Luolc/AdaBound

Health v2 · maintenance only

Adoption not part of the score

2902 stars · 334 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

AdaBound is a PyTorch implementation of the AdaBound optimizer from an ICLR 2019 paper, which behaves like Adam early in training and gradually transitions to SGD for better generalization. It is a small Python library installable via pip or copy-paste, usable like any standard PyTorch optimizer.

Use cases

• train deep learning models faster than SGD with better generalization than Adam
• reduce hyperparameter tuning time for neural network training
• optimize CNNs for computer vision tasks like CIFAR-10 classification
• apply a robust optimizer to NLP model training
• replace Adam in a PyTorch training loop with minimal code changes

When to choose

• you use PyTorch and want an optimizer combining Adam's fast early training with SGD's generalization
• you want to spend less time tuning learning rates
• you want a tiny, dependency-light optimizer you can drop into your project

When to avoid

• you need a TensorFlow implementation (only promised, not shipped)
• you expect a silver bullet that eliminates all hyperparameter tuning
• you need actively developed features or support for newer PyTorch versions

Facets

library · maturity maintenance

machine-learning deep-learning machine-learning deep-learning python optimizer pytorch sgd adam learning-rate iclr-2019 gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="Luolc/AdaBound")