Keras (software)

Keras (software)
Name	Keras
Author	François Chollet
Developer	Google, François Chollet
Initial release	March 2015
Latest release	2025 (example)
Programming language	Python
Operating system	Cross-platform
Genre	Machine learning library
License	Apache License 2.0

Keras (software) is an open-source high-level neural networks API written in Python that facilitates rapid prototyping, experimentation, and deployment of deep learning models. Designed by François Chollet and later developed in conjunction with Google teams, it provides user-friendly abstractions built atop computation engines to support tasks in image recognition, natural language processing, and time series analysis. Keras emphasizes modularity, minimalism, and extensibility, enabling researchers and engineers from organizations such as OpenAI, DeepMind, and academic laboratories at Stanford University and MIT to iterate quickly while interoperating with larger ecosystems like TensorFlow and cloud platforms from Amazon Web Services and Microsoft Azure.

History

Keras originated in March 2015 when François Chollet, a software engineer at Google, released a lightweight neural network library to simplify model construction compared with lower-level frameworks. Early adoption accelerated after integrations with Theano and TensorFlow, and Keras became a reference API for many practitioners, attracting contributions from engineers affiliated with Facebook AI Research and research groups at University of Toronto and University College London. In 2017, TensorFlow adopted Keras as its official high-level API, prompting deeper collaboration with teams at Google Brain and cross-project efforts involving contributors associated with NumPy and SciPy. Over subsequent years, Keras evolved through major redesigns to support eager execution paradigms championed by PyTorch and responsive runtime environments such as Tensor Processing Unit accelerators and heterogeneous infrastructures used by NVIDIA and Intel.

Architecture and Design

Keras implements a layered architecture that separates model specification, optimization, and execution. The central abstraction is a sequential or functional model graph inspired by computation graphs used in libraries developed at University of California, Berkeley and Carnegie Mellon University. Layers in Keras are modular components resembling building blocks from earlier projects at Google Research and conventions from software engineered by teams at Apple and Facebook. The design supports an imperative programming style alongside graph-based execution, enabling interoperability with execution engines like TensorFlow and legacy engines such as Theano. Internally, Keras maps high-level operations to backend primitives provided by accelerator vendors including NVIDIA and runtime teams at Google Cloud to balance portability and performance.

Features and Components

Keras provides a comprehensive set of features that streamline model building and evaluation. Core components include Layer classes, Model containers (both Sequential and Functional APIs), and Callback mechanisms inspired by training utilities used in projects at OpenAI and DeepMind. The library ships with prebuilt layers for convolutional networks and recurrent networks that mirror architectures from landmark works at University of Oxford and Google DeepMind, along with utilities for data preprocessing compatible with datasets curated by ImageNet and COCO consortia. Keras also exposes metrics, loss functions, and optimizers reflecting algorithms popularized by researchers at Courant Institute and University of Montreal. Extensible features include model serialization, visualization hooks used with tools from TensorBoard and third-party platforms like those developed by Weights & Biases and Comet ML, and model conversion utilities interoperable with formats championed by ONNX and deployment toolchains from TensorFlow Lite.

Supported Backends and Platforms

Keras was designed to be backend-agnostic, historically supporting engines such as TensorFlow, Theano, and CNTK from Microsoft. Modern Keras tightly integrates with TensorFlow as the primary execution engine while maintaining compatibility layers for runtimes developed by NVIDIA and standards from Open Neural Network Exchange. Deployments span devices and services provided by Android ecosystems, iOS frameworks, edge computing hardware like Raspberry Pi boards, and cloud infrastructures from Google Cloud Platform, Amazon Web Services, and Microsoft Azure. Hardware acceleration support leverages drivers and libraries from NVIDIA (CUDA, cuDNN) and accelerator stacks from Google (TPU runtime), enabling scalable training on clusters orchestrated with technologies such as Kubernetes.

Development, Adoption, and Community

The development of Keras is coordinated through repositories maintained by engineers from Google and contributors distributed across academia and industry, including members affiliated with Facebook, OpenAI, and universities like Princeton University and Harvard University. The project benefits from an active community on platforms such as GitHub, discussion forums popularized by Stack Overflow, and events including workshops at NeurIPS and tutorials at ICML and CVPR. Adoption spans enterprises in sectors represented by corporations like Uber and Airbnb, research labs at Allen Institute for AI, and independent developers leveraging model zoos curated by groups such as Hugging Face. Community governance involves contributors holding maintainer roles, coordinating with standards bodies and working groups associated with projects like Apache Software Foundation initiatives.

Licensing and Distribution

Keras is distributed under the Apache License 2.0, a permissive license that enables commercial and academic use, contributions, and redistribution. Source code and releases are published via code hosting services maintained by organizations such as GitHub and packaged for installation through repositories curated by Python Software Foundation tools including pip and package indexes influenced by Conda ecosystems. Binary distributions and cloud deployment images are provided by partners and vendors like Google Cloud Platform and Amazon Web Services, facilitating integration into continuous delivery pipelines used by teams employing orchestration systems such as Docker and Kubernetes.

Category:Machine learning libraries