freegrad

Alternative backward rules and gradient transforms alongside PyTorch autograd.

---

🔥 Why freegrad?

freegrad is a lightweight research framework that lets you decouple forward and backward in PyTorch. It provides: - Custom gradient rules (e.g. noise, clipping, jamming) - A clean context manager API for applying rules selectively - Drop-in activation wrappers with alternative backward behavior - Compatibility with vanilla autograd — nothing is patched

---

🚀 Quickstart

pip install -e .[dev]

import torch
import freegrad as xg
from freegrad.wrappers import Activation

x = torch.randn(8, requires_grad=True)
act = Activation(forward="ReLU")

with xg.use(rule="rectangular_jam", params={"a": -1.0, "b": 1.0}, scope="activations"):
    y = act(x).sum()
    y.backward()

print(x.grad)

---

📖 Documentation

• Getting started
• Selective gradient rules
• Defining custom rules

---

🤝 Contributing

Contributions are very welcome! Please see CONTRIBUTING.md.

---

📄 License

Released under the MIT License.