Open Neural Network Exchange (ONNX) is an open file format for machine learning models, introduced in 2017 by Microsoft and Facebook (now Meta) and now hosted by the LF AI & Data Foundation. A model trained in PyTorch, for example, can be exported to ONNX and run with other tools or runtimes such as ONNX Runtime. This makes it easier to switch between frameworks and deploy on different hardware.
PyTorch is an open-source framework for machine learning and deep learning in the Python programming language. It was originally developed at Meta (then Facebook) and has been stewarded since 2022 by the PyTorch Foundation under the umbrella of the Linux Foundation. PyTorch is widely used in research and industry because it lets developers build models flexibly, train them on GPUs and test them step by step; many well-known language and image models are built with it.
TensorFlow is an open-source machine learning framework developed by Google and released in 2015. It is used to build, train and deploy neural networks in applications, on servers or on mobile devices. Keras provides an accessible high-level interface; typical uses include image classification, forecasting and recommendation systems.
Apache MXNet was an open-source deep learning framework that could train models efficiently across multiple GPUs and machines, with interfaces for Python, C++, R, Scala and other languages. Amazon Web Services backed it as its preferred framework for some years. The Apache Software Foundation retired the project in September 2023 and moved it to the Apache Attic by February 2024; it is no longer developed. Existing models are usually migrated to other frameworks.
Scikit-learn is an open-source Python library for classical machine learning. It provides consistent tools for classification, regression, clustering, dimensionality reduction, data preprocessing and model evaluation, such as cross-validation. Built on NumPy and SciPy, it is widely used in education, research and practice thanks to its simple, uniform interface. For deep neural networks, frameworks such as PyTorch or TensorFlow are usually chosen instead.
