MiniTorch

Fast NumPy autograd, compiled dense training, and inspectable neural networks.

MiniTorch is a compact scientific-computing framework for learning and experimenting with reverse-mode automatic differentiation. Its Python surface stays readable while first-order backpropagation uses raw NumPy gradient kernels and static dense classifiers can move the complete training loop into compiled C.

• :material-lightning-bolt:{ .lg .middle } Fast backpropagation

  ---

  A linear-time reverse tape visits every graph node and edge once, without constructing a second graph during ordinary first-order training.

• :material-language-c:{ .lg .middle } Compiled training

  ---

  Dense Linear/ReLU classifiers can run batching, forward, backward, softmax cross-entropy, and Adam in generated C while NumPy calls BLAS.

• :material-graph-outline:{ .lg .middle } Scientific model explorer

  ---

  visualize(model) opens a self-contained full-neuron connection map. Selecting a neuron shows only its latest value and gradient.

Get started Explore the model viewer

Thirty-second example

import numpy as np

from MiniTorch import Variable

x = Variable(np.array([[2.0]], dtype=np.float32), name="x")
w = Variable(np.array([[3.0]], dtype=np.float32), name="w")
b = Variable(np.array([[1.0]], dtype=np.float32), name="b")

loss = (x * w + b - 10.0) ** 2
loss.backward()

print(loss.data)    # [[9.]]
print(w.grad.data)  # [[-12.]]

Choose a training path

Eager autogradCompiled dense loop

Use eager execution for custom modules, dynamic Python control flow, new operations, or higher-order gradients.

logits = model(Variable(features))
loss = softmax_cross_entropy(logits, Variable(labels))
optimizer.zero_grad()
loss.backward()
optimizer.step()

Use the native loop for a static Sequential(Linear, ReLU, ..., Linear) classifier.

from MiniTorch.native import train

history = train(
    model,
    x_train,
    y_train,
    epochs=15,
    batch_size=128,
    lr=1e-3,
)

Inspect the trained network

from MiniTorch import visualize

visualize(model)