Keras 3: Train JAX Models with model.fit, Keras Hub, and Keras Recommenders

Keras Ecosystem and JAX Integration

• 💡 The Keras ecosystem provides multiple entry points for using JAX, from simple model.fit loops to advanced model parallel distribution.
• 🛠️ Keras offers a comprehensive toolbox for deep learning, including model.compile, model.fit, and model.summary for building and training JAX models.
• ✅ Keras supports various optimizers, losses, metrics, and callbacks, all compatible with the model.fit training loop abstraction.

Advanced Training with Keras

• 🚀 Keras follows the principle of progressive disclosure of complexity, allowing users to customize model.fit for advanced training loops.
• 💻 Multi-device distribution can be enabled using Keras's distribution APIs, including DeviceMesh and LayoutMap for sharding model weights.
• ⚙️ For ultimate control, users can override core training steps, such as compute_loss and updates, for custom loss functions and tailored training loops.

Backend Agnosticism and Keras Hub

• 🔄 Keras APIs are backend-agnostic, supporting JAX, TensorFlow, and PyTorch, with the ability to switch backends easily by setting an environmental variable.
• 🧩 You can mix and match training frameworks and data pipelines, for example, using a TF.data pipeline with a JAX backend.
• 📚 Keras Hub offers a rich collection of over 211 model variants based on 44+ architectures like Gemma, Llama, and Stable Diffusion, ready for use with JAX, TensorFlow, or PyTorch.
• 🤝 Keras Hub integrates seamlessly with platforms like Kaggle and Hugging Face, providing access to presets, checkpoints, and community guides.

Fine-tuning and Model Components

• 🎯 Keras Hub models can be fine-tuned using standard model.fit or advanced techniques like LoRA and QLoRA for parameter-efficient training.
• 🧩 Task models in Keras Hub consist of a backbone, a pre-processor, and a task head, simplifying the pipeline from raw data to predictions.
• 🔧 Components like backbones, pre-processors, and tokenizers can be instantiated using the from_preset method for easy integration of pre-trained weights and configurations.

Keras Recommenders

• 🎯 Keras Recommenders is a new library for building state-of-the-art recommendation systems with JAX, TensorFlow, or PyTorch.
• 📈 It supports a two-phase approach (retrieval and ranking) and multitask models, with specialized layers, losses, and metrics for recommender models.
• 🧠 Sequential recommender models, like GRU4Rec, can be built using a two-tower architecture with Keras Recommenders and trained with model.fit.
• ☁️ The distributed embedding layer in Keras Recommenders leverages specialized hardware like Google's TPUs for efficient handling of large embedding tables across multiple devices.