GeN: Generalized Newton's Method for Learning-Rate-Free Optimization 🚀

---

Paper: Gradient Descent with Generalized Newton’s Method (ICLR 2024)

---

📦 Repository Overview

This repository contains the code and examples for Generalized Newton's method as a learning-rate-free optimization. It supports a wide range of models and tasks, including:

• 🖼️ Image classification (CIFAR10/CIFAR100/ImageNet... datasets with ViT/ResNet models)
• 📝 Natural language generation (E2E/DART... datasets with GPT2 models)
• 📊 Natural language understanding (SST2/QNLI/MNLI... datasets with BERT/RoBERTa models)
• 🕵️‍♂️ Object detection / Instance segmentation
• 🎯 Recommendation system

Example scripts are provided for each task in the examples/ directory. The core implementation of GeN optimizer can be found in GeN/, which roughly has the same speed and memory cost as the base optimizers.

⚡ Quickstart

🛠️ Installation

Install the package from PyPI:

pip install gen-optim

🏃 Minimal Training Loop

To use GeN in your PyTorch training loop, simply add two lines between backward() and optimizer.step():

from GeN import lr_parabola
optimizer = AdamW(model.parameters(), lr=1e-4)
tr_iter = iter(train_loader)

# Standard training pipeline
loss = F.cross_entropy(model(batch), labels)
loss.backward()
if (batch_idx+1) % lazy_freq == 0:
    lr_parabola(model, optimizer, tr_iter=tr_iter, task='image_cls', scale=scale)
optimizer.step()
optimizer.zero_grad()

• scale can be used to enable the horizon-aware learning rate (e.g., np.linspace(1,0,epochs+1)).
• Call lr_parabola infrequently (a.k.a. lazy update) by setting lazy_freq>=4 for efficiency.
• Different task values need different forward passes. Can be customized.

🧩 Function Overview

The main function is lr_parabola, which adapts the learning rate based on a quadratic curve fitting to the loss landscape, with minimal code changes and computational overhead. This enables learning-rate-free optimization and leverages the Hessian information, like the Newton–Raphson method.

Mathematically, we turn any base optimizer (e.g. SGD or AdamW) to the GeN optimizer by

where g_t is the stochastic pre-conditioned gradient, G_t is the oracle gradient and H_t is the oracle Hessian.

To enable the horizon-aware GeN, like cosine or linear decay learning rates, we use hyperparameter-free one-to-zero decay (controlled by `scale`):

     

📚 Citation

If you use GeN in your research, please cite:

@inproceedings{bu2024gradient,
  title={Gradient descent with generalized newton’s method},
  author={Bu, Zhiqi and Xu, Shiyun},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2024}
}