Cannot import index_max module

[zeal-up]

My environment:
cuda10.0
pytroch1.2/1.4
python3.6

I have successfully compiled index_max and the module is installed in 'xx/xx/python3.6/site-packages/index_max-0.0.0-py3.6-linux-x86_64.egg/index_max.cpython-36m-x86_64-linux-gnu.so'.
However, when I try to import index_max, it raised up an error:

python3.6/site-packages/index_max-0.0.0-py3.6-linux-x86_64.egg/index_max.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZTIN3c1021AutogradMetaInterfaceE

I have also another question, I am not familiar with Cuda, the index_max seems used to achieve max_pooling along the feature dimension for all point besides one som cluster. Do I understand right? How much a cuda implemention speed up compare with a python implementation?

[lijx10]

Please make sure that the cuda environment is configured properly. For example, proper environment variables. For example, you may consider adding these lines to ~/.bashrc

# cuda-10.1
export CUDA_HOME=/usr/local/cuda-10.1
export PATH=/usr/local/cuda-10.1/bin:${PATH}
export LD_LIBRARY_PATH=/usr/local/cuda-10.1/lib64:${LD_LIBRARY_PATH}

Run nvcc -V to confirm that cuda is working properly.

After making sure that cuda and pytorch is installed properly, delete folder dist, build, and the file index_max.egg-info, then run python setup install to install the index_max module again.

Finally, please make sure that the index_max folder is not in the python path. There may be linking problems.

[lijx10]

For the speed. An alternative implementation is to run for loop in pytorch, to get a max-pooling for each node. In the case that node number is small, e.g., <10, this practice should be fine. However, in most cases, the node number is like 64, 128, which means python for loop will be quite slow.