undefined symbol: _Z23LeakyRelu_Backward_CUDAN2at6TensorES0_f about awesome-semantic-segmentation-pytorch

[Yuxiang1995]

I follow this issue zhanghang1989/PyTorch-Encoding#127 (comment) and upgrade the torch-encoding ,but it doesn't work

[dekura]

I follow this issue zhanghang1989/PyTorch-Encoding#127 (comment) and upgrade the torch-encoding ,but it doesn't work

@kernel-Peanut i meet this problem too. Have you find how to solve it? Thank you so much.

[dekura]

Hi, @kernel-Peanut @Tramac
I solved this problem by changing the files in /core/nn/sync_bn/lib/gpu

first

change the syncbn_kernel.cu as follow

before

#include <vector>
#include <torch/extension.h>
#include <ATen/ATen.h>
#include <ATen/cuda/CUDAContext.h>

after

#include <vector>
#include <torch/serialize/tensor.h>
#include <ATen/ATen.h>
#include <ATen/cuda/CUDAContext.h>

second

and do follow the issue here.

Summary

this error may be caused by the rapid update of pytorch.
i forked this program and you can refer to my code here .

[Tramac]

Good job, thank you very much!

[Yuxiang1995]

@dekura It doesn't work for me .
I change #include <torch/extension.h> to #include <torch/serialize/tensor.h> and rename the torch/torch.h to torch/extension.h , but it doesn't work.

[dekura]

@kernel-Peanut you con refer to my code here and see which torch.h should be changed and which should not.
or you can just change the core/nn/sync_bn/lib/gpu folder with mine~
Have a try~

[dekura]

@kernel-Peanut i tried again and i found that you must delete all the codes about LeakyRelu_Backward and LeakyRelu_Forward in core/nn/sync_bn/lib/gpu

[Yuxiang1995]

@dekura
Hey! The bug disappear after following your method. Thanks a lot!

[Yuxiang1995]

@Tramac
I think this issue can be closed.

[Tramac]

@kernel-Peanut @dekura
Hi, the bug undefined symbol: _Z23LeakyRelu_Backward_CUDAN2at6TensorES0_f in syncbn has been fixed with no remove leakyrelu.
Thank you for your job!