KerasHub provides a curated collection of pre-trained models for Keras 3, simplifying the integration of advanced AI capabilities into applications. It targets developers and researchers working with text, image, and audio data, offering a flexible and fast way to leverage state-of-the-art architectures for tasks like generation and classification.

How It Works

KerasHub offers Keras 3 implementations of popular model architectures, directly usable as Layer and Model components. This approach ensures seamless integration with the existing Keras ecosystem. A key advantage is its multi-framework support, allowing models to run on JAX, TensorFlow, and PyTorch from a single definition, facilitating cross-platform development and deployment.

Quick Start & Requirements

• Install with: pip install --upgrade keras-hub
• For NLP models: pip install --upgrade keras-hub[nlp]
• Backend configuration via KERAS_BACKEND environment variable (e.g., export KERAS_BACKEND=jax).
• Requires Keras 3 installation. TensorFlow is pulled in for tf.data preprocessing, though training can occur on any backend.
• See: Getting Started, KerasHub Home

Highlighted Details

• Supports JAX, TensorFlow, and PyTorch backends from a single model definition.
• Enables fine-tuning on GPUs and TPUs with built-in PEFT techniques.
• Offers model and data parallel training for large-scale fine-tuning.
• Provides access to pre-trained checkpoints on Kaggle Models.

Maintenance & Community

• Developed by the Keras team.
• Follows Semantic Versioning with a plan for backward compatibility.
• Pre-release development (0.y.z) may include breaking changes.
• See: Contributing Guide, Roadmap

Licensing & Compatibility

• KerasHub components are provided under an unspecified license.
• Underlying models (BART, BLOOM, GPT-2, Llama, etc.) are provided by third parties and subject to separate licenses. Compatibility for commercial use depends on these underlying licenses.

Limitations & Caveats

The project is in pre-release development (0.y.z), meaning APIs are not yet stable and backward compatibility is not guaranteed. The specific license for KerasHub itself is not detailed in the README, and users must verify the licenses of the third-party underlying models for commercial use.