onnx: Segmentation Fault on ONNX 1.1.2
I am using onnx 1.1.2 and Caffe2, and when I do the following, I get a segmentation fault:
from caffe2.python import core
import onnx
Swapping the order of the above imports doesn’t change the error. backtrace: https://gist.github.com/pooyadavoodi/53c6d0ce381f9a41e8fbd79bb555176c
I built onnx from branch rel-1.1.2 and it works. So the problem is only the pip package I think.
About this issue
- Original URL
- State: closed
- Created 6 years ago
- Comments: 23 (16 by maintainers)
Commits related to this issue
- Exclude ONNX 1.1.2 ONNX 1.1.2 produces a segfault, see: https://github.com/onnx/onnx/issues/875 — committed to NervanaSystems/ngraph-onnx by postrational 6 years ago
- PyUP Initial Update (#84) * Pin numpy to latest version 1.14.3 * Update onnx from 1.1.1 to 1.1.2 * Pin cachetools to latest version 2.1.0 * Pin pytest to latest version 3.5.1 * Pin tox to... — committed to NervanaSystems/ngraph-onnx by pyup-bot 6 years ago
This is not related to pytorch or caffe2 but the combination of an upstream issue in pybind11 https://github.com/pybind/pybind11/issues/1262 and the way onnx binary packages were built. We don’t have a way to fix a released version (pypi doesn’t allow resubmit another packages for a released version). Installing onnx via
pip install --no-binary onnx onnxwould force pip to use source package to fix this issue, and @raymondxyang is going to help us on updating the release pipeline (see discussions here https://github.com/onnx/onnx/pull/1183) so for future release this issue should go away.