Windows.md

Using Intel® AI Reference Models on Windows Systems

Prerequisites for running on bare metal

Basic requirements for running all TensorFlow models on Windows include:

• Install Python 3.9+ 64-bit release for Windows, and add it to your system %PATH% environment variable.
• Download and Install Microsoft Visual C++ 2022 Redistributable.
• MSYS2. If MSYS2 is installed to C:\msys64, add C:\msys64\usr\bin to your %PATH% environment variable. Then, using cmd.exe, run:

   pacman -S git patch unzip

TensorFlow models

• Install tensorflow-intel
• Set MSYS64_BASH=C:\msys64\usr\bin\bash.exe environment variable to your system. The path may change based on where have you installed MSYS2 on our system.
• Install the common models dependencies:
  • python-tk
  • libsm6
  • libxext6
  • requests

Individual models may have additional dependencies that need to be installed before running it. Please follow the instructions in each model documentation.

The following list of models are tested on Windows, please check each model instructions from the Model Documentation column based on the available precisions.

Note that on Windows systems all of the system cores will be used. For users of Windows desktop/laptops, it is strongly encouraged to instead use the batch file provided here to open a Windows command prompt pre-configured with optimized settings to achieve high AI workload performance on Intel hardware (e.g. TigerLake & AlderLake) for image recognition models.

Use Case	Model	Mode	Model Documentation
Image Recognition	ResNet 50v1.5	Inference	Int8 FP32
Language Modeling	BERT	Inference	FP32
Language Translation	BERT	Inference	FP32
Recommendation	Wide & Deep	Inference	FP32

PyTorch models

Intel® Extension for PyTorch is currently not supported on Windows.

Install PyTorch

pip install torch torchvision

The following list of models are tested on Windows, please check each model instructions from the Model Documentation column based on the available precisions.

Use Case	Model	Mode	Model Documentation
Image Recognition	ResNet 50	Inference	FP32 BFloat16
Object Detection	Mask R-CNN	Inference	FP32