model conversion failed in mo_pytorch.py for fasterrcnn_resnet50_fpn about openvino_contrib

[CameronVetter]

I'm unsure of the best way to share the model in a github issue, so I've uploaded it here: https://1drv.ms/u/s!AofkugmEOKADh_QZvQU0bWQm2kfOcQ?e=7GxEh9

This is the exact model I'm using to reproduce this locally.

[dkurt]

Hi! Can you please share a model so we can reproduce issue locally? Even with randomly initialized weights.

[dkurt]

@CameronVetter, Currently module does not support TorchScript models conversion, so can you please try this code?

import torch
import torchvision.models as models

model_jit = torch.jit.load('github356_model.pt', map_location=torch.device("cpu"))
model = models.detection.mask_rcnn.maskrcnn_resnet50_fpn(num_classes=2)
model.load_state_dict(model_jit.state_dict())
model.eval()

# Convert to OpenVINO IR
import mo_pytorch
mo_pytorch.convert(model, input_shape=[1, 3, 800, 800], model_name='Mask_RCNN',
                   scale=255,
                   reverse_input_channels=True)

I used PyTorch 1.9.1 and Torchvision 0.10.0

[CameronVetter]

Thank you, that example worked well. I will just skip TorchScript in the future, I did not realize that was unsupported.