Pytorch_geometric: Bool tensor does not support the `-` operator

[WilliamCCHuang]

🐛 Bug

Hi, this tool is very powerful and amazing. Thanks a lot!

However, when I ran gcn.py from example for testing, an error occurred:

Traceback (most recent call last):
File "gcn.py", line 55, in
train()
File "gcn.py", line 39, in train
F.nll_loss(model()[data.train_mask], data.y[data.train_mask]).backward()
File "/Users/william/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 547, in __call__
result = self.forward(input, *kwargs)
File "gcn.py", line 25, in forward
x = F.relu(self.conv1(x, edge_index))
File "/Users/william/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 547, in __call__
result = self.forward(input, *kwargs)
File "/Users/william/anaconda3/lib/python3.6/site-packages/torch_geometric/nn/conv/gcn_conv.py", line 100, in forward
self.improved, x.dtype)
File "/Users/william/anaconda3/lib/python3.6/site-packages/torch_geometric/nn/conv/gcn_conv.py", line 78, in norm
edge_index, edge_weight, fill_value, num_nodes)
File "/Users/william/anaconda3/lib/python3.6/site-packages/torch_geometric/utils/loop.py", line 107, in add_remaining_self_loops
inv_mask = 1 - mask
File "/Users/william/anaconda3/lib/python3.6/site-packages/torch/tensor.py", line 325, in __rsub__
return _C._VariableFunctions.rsub(self, other)
RuntimeError: Subtraction, the - operator, with a bool tensor is not supported. If you are trying to invert a mask, use the ~ or bitwise_not() operator instead.

It seems that bool tensor does not support the - operator yet. So I modified the line 107 in loop.py as:

inv_mask = ~mask # inv_mask = 1 - mask

and then re-run gcn.py. The result shows that it works!

...
Epoch: 200, Train: 1.0000, Val: 0.8080, Test: 0.8240

Environment

• OS: macOS 10.14.6
• Python version: 3.6
• PyTorch version: 1.2.0
• CUDA/cuDNN version: 10.1
• GCC version: clang
• Any other relevant information:

[tiancheng-zhi]

I met the same issue. It might be because of the new pytorch version

[rusty1s]

Thank you. This is already fixed in master, so installing from source should fix it. I will craft a new minor release soon.

[Magnetar99]

Hey, @rusty1s thanks for this amazing library.

I met the same issue for Knn graph. I even installed the latest commit but the issue still persisted. Also, when I run tests, the issue repeats at multiple times for knn related functions.

Other than the above errors, I also see failures with the error "ValueError: multi_class should be either multinomial or ovr, got auto" in the file logistic.py and there is also assertion error in the file test_signed_gcn.py.

[rusty1s]

Please update torch-cluster to version 1.4.4. The SignedGCN bug should be fixed in PyG master by now.

[neonb88]

Maybe I'm misunderstood, but doing pip install torch-cluster didn't fix the issue for me

[rusty1s]

Are you sure you have the latest version installed? You can check this via

import torch_cluster
print(torch_cluster.__version__)

And update via

pip install --upgrade torch-cluster

[neonb88]

I ended up solving the problem the way the OP solved it. I may take a look at your solution later. Thanks for your contributions to the open source world, rusty1s!

[tjiagoM]

Hello, I'm running the basic example of GCN under "Learning Methods on Graphs" in the introduction section: https://pytorch-geometric.readthedocs.io/en/latest/notes/introduction.html

And I'm having the same issue. Should I downgrade pytorch, as this was working before? I have version 1.4.4 of torch_cluster already. My versions of pytorch are:

>>> import torch ; print(torch.__version__)
1.2.0

>>> import torchvision ; print(torchvision.__version__)
0.4.0a0+6b959ee

>>> import torch_cluster ; print(torch_cluster.__version__)
1.4.4

import torch_geometric ; print(torch_geometric.__version__)
1.3.0

Any suggestions?

[rusty1s]

This is a bug introduced by PyTorch 1.2. You can install PyTorch Geometric from master to fix this (just clone the repository and run python setup.py install). I will craft a new release soon to fix this.

[tjiagoM]

@rusty1s thanks for the update! I just wanted to add that now, when you execute the example I mentioned, we receive a warning:

UserWarning: indexing with dtype torch.uint8 is now deprecated, please use a dtype torch.bool instead.

According to pytorch/pytorch#24918, in the future this will raise an error, so might be worth it to update the example.

Thanks for such awesome tool!

[rusty1s]

Thank you for the reminder, will work on it.