Fails to convert PyTorch Mask R-CNN Resnet50 FPN model

[JJonahJson]

I'm trying to convert this model:

model = torchvision.models.detection.maskrcnn_resnet50_fpn(pretrained=True)

This is the code of the conversion:

model = torch.load('/path/to/model/trained_model.pt')
model.eval()
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
#dummy input
img = cv2.imread('/path/to/images/something.jpg')
transform = T.Compose([T.ToTensor()])
img = transform(img)
img = img.to(device).unsqueeze_(0)

model = torch2trt(model, [img])

These are the warning I get:

Warning: Encountered known unsupported method torch.Tensor.unbind
Warning: Encountered known unsupported method torch.Tensor.__iter__
Warning: Encountered known unsupported method torch.Tensor.unbind
Warning: Encountered known unsupported method torch.Tensor.__iter__
Warning: Encountered known unsupported method torch.Tensor.is_floating_point
Warning: Encountered known unsupported method torch.as_tensor
Warning: Encountered known unsupported method torch.as_tensor

Then I get this output:

AttributeError                            Traceback (most recent call last)
<ipython-input-2-9e433d9c0209> in <module>
      7 img = transform(img)
      8 img = img.to(device, non_blocking=True).unsqueeze_(0)
----> 9 model = torch2trt(model, [img])
     10 
     11 # Set to evaluation mode

/opt/conda/lib/python3.7/site-packages/torch2trt-0.2.0-py3.7-linux-x86_64.egg/torch2trt/torch2trt.py in torch2trt(module, inputs, input_names, output_names, log_level, max_batch_size, fp16_mode, max_workspace_size, strict_type_constraints, keep_network, int8_mode, int8_calib_dataset, int8_calib_algorithm, int8_calib_batch_size, use_onnx, **kwargs)
    540             ctx.add_inputs(inputs, input_names)
    541 
--> 542             outputs = module(*inputs)
    543 
    544             if not isinstance(outputs, tuple) and not isinstance(outputs, list):

/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
    887             result = self._slow_forward(*input, **kwargs)
    888         else:
--> 889             result = self.forward(*input, **kwargs)
    890         for hook in itertools.chain(
    891                 _global_forward_hooks.values(),

/opt/conda/lib/python3.7/site-packages/torchvision/models/detection/generalized_rcnn.py in forward(self, images, targets)
     76             original_image_sizes.append((val[0], val[1]))
     77 
---> 78         images, targets = self.transform(images, targets)
     79 
     80         # Check for degenerate boxes

/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
    887             result = self._slow_forward(*input, **kwargs)
    888         else:
--> 889             result = self.forward(*input, **kwargs)
    890         for hook in itertools.chain(
    891                 _global_forward_hooks.values(),

/opt/conda/lib/python3.7/site-packages/torchvision/models/detection/transform.py in forward(self, images, targets)
    101                 raise ValueError("images is expected to be a list of 3d tensors "
    102                                  "of shape [C, H, W], got {}".format(image.shape))
--> 103             image = self.normalize(image)
    104             image, target_index = self.resize(image, target_index)
    105             images[i] = image

/opt/conda/lib/python3.7/site-packages/torchvision/models/detection/transform.py in normalize(self, image)
    126         mean = torch.as_tensor(self.image_mean, dtype=dtype, device=device)
    127         std = torch.as_tensor(self.image_std, dtype=dtype, device=device)
--> 128         return (image - mean[:, None, None]) / std[:, None, None]
    129 
    130     def torch_choice(self, k):

/opt/conda/lib/python3.7/site-packages/torch2trt-0.2.0-py3.7-linux-x86_64.egg/torch2trt/torch2trt.py in wrapper(*args, **kwargs)
    287 
    288             #             print('%s' % (converter.__name__,))
--> 289             converter["converter"](ctx)
    290 
    291             # convert to None so conversion will fail for unsupported layers

/opt/conda/lib/python3.7/site-packages/torch2trt-0.2.0-py3.7-linux-x86_64.egg/torch2trt/converters/getitem.py in convert_tensor_getitem(ctx)
     28     output = ctx.method_return
     29 
---> 30     input_trt = input._trt
     31 
     32     # Step 1 - Replace ellipsis with expanded slices

AttributeError: 'Tensor' object has no attribute '_trt'

I don't understand why I get this error from torch2trt/converters/getitem.py nor what I have to do to make it work, if I understand correctly the converter for the getitem function already exists, right?

Is there a quick fix to solve this problem?

[leaf918]

Is there going to support the detection-like models in torchvision?
There is some tests for torchvision's classification and segmentation.

[thebruce87m]

@leaf918 Did you find a fix for this? I see the same on Keypoint R-CNN here: #761