keras-team/keras-hub
Pretrained model hub for Keras 3.
What it solves
KerasHub provides a simple and flexible way to access and use pretrained models across different machine learning frameworks. It eliminates the need to rewrite model architectures and training loops when switching between JAX, TensorFlow, and PyTorch.
How it works
KerasHub extends the core Keras 3 API, providing model architectures as Layer and Model implementations. It pairs these architectures with a collection of pretrained checkpoints available on Kaggle Models. This allows users to load a model from a preset, load data using tf.data for preprocessing, and run inference or fine-tuning on any of the supported backends.
Who it’s for
Developers and researchers who use Keras and want to integrate pretrained models for text, image, and audio data into their applications or research projects across multiple frameworks.
Highlights
- Multi-framework support: Works with JAX, TensorFlow, and PyTorch from a single definition.
- Multimodal capabilities: Supports models for text, image, and audio data for tasks like generation and classification.
- Cross-platform training: Fine-tuning can be performed on GPUs and TPUs out of the box.
- Efficient fine-tuning: Includes built-in Parameter-Efficient Fine-Tuning (PEFT) techniques and support for model and data parallel training at scale.