DirectML, a backend for PyTorch that can be used to train machine learning models, has been released as the first preview by the Microsoft Windows AI team. PyTorch now supports rapid machine learning (ML) training on any DirectX12 GPU and WSL, allowing for new possibilities in mixed reality computing.

This preview package was released as part of a partnership between the Microsoft AI team and the PyTorch framework that provides scoped support for CNNs (convolutional neural networks). With this new device, called “DML,” Tensor primitives and Direct ML APIs are called through by introducing a minimal overhead when calling into operators; they will function similarly as other existing backends already exist.

DirectML is a GPU machine learning library that can be used as the backend for PyTorch-DirectML, allowing direct training of models on DirectX12 GPUs and WSL platforms (Windows Subsystem for Linux). This hardware-accelerated learning experience for PyTorch was developed by Microsoft in cooperation with AMD, Intel, and NVIDIA.