Meta entrusts PyTorch to an open source foundation

Meta entrusts PyTorch to an open source foundation

The development of the deep learning library has been entrusted to a brand new independent organization: the PyTorch Foundation. A facility hosted by the Linux Foundation.

[Mise à jour le lundi 12 septembre 2022 à 16h] Meta is so far orchestrating the development of the PyTorch open source deep learning library from which it originated and announcing the creation of a foundation that will now lead the project: the PyTorch Foundation. An independent organization under the auspices of the Linux Foundation. The steering committee will include representatives from Amazon Web Services, AMD, Google Cloud, Meta, Microsoft Azure and Nvidia. It will also be open to new members in the future. “In the long term, the contributors to the framework will benefit from the governance, the diversity of the management team and the additional investments made by the new partners of the PyTorch foundation,” said Santosh Janardhan, vice president of Infrastructure at Meta. adding: “The foundation is committed to adhering to four principles: remain open source, maintain a neutral branding strategy, remain fair and forge a recognized technical identity.”

What is PyTorch?

PyTorch is an AI library, developed by Meta, written in Python to participate in deep learning (or deep learning) and in the development of artificial neural networks. From various variables, it can be used to perform gradient calculations or to use multidimensional tables obtained using tensors.

PyTorch has been available as open source since 2016 with the modified BSD license. In 2018, the library was merged by Meta with Caffe2, the deep learning infrastructure tailored for implementations and capable of supporting learning algorithms that record up to tens of billions of parameters.

To download PyTorch, go to the official project page on GitHub and select the file from the “Code” tab. Extensions are also accessible here.

For Windows, Linux and macOS operating systems, PyTorch can be installed from the Conda package manager. The associated installation file can be downloaded from You must then follow the procedure and then complete it with this line of code: conda install -c pytorch pytorch.

You can also install PyTorch from the pip environment. Depending on your system, it may ask you to insert the line of code pip installs pytorch.

Why is PyTorch used?

PyTorch is first of all easy to learn. This makes it easily accessible and productive. Subsequently, it is a dynamic architecture. Result: Neural networks are scalable during their learning phase. PyTorch allows you to add new nodes, change the connections between them, even from one level to another. Downstream, where model graph processing is directly supported by the PyTorch execution engine, the framework can easily be used with various third party debugging environments such as PyCharm or ipdb.

Finally, Facebook’s infrastructure supports declarative data parallelism. A technique that allows processing to be split into multiple mini-batches to perform the learning phase in parallel on standardized GPU server clusters.

PyTorch vs TensorFlow: what’s the difference?

PyTorch uses a dynamic neural network, while TensorFlow is based on a static graph architecture. This is one of the main differences between the two deep learning libraries. Result: With PyTorch it is much easier to develop neural network structures natively as the learning phase progresses.

What is DataLoader in PyTorch?

PyTorch comes with two primitives. The first,, is designed to create mini datasets from training datasets. The second,, allows the use of preloaded datasets.

PyTorch is more than a library, it is a real framework that integrates several modules. Among these are TorchServe, which is nothing more than a tool for managing previously trained neural network implementations, or TorchElastic, which allows you to start fault-tolerant PyTorch processing on Kubernetes clusters.

Leave a Reply

Your email address will not be published. Required fields are marked *