Sergey - stock.adobe.com
A new PyTorch update from a collaboration between Facebook and AWS adds experimental features and support for more programming languages to the open source machine learning framework, to make it easier for developers to build machine learning models.
The PyTorch 1.5 update, released April 21, introduces TorchServe, a new model serving library, and TorchElastic, a new Kubernetes controller, as experimental applications for PyTorch. PyTorch 1.5 primarily aims to make it easier for developers and engineers to train models cheaper, more efficiently and with fewer errors.
PyTorch, primarily developed by Facebook's AI Research lab, enables engineers and developers to more easily develop and train deep learning models in Python, similar to TensorFlow. PyTorch has become dramatically more popular since its initial release in 2016 as enterprises increasingly move models out of production and into productization.
TorchServe, a PyTorch model serving library, enables users to train models using a distributed cloud infrastructure more easily and inexpensively, said Kashyap Kompella, CEO and chief analyst of the AI industry analyst firm RPA2AI Research.
Kashyap KompellaCEO and chief analyst, RPA2AI Research
The library addresses the need that many enterprises have for an end-to-end machine learning pipeline, Kompella said.
"Building machine models is really only the tip of the iceberg, while a lot more effort goes into deploying and maintaining the models once deployed," he said. "TorchServe makes the lives of the data scientists and machine learning engineers using PyTorch a bit easier now as they don't have to jump through several hoops to move from research to production."
While the collaboration between AWS and Facebook adds new functionality to PyTorch, TensorFlow, a competing open source product primarily developed by Google, already has it, Kompella noted.
Still, he said, the new release is "good for the machine learning ecosystem as this is a sign of growth and maturity of the industry."
TorchElastic, meanwhile, enables developers to more efficiently execute builds in a fault-tolerant and elastic way, which makes them more resilient.
"PyTorch has been gaining momentum in terms of adoption, and the new release PyTorch 1.5, with the C++ API and components such as TorchServe and TorchElastic, help sustain this momentum," Kompella said.