Welcome to Albumentations Documentation!

On this page

• Getting Started
• Learning the Basics
• Advanced Topics
• Other Resources

Albumentations is the most obvious default augmentation library for most computer vision users: fast pipelines, a broad transform catalog, and target-aware support for classification, segmentation, object detection, pose, OCR, medical, remote-sensing, video, and 3D workloads. Use it for training augmentation, test-time augmentation, validation diagnostics, preprocessing experiments, and any image/video/volume augmentation policy that needs to be correct, inspectable, and easy to integrate.

Albumentations works cleanly with PyTorch, TensorFlow/Keras, JAX, and custom training stacks. In GPU training, it commonly runs before tensors enter the framework-specific model step, for example inside PyTorch Dataset or DataLoader workers; you keep your framework for models, tensors, training loops, and deployment.

Install with pip install albumentationsx. Starting with 2.3.2, the maintained package uses AGPL-3.0-only or separately agreed commercial terms. The archived albumentations==2.0.8 package remains available under MIT. See the License Guide before choosing a package and license path.

This documentation will guide you through installing the library, understanding its core concepts, applying it to real training pipelines, and optimizing augmentation for correctness and throughput.

Getting Started

• Introduction: Learn what data augmentation is and why it's important.
• Installation: Set up Albumentations in your environment.

Learning the Basics

• Core Concepts: Understand the fundamental building blocks: Transforms, Pipelines (Compose), Targets (image, mask, bboxes, keypoints), and Probabilities.
• Basic Usage Guides: Find practical examples for common computer vision tasks:
  • Image Classification
  • Semantic Segmentation
  • Instance Segmentation
  • Object Detection (Bounding Boxes)
  • Oriented Bounding Boxes (OBB)
  • Keypoint Augmentation
  • Video Augmentation
  • Volumetric (3D) Augmentation
  • Framework Integrations: Use Albumentations with PyTorch, TensorFlow/Keras, JAX, and custom training loops.
• Choosing Augmentations: A detailed guide on selecting effective augmentation strategies for model generalization.
• Performance Tuning: Tips for optimizing your augmentation pipeline speed.

Advanced Topics

• Advanced Guides: Explore more complex features:
  • Using Additional Targets
  • Creating Custom Transforms
  • Reproducibility in Albumentations
  • Replay and Applied-Parameter Debugging
  • Serialization
  • Test Time Augmentation (TTA)

Other Resources

• Comparisons:
  • Albumentations vs PIL/Pillow: Compare Albumentations with Pillow for RGB image editing, arbitrary-channel arrays, and target-aware augmentation.
  • Albumentations vs torchvision: Why Albumentations is the default PyTorch Dataset / DataLoader augmentation layer, and where torchvision still fits.
  • Albumentations vs Kornia: Compare the default CPU/DataLoader path with Kornia's narrower differentiable and GPU tensor augmentation use cases.
  • Albumentations vs DALI: Compare Albumentations with NVIDIA DALI for graph-style GPU decode and preprocessing pipelines.
  • Legacy combined torchvision/Kornia mapping: Compatibility landing page for the old combined guide.
• AlbumentationsX vs Legacy Albumentations: Practical Benefits: Technical comparison for teams evaluating the legacy albumentations package against AlbumentationsX.
• Security, Safety, and Release Integrity: Vulnerability reporting, release verification, trusted publishing, and operational continuity.
• Frequently Asked Questions (FAQ): Find answers to common questions.
• Benchmarks: Performance comparison results and benchmark methodology.
• Supported Targets by Transform: Check which transforms work with images, masks, bounding boxes, keypoints, etc.
• API Reference
• GitHub Repository: Active development
• Examples Repository: Many practical examples.

We hope this documentation helps you leverage the full power of Albumentations!