CUDA error while runing unsupervised_TU on GITHUB dataset about graphcl

[yyou1996]

Hi @c747903646,

We didn't conduct unsupervised_TU on GITHUB dataset (but do in semisupervised_TU). It looks like the edge index get mismatches when x = F.relu(self.convs[i](x, edge_index)) is executed. Would you confirm that via (i) trying other datasets e.g. NCI1 with the same script, and (ii) running on GITHUB dataset semisupervised setting?

[c747903646]

Hi @c747903646,

We didn't conduct unsupervised_TU on GITHUB dataset (but do in semisupervised_TU). It looks like the edge index get mismatches when x = F.relu(self.convs[i](x, edge_index)) is executed. Would you confirm that via (i) trying other datasets e.g. NCI1 with the same script, and (ii) running on GITHUB dataset semisupervised setting?

Yes, I have run the same script on other TU datasets, such as COLLAB, DD, NCI1, RDT-5K and RDT-B, and all work except run on GITHUB. Meanwhile, semisupervised codes work all TU datasets, including GITHUB.

In debugging, I found that before x = F.relu(self.convs[i](x, edge_index)) is excuted, the dimension of x is matched with self.convs[i]. So I could not understand the error info, srcIndex < srcSelectDimSize.

[yyou1996]

Mismatch could happen if edge_index.max()+1 > x.shape[0], which is my suspect.