Facebook's open-source deep learning framework is PyTorch. Facebook's artificial-intelligence researchers created PyTorch to make machine-learning development more flexible and intuitive, especially for research and experimentation.
PyTorch is built around tensors, multidimensional arrays that can be processed efficiently on CPUs and GPUs. Its major feature is automatic differentiation, which calculates the gradients needed to train neural networks. Its eager, Python-friendly execution model also made it easier for researchers to inspect and change models while they were running.
Facebook released PyTorch as open source in 2016, and the project quickly became popular in universities, laboratories, and industry. It is used for computer vision, natural-language processing, recommendation systems, robotics, and generative AI. The framework is closely associated with the wider Python machine-learning ecosystem, including tools for datasets, distributed training, and model deployment.
A common mix-up is TensorFlow, another major deep-learning framework originally developed by Google. Caffe and Keras are also associated with neural-network development, but neither is Facebook's framework. PyTorch is now maintained under the Linux Foundation after Meta transferred governance of the project.