A few months ago, we released the first preview of PyTorch-DirectML: a hardware accelerated backend for training PyTorch models on any DirectX12 GPU on Windows and the Windows Subsystem for Linux (WSL). Today, we are releasing the Second Preview with significant performance improvements and greater coverage for computer vision models.

PyTorch-DirectML now works with Python versions 3.6, 3.7, and 3.8, and includes support for GPU device selection. On the DirectML GitHub, you’ll find a new public Operator Roadmap indicating current and planned future support for PyTorch operators. We also published updated samples on GitHub including a new YOLO object detection sample.

Try it out today! Download from PyPI and leave feedback on the DirectML GitHub.

Interested in learning more about DirectML for training and deploying ML models? Check out our new YouTube video!