PyRae

By Rae Chipera | Jaxorik AI Research Group

A neural network framework built from scratch in pure NumPy, modeled after PyTorch's API. PyRae is designed for anyone who wants full visibility into every line of a forward pass, backward pass, and weight update — no black boxes.

Originally developed as part of Carnegie Mellon University's Deep Learning bootcamp curriculum.

---

🧠 What's Included

nn — Neural Network Layers (torch.nn equivalent)

Class	Description
Linear	Fully connected layer with Xavier initialization
ReLU	ReLU activation function
Sequential	Chains layers into a feed-forward network
CrossEntropyLoss	Cross-entropy loss with numerical stability (LogSumExp)

optim.py — Optimizers (torch.optim equivalent)

Class	Description
SGD	Stochastic Gradient Descent with optional momentum
Adam	Adam optimizer with bias correction

---

🚀 Quick Start

from PyRae import nn, optim

# Build a model
model = nn.Sequential(
    nn.Linear(3, 4),
    nn.ReLU(),
    nn.Linear(4, 2)
)

# Define loss and optimizer
loss_fn = nn.CrossEntropyLoss()
optimizer = optim.SGD(model, lr=0.01, momentum=0.9)

# Training loop
for x_batch, y_batch in dataloader:
    optimizer.zero_grad()
    out = model.forward(x_batch)
    loss = loss_fn.forward(out, y_batch)
    model.backward(loss_fn)
    optimizer.step()

---

💡 Why PyRae?

PyTorch and TensorFlow are great — until you need to understand or control exactly what's happening under the hood. PyRae is useful for:

• Learning how backpropagation actually works
• Research requiring custom gradient logic
• Teaching deep learning fundamentals without framework magic

---

🔧 Requirements

• Python 3.x
• NumPy

---

License

MIT © Rae Chipera / Jaxorik AI Research Group