Error on training phase

[GiuliCler]

Hi and thank you for the work! I tried the tool on Google Colaboratory with the cooked dataset https://drive.google.com/file/d/1AHxXQhS2UVKOxNUfuetrM-uVKHjpjSOs/view?usp=sharing and checkpoint https://drive.google.com/file/d/1xAnJwPEd1DzsxHW2Z_SLZikgiUwS6_zW/view?usp=sharing, loaded through torch.

After unzipped the cooked dataset and load the checkpoint of the model, run_main.sh script (with variables cooked_root, data_name and save_dir modified accordingly) has been launched for the training phase. After half an hour of training, this error appears:

ValueError: MessagePassing.propagate only supports torch.LongTensor of shape [2, num_messages] or torch_sparse.SparseTensor for argument edge_index.
0% 0/100 [00:04<?, ?it/s]

Any suggestions about it?

[Hanjun-Dai]

I guess this might be the version mismatch of pytorch_geometric.

You may want to downgrade the pytorch_geometric and its dependencies (torch_scatter, etc)

[GianlucaDeStefano]

In alternative, this error is addressable by simply stacking the edge indexes into a single tensor:
edge_indexes = torch.stack((list_edge_idx[i][0],list_edge_idx[i][1]))
t = self.conv_layers[lv](chunks[i], edge_indexes)