Segmentation fault (core dumped)

[xiaojinglu]

When I run python reddit.py in pytorch_geometric, which use neighbor_sampler in torch_cluster, it occur 'Segmentation fault (core dumped)'. My machine is Ubuntu 16.04.5, use python3.6 and cuda10.0

[rusty1s]

Can you test if the test suite of torch-cluster works for you?

[xiaojinglu]

Can you test if the test suite of torch-cluster works for you?

I just run python setup.py install and then python setup.py test, all tests passed.

test/test_fps.py ....
test/test_graclus.py ............................
test/test_grid.py ........................................................
test/test_knn.py ........
test/test_nearest.py ....
test/test_radius.py ........
test/test_rw.py ..
test/test_sampler.py .

----------- coverage: platform linux, python 3.6.8-final-0 -----------
Name                        Stmts   Miss  Cover
-----------------------------------------------
torch_cluster/__init__.py      10      0   100%
torch_cluster/fps.py           10      0   100%
torch_cluster/graclus.py       10      0   100%
torch_cluster/grid.py           9      0   100%
torch_cluster/knn.py           37      0   100%
torch_cluster/nearest.py       21      0   100%
torch_cluster/radius.py        31      0   100%
torch_cluster/rw.py            11      0   100%
torch_cluster/sampler.py        8      0   100%
-----------------------------------------------
TOTAL                         147      0   100%


============================= 111 passed in 24.19s =============================

[rusty1s]

Okay, I did not expect this. So installation succeed. The reddit.py example works for me, so I am not sure what may causes this issue. Is it possible to do some debugging on your side?

[Chandrahasd]

I am also facing the same issue.
My graph has 643474 nodes and 1222204 edges.
I am trying to use NeighborSampler to generate batches of size 128. I tried using full neighborhood (size=1.0) as well as subsampling (e.g. size=16).
In both cases, the code terminates with Segmentation Fault at this line of pytorch_cluster/torch_cluster/sampler.py file.

The example code reddit.py from Pytorch Geometric works fine for me.

[rusty1s]

Can you send me your data.edge_index per mail? Curious to debug this!

[xiaojinglu]

My code terminates with Segmentation Fault at this line too when I run python reddit.py

[Chandrahasd]

Can you send me your data.edge_index per mail? Curious to debug this!

@rusty1s, I shared the data.edge_index file over email.
Thank you for looking into this issue.

[rusty1s]

I received your file and will look into it. An initial guess is that it may be due to isolated nodes.

[Chandrahasd]

I thought that too.
I did try adding the self-loops to avoid zero-degree nodes, but it wasn't enough to avoid the error.

[rusty1s]

Works flawless for me:

import torch
from torch_geometric.read import read_txt_array
from torch_geometric.data import Data
from torch_geometric.data import NeighborSampler

edge_index = read_txt_array('edge_indices.txt', sep=' ', dtype=torch.long)
edge_index = edge_index.t().contiguous()
data = Data(edge_index=edge_index)
print(data)

loader = NeighborSampler(data, size=1.0, num_hops=2, batch_size=100,
                         shuffle=False, add_self_loops=False)

for data_flow in loader():
    print(data_flow)

Can you run the code and see if it works for you?

[Chandrahasd]

This code works.
The problem happens when either edge_index or data object or both are in GPU memory.

import torch
from torch_geometric.read import read_txt_array
from torch_geometric.data import Data
from torch_geometric.data import NeighborSampler

# Put edge_index in GPU memory
edge_index = read_txt_array('edge_indices.txt', sep=' ', dtype=torch.long).cuda()
edge_index = edge_index.t().contiguous()
# Put data object to GPU memory
data = Data(edge_index=edge_index).to('cuda')
print(data)

loader = NeighborSampler(data, size=1.0, num_hops=2, batch_size=100,
                         shuffle=False, add_self_loops=False)

for data_flow in loader():
    print(data_flow)

Output:

torch_geometric/data/data.py:177: UserWarning: The number of nodes in your data object can only be inferred by its edge indices, and hence may result in unexpected batch-wise behavior, e.g., in case there exists isolated nodes. Please consider explicitly setting the number of nodes for this data object by assigning it to data.num_nodes.
  warnings.warn(__num_nodes_warn_msg__.format('edge'))
Segmentation fault (core dumped)

[rusty1s]

This makes sense, and we should add an assertion to prevent this. The NeighborSampler is a CPU-op, and hence expects data being on the CPU.

[github-actions]

This issue had no activity for 6 months. It will be closed in 2 weeks unless there is some new activity. Is this issue already resolved?