pytorch version compatible for training phase, undefined symbol when importing resample2d_cuda
WeisiX opened this issue · comments
Hi!
I cloned the latest code and followed the instructions for installation.
I upgraded Pytorch from 0.4 to 1.0 in order to successfully compile the flownet2. I could also run the test with cityscape data under Pytorch 1.0. However, when I switch to the training phase of a single GPU, I receive the error of:
"segmentation error (core dumped)".
I figured that it was a pytorch/cuda version mismatch so I downgraded my Pytorch back to 0.4.0 and run the training. Unfortunately, this time I got the error when importing resample2d_cuda:
"resample2d_cuda.cpython-35m-x86_64-linux-gnu.so: undefined symbol: _ZN2at19UndefinedTensorImpl10_singletonE".
It seems like it is still a version mismatch problem but usually people could fix that when they are on pytorch 0.4.0 or 0.4.1, both of which I tried but with the same error coming out.
Many thanks for any clue to solve this problem!!
FYI, I am on:
ubuntu 16.04
cuda 9.0
Pytorch 0.4/1.0
Solved by going back to the pytorch 0.4 version of the code (before pytorch 1.0 support is added) with the pytorch 0.4.0 version of the flownet2 compiled.