Training fails with: Expected a value of type 'Tensor'

Question

Training fails with: Expected a value of type 'Tensor'

Closed this issue 4 years ago · 3 comments

Hi,
I'm trying to to train the model (on the CGCNN bulk-mod dataset), to do some comparisons between this and other models on transfer learning expensive properties.

Traceback (most recent call last):
  File "train.py", line 111, in <module>
    predictions  = net(data)  
  File "/home/ash/anaconda3/envs/torch/lib/python3.7/site-packages/torch/nn/modules/module.py", line 722, in _call_impl  
    result = self.forward(*input, **kwargs)  
  File "/media/ash/Windows/Users/awsom/Documents/ResearchProjects/Anisotropy ML/GATGNN2/gatgnn/model.py", line 214, in forward  
    x     = self.node_att[a_idx](x,edge_index,edge_attr)  
  File "/home/ash/anaconda3/envs/torch/lib/python3.7/site-packages/torch/nn/modules/module.py", line 722, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/media/ash/Windows/Users/awsom/Documents/ResearchProjects/Anisotropy ML/GATGNN2/gatgnn/model.py", line 145, in forward
    return self.propagate(edge_index, x=x,edge_attr=edge_attr)
  File "/home/ash/anaconda3/envs/torch/lib/python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 236, in propagate
    out = self.message(**msg_kwargs)
  File "/media/ash/Windows/Users/awsom/Documents/ResearchProjects/Anisotropy ML/GATGNN2/gatgnn/model.py", line 158, in message
    alpha = softmax(alpha, edge_index_i, size_i)
  File "/home/ash/anaconda3/envs/torch/lib/python3.7/site-packages/torch_geometric/utils/softmax.py", line 34, in softmax
    out = src - gather_csr(segment_csr(src, ptr, reduce='max'), ptr)
  File "/home/ash/anaconda3/envs/torch/lib/python3.7/site-packages/torch_scatter/segment_csr.py", line 110, in segment_csr
    return segment_max_csr(src, indptr, out)[0]
RuntimeError: segment_max_csr() Expected a value of type 'Tensor' for argument 'indptr' but instead found type 'int'.
Position: 1
Value: 2135
Declaration: segment_max_csr(Tensor src, Tensor indptr, Tensor? out=None) -> ((Tensor, Tensor))
Cast error details: Unable to cast Python instance to C++ type (compile in debug mode for details)

I'm using pytorch 1.6.0, since I couldn't get a working version of geometric with 1.3.0.
Should I try using pytorch 1.5.0 or 1.4.0 maybe, or is there a simple fix?

Very excited to see attention in the cg-nn space; Hopfields is All You Need would seem to indicate that this type of model could have the same kind of aggressive generalizability as transformers - good luck!

Answer 1 · 2020-08-26T18:55:28.000Z

Hey Ash Pera, thanks for reaching out. I certainly understand your frustration with being unable to run the model on Windows. A simple fix would be to use Linux instead of Windows. I doubt that you would encounter any issues. Otherwise, I advise you to use Pytorch 1.5 along with Pytorch-geometric version 1.4.3. Additionally, another thing to be aware of is that Pytorch-geometric won't work on gpu depending on your cuda version, see documentation<https://github.com/rusty1s/pytorch_geometric#installation>. Please let me know how I can be of any assistance.

________________________________ From: Ash Pera <notifications@github.com> Sent: Wednesday, August 26, 2020 2:01 PM To: superlouis/GATGNN <GATGNN@noreply.github.com> Cc: Subscribed <subscribed@noreply.github.com> Subject: [superlouis/GATGNN] Training fails with: Expected a value of type 'Tensor' (#1) Hi, I'm trying to to train the model (on the CGCNN bulk-mod dataset), to do some comparisons between this and other models on transfer learning expensive properties.

TRAINING MODEL ... Traceback (most recent call last): File "train.py", line 111, in <module> predictions = net(data) File "/home/ash/anaconda3/envs/torch/lib/python3.7/site-packages/torch/nn/modules/module.py", line 722, in _call_impl result = self.forward(*input, **kwargs) File "/media/ash/Windows/Users/awsom/Documents/ResearchProjects/Anisotropy ML/GATGNN2/gatgnn/model.py", line 214, in forward x = self.node_att[a_idx](x,edge_index,edge_attr) File "/home/ash/anaconda3/envs/torch/lib/python3.7/site-packages/torch/nn/modules/module.py", line 722, in _call_impl result = self.forward(*input, **kwargs) File "/media/ash/Windows/Users/awsom/Documents/ResearchProjects/Anisotropy ML/GATGNN2/gatgnn/model.py", line 145, in forward return self.propagate(edge_index, x=x,edge_attr=edge_attr) File "/home/ash/anaconda3/envs/torch/lib/python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 236, in propagate out = self.message(**msg_kwargs) File "/media/ash/Windows/Users/awsom/Documents/ResearchProjects/Anisotropy ML/GATGNN2/gatgnn/model.py", line 158, in message alpha = softmax(alpha, edge_index_i, size_i) File "/home/ash/anaconda3/envs/torch/lib/python3.7/site-packages/torch_geometric/utils/softmax.py", line 34, in softmax out = src - gather_csr(segment_csr(src, ptr, reduce='max'), ptr) File "/home/ash/anaconda3/envs/torch/lib/python3.7/site-packages/torch_scatter/segment_csr.py", line 110, in segment_csr return segment_max_csr(src, indptr, out)[0] RuntimeError: segment_max_csr() Expected a value of type 'Tensor' for argument 'indptr' but instead found type 'int'. Position: 1 Value: 2135 Declaration: segment_max_csr(Tensor src, Tensor indptr, Tensor? out=None) -> ((Tensor, Tensor)) Cast error details: Unable to cast Python instance to C++ type (compile in debug mode for details)

I'm using pytorch 1.6.0, since I couldn't get a working version of geometric with 1.3.0. Should I try using pytorch 1.5.0 or 1.4.0 maybe, or is there a simple fix? Very excited to see attention in the cg-nn space; Hopfields is All You Need would seem to indicate that this type of model could have the same kind of aggressive generalizability as transformers - good luck! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<https://protect2.fireeye.com/v1/url?k=5a153796-049a9015-5a157957-86788dc299f7-e472da2623752d9e&q=1&e=843b7a4e-b1d6-4022-9428-f89ed9b72653&u=https%3A%2F%2Fgithub.com%2Fsuperlouis%2FGATGNN%2Fissues%2F1>, or unsubscribe<https://protect2.fireeye.com/v1/url?k=a9deac80-f7510b03-a9dee241-86788dc299f7-9789999f1e91b1ae&q=1&e=843b7a4e-b1d6-4022-9428-f89ed9b72653&u=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAKMQN72U2C4LXP7X5KSSSN3SCVEYHANCNFSM4QMDGW2A>.

Answer 2 · 2020-08-26T20:30:04.000Z

Hi,
Thank you for your reply. Sorry for the confusion, I'm running on Linux, I just keep my research files in a dual-booted windows directory. I'll check if moving it to the Linux partition will fix the problem. I'll also try it with the versions you suggested. Thanks!

Answer 3 · 2020-08-27T00:33:53.000Z

It was a versioning issue - installing pytorch 1.5.0, pytorch-geometric 1.4.3, and appropriate 1.5.x pytorch-geometric dependencies fixed it. Thank you for your help!