Maskrcnn-benchmark: torchvision ImportError
ImportError: /home/jiang/anaconda3/envs/maskrcnn_benchmark/lib/python3.6/site-packages/torchvision/_C.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN2at19NonVariableTypeMode10is_enabledEv
jiangkeng123
All 6 comments
@jiangkeng123 before you build, try to export CUDA_HOME=/cuda_dir, and build again.
githubgsq
on 24 May 2019
The same thing happens for me. Whenever I open python and attempt to import torch vision, this error pops up.
AlexG1105
on 24 May 2019
Ok, I think you have to downgrade your torch vision version. Do
conda uninstall pytorch
conda uninstall pytorch-nightly
conda uninstall torchvision
then redo this:
conda install -c pytorch pytorch-nightly torchvision=0.2.2 cudatoolkit=10.0
AlexG1105
on 24 May 2019
What happens is that you compiled with one version of PyTorch, and when you updated torchvision it reinstalled a new version of PyTorch.
Another solution is just to recompile maskrcnn_benchmark after deleting the build folder.
fmassa
on 24 May 2019
Hello, I have the same error, and finally how do you solve this problem?
1032697379
on 25 May 2019
@jiangkeng123 before you build, try to export CUDA_HOME=/cuda_dir, and build again.
can you give more detail about the cuda_dir?such as ?
Liusandian
on 28 Jun 2020
Related issues
adityaarun1
路
3Comments
CF2220160244
路
3Comments
IenLong
路
4Comments
BelhalK
路
4Comments
YuShen1116
路
4Comments
Most helpful comment
Ok, I think you have to downgrade your torch vision version. Do
conda uninstall pytorchconda uninstall pytorch-nightlyconda uninstall torchvisionthen redo this:
conda install -c pytorch pytorch-nightly torchvision=0.2.2 cudatoolkit=10.0