---
id: pytorch-ranger
version: "0.1.1"
license: Apache
license_treatment: permissive
maintenance: abandoned
---

# pytorch-ranger — Ranger - a synergistic optimizer using RAdam (Rectified Adam) and LookAhead in one codebase

License: permissive · Maintenance: abandoned · Downloads: 160.0K/mo

## What it is and what it does

Ranger is a PyTorch optimizer that merges two optimization techniques—RAdam for rectified adaptive momentum and LookAhead for weight averaging—into a single implementation. It was designed to improve training stability and generalization in deep learning models. The package provides three optimizer variants: Ranger (the main version), RangerVA (using calibrated anisotropic adaptive learning rates), and RangerQH, each with different parameter configurations.

You use it like any PyTorch optimizer: instantiate it with your model's parameters and pass it to your training loop. The package depends only on torch and installs as a lightweight pure-Python wheel. However, the project is no longer maintained—the last commit was 2021-12-01—so it receives no updates, bug fixes, or compatibility patches for newer PyTorch releases.

Use it for:

- Training convolutional neural networks where you want better convergence than standard Adam, especially with Mish activation functions.
- Fine-tuning large models where weight averaging via LookAhead helps stabilize learning and reduce overfitting.
- Experimenting with alternative optimization strategies for image classification tasks on platforms like FastAI.
- Comparing optimizer performance in research or benchmarking studies of deep learning training dynamics.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Ranger is a PyTorch optimizer that combines RAdam (Rectified Adam) and LookAhead techniques to improve training stability and convergence in deep learning models.

No—install only if you are reproducing historical research or maintaining legacy code. The package is abandoned (last commit 2021-12-01) and receives no maintenance, bug fixes, or compatibility updates for modern PyTorch versions. For new projects, use actively maintained optimizers like torch.optim.AdamW or community-supported alternatives. The permissive Apache license poses no legal barrier, but the lack of maintenance makes it a liability.

## Install

pip install pytorch-ranger
uv add pytorch-ranger
poetry add pytorch-ranger

## Installing pytorch-ranger

Before you install: Low install friction with a pure Python wheel. However, the package is abandoned—last commit was 2021-12-01. No active maintenance or updates.

License in practice: Licensed under Apache (permissive), so you can use it freely in commercial and open-source projects without restrictive obligations.

Quickstart:

pip install pytorch_ranger

from pytorch_ranger import Ranger
model = ...  # your PyTorch model
optimizer = Ranger(model.parameters())

Requires torch as a runtime dependency; PyTorch must be installed separately.

Verify before relying:

- Whether the optimizer variants (RangerVA, RangerQH) are stable or still experimental given the beta status.
- Current compatibility with recent PyTorch versions, given the last commit was in 2021.
- Whether memory issues mentioned in the description (save/load stranding weights) are fully resolved in version 0.1.1.

## Package facts

- License: Apache (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: abandoned
- Downloads: 160.0K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

pytorch optimizer radam lookahead, deep learning optimizer ranger, adaptive learning rate optimizer, neural network training optimizer, radam lookahead combination, pytorch ranger optimizer, optimizer, abandoned

[View on SkillFed](https://skillfed.io/packages/pytorch-ranger) · [View on PyPI](https://pypi.org/project/pytorch-ranger/)