mo_pytorch

PyTorch extension for Model Optimizer

This module let you convert PyTorch models directly to OpenVINO IR without using ONNX

Usage

1. Clone repository

   git clone --depth 1 https://github.com/openvinotoolkit/openvino_contrib

2. Setup environment

   source /opt/intel/openvino_<VERSION>/bin/setupvars.sh
   export PYTHONPATH=openvino_contrib/modules/mo_pytorch:$PYTHONPATH

3. Convert PyTorch model to OpenVINO IR

   import torchvision.models as models
   
   # Create model
   model = models.alexnet(pretrained=True)
   
   # Convert to OpenVINO IR
   import mo_pytorch
   mo_pytorch.convert(model, input_shape=[1, 3, 227, 227], model_name='alexnet')

Supported networks

• torchvision.models.alexnet
• torchvision.models.resnet18
• torchvision.models.squeezenet1_1
• torchvision.models.inception_v3
• torchvision.models.segmentation.deeplabv3_resnet50
• torchvision.models.detection.mask_rcnn.maskrcnn_resnet50_fpn
• Detectron2 RetinaNet
• 3D-ResUNet
• ruGPT3Medium
• USRNet