keras-cv

Introduction to KerasCV

KerasCV is a versatile library that provides a collection of modular computer vision components compatible with TensorFlow, JAX, or PyTorch. Built on the foundations of Keras 3, it allows models, layers, metrics, and callbacks to be seamlessly trained and transferred across different frameworks without requiring complex migrations.

KerasCV doesn't replace the core Keras API but rather extends it horizontally by providing specialized components dedicated to computer vision tasks. These components are officially recognized and maintained by the Keras team, ensuring the same level of reliability and backward compatibility as Keras itself.

Key Features and Applications

KerasCV offers a wide range of APIs that assist with common computer vision tasks, including:

• Data Augmentation: Effortlessly enhance datasets to improve model training.
• Classification: Identify and categorize objects within images.
• Object Detection: Locate and recognize objects in various scenes.
• Segmentation: Differentiate and label regions within an image.
• Image Generation: Create new images using state-of-the-art generative models.

Computer vision engineers can quickly build robust and innovative training and inference pipelines using KerasCV for these tasks.

Installation Guidelines

KerasCV supports both Keras 2 and Keras 3, with a preference for Keras 3 for new users to leverage compatibility with various backends like JAX, TensorFlow, and PyTorch.

• Keras 2 Installation: Use the command pip install --upgrade keras-cv tensorflow to install the latest release.

• Keras 3 Installation: Use nightly packages via pip install --upgrade keras-cv-nightly tf-nightly, or ensure Keras 3 is installed after KerasCV using stable versions:

  pip install --upgrade keras-cv tensorflow
  pip install --upgrade keras
  

Configuring Your Backend

Once Keras 3 is installed, KerasCV can be set up to work with your preferred backend. This can be achieved by configuring the KERAS_BACKEND environment variable to JAX, TensorFlow, or PyTorch before importing any Keras libraries.

Example setup for JAX:

export KERAS_BACKEND=jax

Quickstart Guide

KerasCV simplifies setting up a data preprocessing pipeline and model training:

• Utilize various layers for data augmentation such as RandomFlip, RandAugment, and CutMix.
• Leverage pre-trained models, like the EfficientNetV2 backbone, for tasks like image classification.
• Compile models with appropriate loss functions, optimizers, and metrics.
• Train the model using datasets from sources like TensorFlow Datasets (e.g., rock-paper-scissors dataset).

Example code snippets provide an illustration of setting up data augmentation, processing datasets, and training models with KerasCV components.

Contribution and Community

KerasCV encourages contributions from the community to help enhance its features, fix issues, and expand its capabilities. Contributors are guided on how to submit new model architectures or tasks, aligning them with the project roadmap.

Pretrained Models and Custom Ops

KerasCV offers pretrained model weights for numerous models, facilitating easier implementation and application in various projects. For more advanced usage, such as 3D Object Detection layers, it supports custom TensorFlow operations, though these require installation from source.

Citing KerasCV

For use in academic or research contexts, users are encouraged to cite KerasCV using the provided BibTeX entry.

In conclusion, KerasCV is a powerful asset for developers and researchers in the field of computer vision, providing robust, flexible tools to accelerate the development and deployment of advanced visual models across multiple frameworks.