PyTorch is an open-source package to support deep learning, originally developed at Facebook. Like TensorFlow, PyTorch provides at its core a tensor data structure, which is very similar to the numpy array but which can used on GPUs and other hardware accelerators, and which also supports automatic differentiation needed for backpropagation in neural network training.

The design philosophy of the PyTorch API is a bit different than those presented by TensorFlow and Keras. Stated design principles are "Usability Over Performance", "Simple Over Easy", and "Python First with Best In Class Language Interoperability". The emphasis on "Simple Over Easy" stresses the importance of having explicit control over many of the internals of the system.

PyTorch version 2 was released in late 2022. While many programs developed to use an earlier PyTorch v1.x version will continue to run under v2, there may be circumstances where that is not so, in which case you should consult documentation about relevant changes to the API or to version compatibilities.

In addition to the core PyTorch system itself, there is a rich PyTorch Ecosystem, a collection of packages and tools that build upon the base provided by PyTorch, built and supported by a developer community spread across a variety of application domains. For many applications, you might want to use one or more of those ecosystem packages rather than building up analogous functionality on your own using PyTorch.