Ranger is an open-source deep learning optimizer for PyTorch that combines three proven techniques โ RAdam, LookAhead and Gradient Centralization โ into a single drop-in optimizer. Maintained by Less Wright on GitHub, where it has collected around 1,200 stars, Ranger aims to give practitioners faster, more stable convergence without extensive learning-rate tuning.
Optimizer choice can make or break a training run. RAdam rectifies the variance of adaptive learning rates early in training, removing the need for a careful warm-up in many cases. LookAhead keeps a slow set of weights that periodically pulls the fast training weights back, improving stability and final accuracy. Gradient Centralization, added in later versions, normalizes gradient vectors to improve generalization. Ranger fuses all three: you swap it in wherever you would use Adam or SGD in a PyTorch training loop and get the combined behaviour. The repository includes usage notes, parameter guidance and results reported by community members across image classification and other benchmarks. The codebase was last updated in late 2023 but remains a widely referenced implementation of the technique.
Ranger gained a following through community benchmarks and its inclusion in popular training libraries and fastai-style workflows. Its author documents the reasoning behind each component, making the repository a useful read even for practitioners who ultimately train with a different optimizer.
Ranger is free and open source under the Apache-2.0 license. There are no paid tiers or hosted services; you install the code from GitHub and use it in your own training pipelines.
Machine learning researchers and PyTorch practitioners who want a strong general-purpose optimizer that is forgiving of hyperparameters and simple to drop into existing training code.