core dumped issue

[wuhuikai]

core dumped issue

[wuhuikai]

I cannot figure it out without further details. Maybe try to run it again : )

[danxuhk]

the error comes at this line: x = self.pretrained.conv1(x) in the base.py
If I replace self.pretrained.conv1 with a normal self.conv, it passes. What do you think could be the issue?

[wuhuikai]

Can you provide the details of your environment and base.py before and after modification?

[danxuhk]

I did not change any environment configuration and anything. I use Encnet with JPU and ResNet50. Just used the training sh script to change it...

[danxuhk]

My pytorch version is 1.2.0 and CUDA 9.0.

[wuhuikai]

Please use PyTorch==1.0.x

[danxuhk]

I changed to Pytorch==1.0.0. The problem is still there. I did not change anything on the base.py...
When I change BatchNorm2d with torch.nn.BatchNorm2d, it passes. The issue is due to the batch norm layer. Any idea how could solve this?

[wuhuikai]

Have you reinstalled encoding after changing to 1.0.0? If you did, then I think everything is the same as my environment :(
Ubuntu 16.04, Python 3.5, PyTorch 1.0.0, Cuda 9.0

[danxuhk]

I reinstalled encoding after changing to 1.0.0. It is a segmentation fault caused by using encoding.nn.BatchNorm2D.

[wuhuikai]

I have no idea why there's a segmentation fault since you have the same configurations as mine :(
But for a reference, there's a solution in the original Encoding repo.

[wuhuikai]

@danxuhk The code in branch latest is now ready, which uses official SyncBatchNorm and can run on all OS with PyTorch>=1.1.0.