Keras

Keras is a human-centered deep learning API that emphasizes code clarity, fast debugging, and deployability across multiple backends. It provides high-level abstractions for layers and models (Sequential, Functional, and subclassing) while remaining extensible for advanced research. Keras supports multi-backend execution (JAX, TensorFlow, PyTorch), distributed and mixed-precision training, and includes KerasHub with pretrained model implementations and checkpoints.

• Multi-backend support — Use JAX, TensorFlow, or PyTorch backends to train and run models without changing Keras code; to get started, choose a backend and follow the backend configuration guides.
• Model APIs (Sequential, Functional, Subclassing) — Build simple to complex architectures using intuitive APIs; begin with Sequential for linear stacks and the Functional API for arbitrary graphs.
• Training & callbacks — Train with model.fit, evaluate, and use callback utilities (Checkpointing, EarlyStopping, TensorBoard) to manage experiments.
• KerasHub & pretrained models — Access implementations and checkpoints for language, vision, and generative models for fine-tuning or inference.
• Data loading and utilities — Built-in data loaders, preprocessing layers, and utilities streamline dataset handling and experiment management.
• Production & deployment features — Model saving, serialization, mixed-precision, and distribution primitives support research-to-production workflows.

To get started, install Keras from source/package, follow the API guides, and run the provided examples and quickstarts to train or load pretrained models.