FANoS Optimizer (PyTorch)

FANoS = Friction-Adaptive Nosé–Hoover Symplectic momentum.

This package provides a PyTorch torch.optim.Optimizer implementation of FANoS:

• semi-implicit (symplectic-Euler) momentum update
• a Nosé–Hoover-inspired thermostat variable that adapts friction using kinetic-energy feedback
• optional diagonal RMS “mass” (preconditioner) and optional global gradient clipping

The accompanying paper is included in the release repo; this library is just the clean optimizer code.

Install

pip install fanos-optimizer

Quickstart

import torch
from fanos import FANoS

model = torch.nn.Linear(10, 1)
opt = FANoS(model.parameters(), lr=1e-3, grad_clip=1.0)

x = torch.randn(64, 10)
y = torch.randn(64, 1)

loss = torch.nn.functional.mse_loss(model(x), y)
loss.backward()
opt.step()
opt.zero_grad()

Notes (read this before hype happens)

FANoS is a research optimizer. In the paper's reported protocols it:

• helps vs unclipped AdamW/RMSProp on Rosenbrock-100D,
• but is not a general replacement for strong baselines like AdamW + clipping,
• and can be unstable or high-variance on some problems without tuning.

So: treat it as a tool for experiments, not a default choice for production.

Citation

Add the paper citation from CITATION.cff in the repo.

License

MIT (see LICENSE).