search

NVIDIA / vid2vid

Pytorch implementation of our method for high-resolution (e.g. 2048x1024) photorealistic video-to-video translation.
Other
8.57k stars 1.2k forks source link

pytorch version compatible for training phase, undefined symbol when importing resample2d_cuda #134

Open WeisiX opened 4 years ago

WeisiX commented 4 years ago

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

WeisiX commented 4 years ago

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.