Windows AI team debuts 'PyTorch-DirectML' for training machine learning models on any GPU
A framework that doesn't require you to have an Nvidia GPU
October 26, 2021
A framework that doesn't require you to have an Nvidia GPU
Microsoft's Windows AI team has announced the first preview of 'PyTorch-DirectML.'
The software allows for DirectML to be used as a backend to PyTorch for training machine learning models, with the full release allowing for accelerated machine learning training for PyTorch on any DirectX12 GPU on Windows and the Windows Subsystem for Linux (WSL).
This allows for GPU acceleration for common machine learning tasks across a broad range of supported hardware and drivers, including those from AMD, Intel, Qualcomm, and Nvidia.
Put your AMD GPUs to use
Microsoft introduced DirectML earlier this year as a low-level API for machine learning, spun out of its work on video games. PyTorch-DirectML builds upon that, integrating open source machine learning library PyTorch – which is primarily developed by Facebook's AI Research lab.
"We co-engineered with AMD, Intel, and Nvidia to enable this hardware-accelerated training experience for PyTorch," Microsoft said. The early preview supports convolutional neural networks, ahead of a wider rollout.
The project will allow AI developers to make better use of non-Nvidia GPUs, with the US chipmaker dominating the machine learning market thanks to its CUDA platform, APIs, and software set – which only runs on Nvidia hardware.
The PyTorch-DirectML package only requires changing one line of code in an existing script, Microsoft said. The latest version can be downloaded here, with a tutorial for training SqueezeNet and ResNet available on GitHub.
The Microsoft-owned software repository can also be used for reporting issues with PyTorch-DirectML here.
About the Author
You May Also Like