Maskrcnn-benchmark: ImportError: ../maskrcnn_benchmark/_C.cpython-35m-x86_64-linux-gnu.so: undefined symbol: _ZN2at15UndefinedTensor10_singletonE

[ypxie]

❓ Questions and Help

pytorch 1.0
CUDA 9.0
PYTHON 3.5
Ubuntu 16.04

[fmassa]

This is a duplicate of https://github.com/facebookresearch/maskrcnn-benchmark/issues/119
You are probably not using PyTorch 1.0, and there is an old PyTorch version laying around in your machine that is conflicting.
Try doing in your interpreter

import torch
print(torch.__version__)

and it will probably show that you are not using 1.0

[Dandi007]

I have the same issue too.
cuda 9.2
pytorch-nightly 1.0.0.dev20181211 py3.7_cuda9.0.176_cudnn7.4.1_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch

when I do print torch.__version__,it outpus 1.0.0 so my torch version is correct.

but I still got this issue.

[Dandi007]

this is my env info

Collecting environment information...
PyTorch version: 1.0.0.dev20181211
Is debug build: No
CUDA used to build PyTorch: 9.0.176

OS: Ubuntu 17.10
GCC version: (Ubuntu 7.2.0-8ubuntu3.2) 7.2.0
CMake version: version 3.9.1

Python version: 3.7
Is CUDA available: Yes
CUDA runtime version: 9.2.88
GPU models and configuration: GPU 0: GeForce MX150
Nvidia driver version: 410.48
cuDNN version: Could not collect

Versions of relevant libraries:
[pip] Could not collect
[conda] blas 1.0 mkl https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
[conda] mkl 2018.0.3 1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
[conda] mkl_fft 1.0.4 py37h4414c95_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
[conda] mkl_random 1.0.1 py37h4414c95_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
[conda] pytorch-nightly 1.0.0.dev20181211 py3.7_cuda9.0.176_cudnn7.4.1_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch
[conda] torchvision 0.2.1

[Dandi007]

I can run web.py successfully，
python webcam.py --min-image-size 800

but when I run nms.py,I got
ImportError: /media/disk1/PycharmProjects/maskrcnn-benchmark/maskrcnn_benchmark/_C.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN2at19UndefinedTensorImpl10_singletonE

[fmassa]

what is nms.py?

[Dandi007]

I just run it to run this line of code:

from maskrcnn_benchmark import _C

it raise error when trying to import _C

[fmassa]

This means that maskrcnn-benchmark was not properly installed.

Can you try rm -rf build/ from maskrcnn-benchmark folder, and try installing it again according to the instructions in INSTALL.md?

[niaoyu]

This means that maskrcnn-benchmark was not properly installed.

Can you try rm -rf build/ from maskrcnn-benchmark folder, and try installing it again according to the instructions in INSTALL.md?

Oh... Forget to delete all the files. It works for me

[qushan326]

actually this maybe not some wrong with installation, as it's said in pytorch's tutorial:
https://pytorch.org/tutorials/advanced/cpp_extension.html

[sdw95927]

Usually reinstalling packages and some important depencencies (e.g. gcc) solves most of the problems lol

[futureisatyourhand]

conda install -c pytorch pytorch-nightly torchvision cudatoolkit=10.0
issue:
from maskrcnn_benchmark import _C ImportError: maskrcnn_benchmark/_C.cpython-36m-x86_64-linux-gnu.so: undefined symbol: __cudaRegisterFatBinaryEnd
or from torchvision import _C.
this do run by:
conda uninstall pytorch
conda uninstall pytorch-nightly
conda uninstall torch
codna uninstall torchvision
then:
conda install -c pytorch pytorch-nightly torchvision cudatoolkit=10.0
pip install torchvision==0.2.2
successfully!!

[i-chaochen]

conda install -c pytorch pytorch-nightly torchvision cudatoolkit=10.0
issue:
from maskrcnn_benchmark import _C ImportError: maskrcnn_benchmark/_C.cpython-36m-x86_64-linux-gnu.so: undefined symbol: __cudaRegisterFatBinaryEnd
or from torchvision import _C.
this do run by:
conda uninstall pytorch
conda uninstall pytorch-nightly
conda uninstall torch
codna uninstall torchvision
then:
conda install -c pytorch pytorch-nightly torchvision cudatoolkit=10.0
pip install torchvision==0.2.2
successfully!!

how do you success with cuda 10.0?
I tried python webcam.py --min-image-size 800

and I got
ImportError: libcudart.so.9.0 cannot open shared object file: No such file or directory

[futureisatyourhand]

the version of cuda on my server is 10.0, if you use cuda9.0, I suggest you try cuda9.2, the version can run well, since I find cuda9.0 often doesn't work very well.

[i-chaochen]

the version of cuda on my server is 10.0, if you use cuda9.0, I suggest you try cuda9.2, the version can run well, since I find cuda9.0 often doesn't work very well.

My machine is cuda 10.0 as well, but when I followed your steps uninstall everything and conda install again.

I met the cuda library problem. Does it mean I need to uninstall cuda10.0 and reinstall cuda9.2?
ImportError: libcudart.so.9.0 cannot open shared object file: No such file or directory

[milkytipo]

My machine is cuda 10.0 as well, but when I followed your steps uninstall everything and conda install again.

I met the cuda library problem. Does it mean I need to uninstall cuda10.0 and reinstall cuda9.2?
ImportError: libcudart.so.9.0 cannot open shared object file: No such file or directory

you just need to reinstall it again as the instruction, I met the the same problem and reinstall it, it worked now

[futureisatyourhand]

the version of cuda on my server is 10.0, if you use cuda9.0, I suggest you try cuda9.2, the version can run well, since I find cuda9.0 often doesn't work very well.

My machine is cuda 10.0 as well, but when I followed your steps uninstall everything and conda install again.

I met the cuda library problem. Does it mean I need to uninstall cuda10.0 and reinstall cuda9.2?
ImportError: libcudart.so.9.0 cannot open shared object file: No such file or directory

you can try to reinstall "conda install -c pytorch pytorch-nightly torchvision cudatoolkit=9.2" or reinstall by my instruction.

[zimenglan-sysu-512]

hi @fmassa
it seems that torch_nightly-1.2.0 is not compatible with torchvision-0.3