Target size (torch.Size([128, 12])) must be the same as input size (torch.Size([1, 12]))

[rajarshiche]

I am trying to binarily classify Tox21 data using dgllife GATPredictor. The code link is attached below and I can run only when 'batch_size': 1.
Whenever, I am using 'batch_size': 128 (or any >1 value), I am getting the error 'Target size (torch.Size([128, 12])) must be the same as input size (torch.Size([1, 12]))'. This is the case even when I am defining the batch_size in DataLoader which uses a collate function to batch the data as per the defined batch size.
How and where can I change the input size (or target size) so that the above discrepancy does not arise?

Code link: https://github.com/rajarshiche/GNNs/blob/main/GAT_trial1.py

[mufeili]

I suspect the issue is related to L325. If you call dgl.add_self_loop(bg), the new graph returned will have batch size 1. As a result, the first dimension of the model prediction will be 1. For a workaround, add self loops for individual graphs before batching them.

[rajarshiche]

Thanks a lot Mufeili. You pointed me towards the right direction. Added self-loop on the graphs before batching under the collate_molgraphs function in the code:
"...........
for i in range(len(smiles)):
graphs[i] = dgl.add_self_loop(graphs[i])
bg = dgl.batch(graphs)
............."

It started running with any batch_size. However, I had to do a minor correction in roc_auc_score function to absorb the effect of batch_size. Will update the corrected code on my github page.

Best!
Raj