ImportError: ./modules/roi_align/_ext/roi_align/_roi_align.so: undefined symbol: state about rt-mdnet HOT 21 OPEN

@jimiolaniyan $ python build.py -> $ make -C src/cuda -> make -C src/cuda clean

from rt-mdnet.

any progress?

from rt-mdnet.

None. I am researching the SiamRPN recently.

from rt-mdnet.

okk, thanks a lot.

from rt-mdnet.

Thank you for your interest.
I think... it may be a problem related to python version or pytorch version. In my memorial, ffi package is not applicable to python 3.0?. Could you let me know what version do you use?

from rt-mdnet.

In fact, I found out that it was because that the .so file had not been compiled. It should follow the steps in makefile .

from rt-mdnet.

Thank you for your interest.
I think... it may be a problem related to python version or pytorch version. In my memorial, ffi package is not applicable to python 3.0?. Could you let me know what version do you use?

Thanks for your reply. My python version is 3.6.8.

from rt-mdnet.

In fact, I found out that it was because that the .so file had not been compiled. It should follow the steps in makefile .

So, should I compile the .so file manually ?

from rt-mdnet.

I have also run into this issue. I am using python 2.7.12.

from rt-mdnet.

I have also run into this issue. I am using python 2.7.12.

So, this is a very interesting problem. @IlchaeJung
@IlchaeJung Can you reimplement those C++ code in Python? I think that it's a very significant work for all of us. Thanks.

from rt-mdnet.

In fact, I found out that it was because that the .so file had not been compiled. It should follow the steps in makefile .

So, should I compile the .so file manually ?

yes, I think so.

from rt-mdnet.

@siyueyu Kindly share how to compile the .so file if you were able to? Thank you

from rt-mdnet.

@jimiolaniyan Have you followed the steps in ./modules/roi_align/makefile. After I do these steps first, this problem solved in my case.

from rt-mdnet.

@siyueyu I was able to get past that error. However, i quickly ran into this issue:

cudaCheckError() failed : no kernel image is available for execution on the device

Looks like this is because I have a Tesla K80 gpu. I have tried to recompile the .so with -arch=sm_37 (for K80) but this did not work.

Let me know if you have encountered this and if you have a solution.

from rt-mdnet.

@jimiolaniyan I hadn't met that problem. I am very sorry. But I think that may be caused by cuda version or something about the environment.

from rt-mdnet.

@siyueyu Yes I tried cuda 8.0 but then had a different issue:

cudaCheckError() failed: invalid device function

The same issue is reported here. What is your current set up?

Also @IlchaeJung I have recompiled the kernel files with arch=sm_37 (for Tesla K80) but the problem persists. Do you have any guidance on this?

from rt-mdnet.

@jimiolaniyan what about the version of pytorch？

from rt-mdnet.

@siyueyu I have used versions 0.3.0, 0.4.0 and 0.4.1. Which version did you use?

from rt-mdnet.

Dear all,

Sorry, I forgot to mention that recompile of RoIAlign may be need to some other GPU users.
As @siyueyu mentioned, we can recompile RoIAlign layer.
I used Titan Xp, Titan X maxwell with cuda 8.0 / 9.0, python2.7 and pytorch < 0.4.1.
I think that users with these settings will be okay after recompiling. However, unfortunately, I do not have an experience with other type of GPU. Moreover, I think that the GPU having memory lower than 10GB should reduce the batch-size for pre-training while it needs at least 2GB~3GB for test. Could you check this memory capacity please?

To @fzh0917, I really want to extend our code to C++ version. However, it needs some time because I do not familiar with linking c++ to python and the other project currently I do. But I keep in mind to do that. Thank you.

from rt-mdnet.

@IlchaeJung OK. I quite understand your troubles. Actually, my main purpose is to understand the RT-MDNet tracker's core view through reading the corresponding code implementation.

from rt-mdnet.

I was able to get it to work with a P100 GPU on Google cloud. I used:

PyTorch: 0.3.0
Python: 2.7
Cuda: 8.0

from rt-mdnet.