Vision: Mask-rcnn torchscript doesn't load in c++: Unknown builtin op: torchvision::_new_empty_tensor_op.

Created on 5 Dec 2019 · 6Comments · Source: pytorch/vision

I was trying to load a torchscript version of mask-rcnn into c++ but unfortunately this failed. Here is the error message / stacktrace I got:

./run_model ../torchscript.pt 
terminate called after throwing an instance of 'torch::jit::script::ErrorReport'
  what():  
Unknown builtin op: torchvision::_new_empty_tensor_op.
Could not find any similar ops to torchvision::_new_empty_tensor_op. This op may not exist or may not be currently supported in TorchScript.
:
  File "/home/user/.local/lib/python3.6/site-packages/torchvision/ops/new_empty_tensor.py", line 16
        output (Tensor)
    """
    return torch.ops.torchvision._new_empty_tensor_op(x, shape)
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
Serialized   File "code/__torch__/torchvision/ops/new_empty_tensor.py", line 3
def _new_empty_tensor(x: Tensor,
    shape: List[int]) -> Tensor:
  _0 = ops.torchvision._new_empty_tensor_op(x, shape)
       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
  return _0
'_new_empty_tensor' is being compiled since it was called from 'interpolate'
Serialized   File "code/__torch__/torchvision/ops/misc.py", line 123
    align_corners: Optional[bool]=None) -> Tensor:
  _31 = __torch__.torchvision.ops.misc._output_size
  _32 = __torch__.torchvision.ops.new_empty_tensor._new_empty_tensor
  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
  _33 = uninitialized(Tensor)
  if torch.gt(torch.numel(input), 0):
'interpolate' is being compiled since it was called from 'GeneralizedRCNNTransform.resize'
Serialized   File "code/__torch__/torchvision/models/detection/transform.py", line 78
    target: Optional[Dict[str, Tensor]]) -> Tuple[Tensor, Optional[Dict[str, Tensor]]]:
    _18 = __torch__.torchvision.models.detection.transform.resize_boxes
    _19 = __torch__.torchvision.ops.misc.interpolate
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
    _20 = __torch__.torchvision.models.detection.transform.resize_keypoints
    _21 = uninitialized(Tuple[Tensor, Optional[Dict[str, Tensor]]])
'GeneralizedRCNNTransform.resize' is being compiled since it was called from 'GeneralizedRCNNTransform.forward'
  File "/home/user/.local/lib/python3.6/site-packages/torchvision/models/detection/transform.py", line 47
                                 "of shape [C, H, W], got {}".format(image.shape))
            image = self.normalize(image)
            image, target_index = self.resize(image, target_index)
                                  ~~~~~~~~~~~ <--- HERE
            images[i] = image
            if targets is not None and target_index is not None:
Serialized   File "code/__torch__/torchvision/models/detection/transform.py", line 28
        pass
      image0 = (self).normalize(image, )
      _2 = (self).resize(image0, target_index, )
                                 ~~~~~~~~~~~~ <--- HERE
      image1, target_index0, = _2
      _3 = torch._set_item(images0, i, image1)

Aborted (core dumped)

All the details:
I tried it with the latest versions of pytorch/libtorch

import torch, torchvision
torch.__version__
'1.4.0.dev20191205+cpu'
torchvision.__version__
'0.5.0.dev20191204+cpu'

The export (and import) in python works:

model = torchvision.models.detection.maskrcnn_resnet50_fpn(pretrained=True)
model.eval()
script_model = torch.jit.script(model)
script_model.save("torchscript.pt")

After which I followed the basic steps of importing a torchscript module into c++ from https://pytorch.org/tutorials/advanced/cpp_export.html .

Also here I've downloaded the latest version of libtorch:
```file=build-hash
/pytorch /pytorch ~/project
3fd4c696d9b869b03c47dae35ba908c07e6fa1b5

```file=build-version
1.4.0.dev20191205+cpu

The c++ code is a just a very basic loading operation:

#include <torch/script.h> // One-stop header.

#include <iostream>
#include <memory>
#include <chrono>

int main(int argc, const char* argv[]) {
  if (argc != 2) {
    std::cerr << "usage: example-app <path-to-exported-script-module> \n";
    return -1;
  }
  torch::jit::script::Module module;
  module = torch::jit::load(argv[1]);

  std::cout << "ok\n";
}

For completion, the cmakelist (since the c++ version got changed to 14 https://github.com/pytorch/vision/pull/1635 ):

cmake_minimum_required(VERSION 3.0 FATAL_ERROR)

project(custom_ops)

find_package(Torch REQUIRED)

add_executable(run_model run_model.cpp)
target_link_libraries(run_model "${TORCH_LIBRARIES}")
set_property(TARGET run_model PROPERTY CXX_STANDARD 14)

enhancement help wanted c++ frontend documentation ops

Source

gslotman

Most helpful comment

Thanks! I after some (cmake) issues I managed to get the loading of the networks running with the code in https://github.com/pytorch/vision/pull/1407#issuecomment-562096191 . Not yet tried forward passing or anything

I had to fix some headers though: https://github.com/pytorch/vision/pull/1644

gslotman on 6 Dec 2019

👍2

All 6 comments

Hi,

For this particular error, I think you can follow the comments in https://github.com/pytorch/vision/pull/1407#issuecomment-562096191 to fix this issue