Keras

Keras is an open-source software library that provides a Python interface for building and training neural networks. It is designed to enable fast experimentation with deep learning models through a user-friendly, modular API that can be extended and deployed across platforms. Keras emphasizes readability and simplicity while supporting common deep learning workflows, from prototyping to production.

Keras was created by François Chollet and released in 2015 as an independent project. Originally it provided a high-level API that could run on top of several backends, including TensorFlow, Theano, and CNTK. With the release of TensorFlow 2.0, Keras became the recommended high-level API for TensorFlow, and the integration is exposed as tf.keras. In recent years, the Keras project has continued to evolve with versions that emphasize backend-agnostic design, while TensorFlow users commonly interface with Keras through tf.keras.

The library offers two primary APIs for model construction: a simple Sequential model for linear stacks of layers, and the Functional API for complex architectures with multiple inputs, outputs, and shared layers. Core concepts include layers, models, loss functions, optimizers, metrics, callbacks, and utilities for data preprocessing, serialization, and deployment. Keras provides a growing set of built-in layers (convolutional, recurrent, normalization, regularization, etc.) and integrates with TensorFlow tools for training, evaluation, and distributed computing.

Keras is used for rapid prototyping and teaching as well as for production systems. It is released under the Apache 2.0 license and has a large, active community contributing tutorials, examples, and extensions. It remains a central component of the TensorFlow ecosystem, with extensive documentation and ecosystem compatibility.