Failed to register operator torchvision::_new_empty_tensor_op

[h6197627]

Trying to import
import torchvision
results in exception:

Traceback (most recent call last):
File "", line 1, in
File "/usr/local/lib/python3.6/dist-packages/torchvision/init.py", line 3, in
from torchvision import models
File "/usr/local/lib/python3.6/dist-packages/torchvision/models/init.py", line 12, in
from . import detection
File "/usr/local/lib/python3.6/dist-packages/torchvision/models/detection/init.py", line 1, in
from .faster_rcnn import *
File "/usr/local/lib/python3.6/dist-packages/torchvision/models/detection/faster_rcnn.py", line 7, in
from torchvision.ops import misc as misc_nn_ops
File "/usr/local/lib/python3.6/dist-packages/torchvision/ops/init.py", line 13, in
_register_custom_op()
File "/usr/local/lib/python3.6/dist-packages/torchvision/ops/_register_onnx_ops.py", line 49, in _register_custom_op
register_custom_op_symbolic('torchvision::_new_empty_tensor_op', new_empty_tensor_op, _onnx_opset_version)
File "/usr/local/lib/python3.6/dist-packages/torch/onnx/init.py", line 200, in register_custom_op_symbolic
return utils.register_custom_op_symbolic(symbolic_name, symbolic_fn, opset_version)
File "/usr/local/lib/python3.6/dist-packages/torch/onnx/utils.py", line 793, in register_custom_op_symbolic
.format(symbolic_name))
RuntimeError: Failed to register operator torchvision::_new_empty_tensor_op. The symbolic name must match the format Domain::Name, and sould start with a letter and contain only alphanumerical characters

Torchvision version (built from source): 0.6.0a0+28b7f8a
PyTorch version (built from source): 1.4.0a0+7f73f1d

[fmassa]

Hi,

You need a newer version of PyTorch to be able to compile latest torchvision (I had myself the same issue yesterday and an upgrade of PyTorch fixed it)

[saareliad]

Hi,

You need a newer version of PyTorch to be able to compile latest torchvision (I had myself the same issue yesterday and an upgrade of PyTorch fixed it)

By newer you mean the nightly version?
He already uses 1.4 (And I got the same problem with v1.4.0)

Which torchvision should we use with pytorch v1.4.0?

[joefutrelle]

I'm using PyTorch 1.4.0. During building torchvision (0.6.0a0+e2573a7) from source I see this message:

Searching for torch==1.4.0
Best match: torch 1.4.0
Adding torch 1.4.0 to easy-install.pth file

So it looks to me like torchvision depends on 1.4.0. Is that dependency not correct?

[joefutrelle]

5ac5ab9 seems to have introduced _new_empty_tensor_op which is the identifier that torch is complaining about. Rolling back torchvision to the parent commit, 40c99ea, causes the import to succeed, and I was successful in running an image recognition demo.

[fmassa]

@saareliad newer means a nightly version of PyTorch and torchvision. Using 1.4.0 and 0.5.0 is not enough.

[fmassa]

If you want to use torchvision nightly, you need to use PyTorch nightly.

If you want to use a PyTorch from a release (say 1.4.0), you should use a torchvision from a matching release as well (in this case, 0.5.0)

[mknw]

I'm encountering the same problem on a linux server, nightly installed with pip:
pip install --pre torch torchvision -f https://download.pytorch.org/whl/nightly/cu101/torch_nightly.html

Solved by switching to python 3.7 (was 3.8 before).

edit: but gpu isn't detected as available (it is with upstream pytorch + torchvision).

[fmassa]

@mknw can you open a new issue with the problem you are facing? From the description it seems that Python 3.8 with CUDa 10.1 is not working for our pip binaries, is that correct?

Can you also post when you tried downloading the binaries, and which version they are?

[cuongpxu]

Hi,
I'm facing this issue on the lastest PyTorch version (1.5.0).

Traceback (most recent call last): File "main.py", line 14, in <module> from torchvision import transforms File "/home/alan/anaconda3/envs/cuongpx/lib/python3.6/site-packages/torchvision/__init__.py", line 3, in <module> from torchvision import models File "/home/alan/anaconda3/envs/cuongpx/lib/python3.6/site-packages/torchvision/models/__init__.py", line 12, in <module> from . import detection File "/home/alan/anaconda3/envs/cuongpx/lib/python3.6/site-packages/torchvision/models/detection/__init__.py", line 1, in <module> from .faster_rcnn import * File "/home/alan/anaconda3/envs/cuongpx/lib/python3.6/site-packages/torchvision/models/detection/faster_rcnn.py", line 7, in <module> from torchvision.ops import misc as misc_nn_ops File "/home/alan/anaconda3/envs/cuongpx/lib/python3.6/site-packages/torchvision/ops/__init__.py", line 13, in <module> _register_custom_op() File "/home/alan/anaconda3/envs/cuongpx/lib/python3.6/site-packages/torchvision/ops/_register_onnx_ops.py", line 51, in _register_custom_op register_custom_op_symbolic('torchvision::_new_empty_tensor_op', new_empty_tensor_op, _onnx_opset_version) File "/home/alan/anaconda3/lib/python3.6/site-packages/torch/onnx/__init__.py", line 195, in register_custom_op_symbolic return utils.register_custom_op_symbolic(symbolic_name, symbolic_fn, opset_version) File "/home/alan/anaconda3/lib/python3.6/site-packages/torch/onnx/utils.py", line 740, in register_custom_op_symbolic .format(symbolic_name)) RuntimeError: Failed to register operator torchvision::_new_empty_tensor_op. The symbolic name must match the format Domain::Name, and sould start with a letter and contain only alphanumerical characters
Environment
OS: Ubuntu 16.04.3 LTS (Xenial Xerus)"
PyTorch version: 1.5.0
torchvision version: 0.6.0
Python version: 3.6.9
CUDA version: 10.1

I have tried to degrade torchvision version to 0.5.0 but it raises another issue. How can I solve this issue?

Thanks

[qhaas]

Seeing this error with Torchvision 0.6.0 (built from source) on ppc64le (CUDA 10.0) and x86-64 (CUDA 10.2) with PyTorch 1.4.0 (built from source), python 3.6.9 on Ubuntu 18.04. I'll attempt to downgrade TorchVision to 0.5.0 since ppc64le is currently broken with PyTorch 1.5.0.

UPDATE: Downgrading cleared this issue in x86-64, testing ppc64le
UPDATE2: Downgrade cleared it in ppc64le as well