hobbitgrad 0.0.1

A tiny educational autograd library.

HobbitGrad

HobbitGrad is a tiny educational autograd library written in pure Python. It includes a small NDArray, a Tensor type with backpropagation, a linear layer, SGD, and mean squared error loss.

It is intended for learning how tensor operations and backpropagation fit together.

Install

pip install hobbitgrad

Example

from hobbitgrad import Linear, SGD, Tensor, mse

x = Tensor([[0, 0], [0, 1], [1, 0], [1, 1]])
y = Tensor([[0], [0], [0], [1]])

model = Linear(2, 1)
optimizer = SGD(model.parameters(), lr=0.1)

for _ in range(100):
    pred = model.forward(x)
    loss = mse(pred, y)
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

print(loss.data.data[0])

Current Scope

HobbitGrad currently supports a small set of tensor operations, broadcasting, matrix multiplication, scalar reductions, a linear layer, SGD, and MSE loss.

The API is early and may change.