Segmentation fault in SplineConv about pytorch_geometric HOT 13 CLOSED

A segmentation fault may be a sign of an installation error in pytorch_spline_conv. What PyTorch version are you using?

Please confirm that

python setup.py install
python setup.py test

in pytorch_spline_conv works and report again.

from pytorch_geometric.

0.4.1

Install runs fine but it's the same problem in test.

from pytorch_geometric.

Are you sure the kernel did rebuild? You can force this with

rm -rf build/ && python setup.py install

from pytorch_geometric.

still, the test fails with this:

test/test_backward.py ....Segmentation fault (core dumped)

from pytorch_geometric.

I had the same issue with all the mnist examples.

I ran python setup.py test in pytorch_geometric and got:
test/transforms/test_local_cartesian.py Segmentation fault

from pytorch_geometric.

I've got the same issue w/ the qm9 example: Segmentation fault (core dumped)

from pytorch_geometric.

Could this be resolved by anyone?

from pytorch_geometric.

I have that all tests run successfully, but get problems when running my own model: It always runs for some epochs, but then i get problems:

Segmentation fault (core dumped)
or

free(): invalid size
Aborted (core dumped)
or

free(): invalid next size (normal)
Aborted (core dumped)
or

double free or corruption (out)
Aborted (core dumped)
or

Traceback (most recent call last):
File "first_test.py", line 101, in
loss = train(epoch)
File "first_test.py", line 83, in train
loss.backward()
File "/home//anaconda3/lib/python3.7/site-packages/torch/tensor.py", line 102, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph)
File "/home//anaconda3/lib/python3.7/site-packages/torch/autograd/init.py", line 90, in backward
allow_unreachable=True) # allow_unreachable flag
RuntimeError: could not compute gradients for some functions

or
corrupted size vs. prev_size
Aborted (core dumped)

or
munmap_chunk(): invalid pointer
Aborted (core dumped)

I think those are all the different error messages I get without changing any of my code.

I was not able to track down the error further than knowing it is somewhere in the backward step if I use SplineConv. When doing with torch.no_grad for the SplineConv layer I don't have issues.

p.s. running Ubuntu 18.04 with CPU.

from pytorch_geometric.

I think I know the issue. When feeding pseudo values which are not in [0,1], there is a disaster. Maybe there should be a test somewhere warning the user?

from pytorch_geometric.

Yes, this is not allowed! I could add an assertion, but this would result in unnecessary computations. Any idea?

from pytorch_geometric.

Mhh true. I think in the documentation there should be a warning.
And maybe when initializing the spline conv you could add a safe=True/False option. Depending on this it will do either
if safe:
def forward(...) safely
else:
def forward(...) as is...

this is ofc ugly and a dupication of code.
Then again i am bad at programming.

from pytorch_geometric.

I will definitively add a note in the documentation. I think i will add an assertion in the forward pass which gets only called once.

from pytorch_geometric.

Done!

from pytorch_geometric.