MiniTorchBRAutograd from scratch

A lightweight PyTorch-inspired deep learning framework built on NumPy — learn how backpropagation really works.

MiniTorchBR neural network computation graph

⚡

Reverse-mode Autograd

Full automatic differentiation engine with dynamic computation graphs. Every operation tracks its gradient function for efficient backprop.

🧠

Neural Network Modules

Familiar nn.Module API with Linear layers, Sequential containers, activation functions, and loss functions — just like PyTorch.

🔧

Optimizers

SGD with momentum and Adam with bias correction, ready to train any model you build.

📊

Rich Visualizations

Interactive HTML computation-graph rendering, training-curve plots, weight histograms, and confusion matrices out of the box.

📦

Install with pip

Available on PyPI. Zero heavy dependencies — just NumPy, Matplotlib, and PyVis.

🎓

Educational Design

Every module is readable source code. Ideal for learning or teaching the internals of modern deep learning frameworks.

Quick Install

bash

pip install minitorchbr

30-Second Example

python

import numpy as np
from MiniTorch.core.variable import Variable

# Create tensors with gradient tracking
x = Variable(np.array([[1.0, 2.0, 3.0]]))
w = Variable(np.random.randn(3, 1))

# Forward pass — graph is built automatically
y = x @ w          # matmul
loss = (y ** 2).sum()

# Backward pass
loss.backward()

print(w.grad)      # dL/dw computed via reverse-mode AD