 0.2.0

A tiny, elegant automatic differentiation engine built from scratch.

🚀 GradLite

A tiny, elegant automatic differentiation engine built from scratch.

GradLite is a lightweight module for automatic gradient computation, the core mechanism behind training neural networks using backpropagation.

Inspired by:

• 🔬 MicroGrad
• 🔥 PyTorch Autograd

...but extended to support more features than MicroGrad, without all the complexity of PyTorch. This module supports building complete neural networks and has a clear educational purpose.

---

✨ Features

• 🧠 Reverse-mode automatic differentiation (backpropagation)
• 🔢 Scalar-based computational graph
• 🔄 Dynamic graph construction
• 🏗️ Build full neural networks (Linear layer support for now! 🗓️)
• 🪶 Lightweight and easy to read
• 📚 Fully Educational

---

📦 Installation

You can run the following command to install gradlite in your system/virtual environment:

pip install gradlite@git+https://github.com/RaulBarbaRojas/gradlite.git

🧩 Example Usage

The usage of gradlite to build expressions and run backpropagation (automatic differentation) is shown next:

from gradlite.core import Parameter

a = Parameter(2.0)
b = Parameter(0.3)
c = Parameter(1.0)

ab = a * b
f = ab + c
print(f'f={f.value:.4f}')
f.backward()
print(f'grad_f={f.grad:.4f}')
print(f'grad_a={a.grad:.4f}')
print(f'grad_b={b.grad:.4f}')
print(f'grad_c={c.grad:.4f}')

An example training loop of gradlite is given next:

from gradlite.nn.linear import Linear
from gradlite.nn.activations import ReLU
from gradlite.nn.loss.mse import MSE
from gradlite.optimizer.sgd import SGD

X = [2.00, 2.10, 2.20, 2.30, 2.40]
y = [3.00, 3.10, 3.13, 3.16, 3.19]

epochs = 2000
model = Linear(input_features=1, output_features=1, activation_fn=ReLU())
model.neurons[0].weights[0].value = 0.5  # Using positive weight to prevent Dying ReLU
optimizer = SGD(model.parameters(), lr=1e-2)
loss_fn = MSE()

for _ in range(epochs):
    for x_value, y_true in zip(X, y):
        optimizer.zero_grad()
        y_pred = model([x_value])
        loss = loss_fn(y_pred, [y_true])
        loss.backward()
        optimizer.step()
        print(f'Loss: {loss.value:.4f}')