I'll write more details later, but in MultitaskDataset.merge
, it merges the molecules that are identical across datasets. But if the ordering of the nodes/edges/nodepairs are different, then nothing guarantees that the order will be the same, so the nodes that ends up being used might be in a different orders than the labels.
In that function, if there is a node/edge/nodepair task, all the labels must be re-ordered according to the ordering of the unique graph selected.