When using the ProteinGraphDataset class, if a single graph fails construction in construct_graphs_mp, the graph will be passed to transformation functions as None and this will often cause them to fail.
If None graphs are filtered out of the data_list, then a list index out of range error will result as the loop that saves the torch Data objects, as it saves the filenames from the original pdb list (list of uniprot ids) and is not aware which specific IDs had failed graph constructions.
Expected behavior
The ProteinGraphDataset should be robust in handling failed graph constructions and store a list of failed / successful graphs, so that the returned object can be used with valid indexes still (and the user can be aware of which samples in the newly constructed dataset are not able to be used).
When using the
ProteinGraphDatasetclass, if a single graph fails construction inconstruct_graphs_mp, the graph will be passed to transformation functions asNoneand this will often cause them to fail.If
Nonegraphs are filtered out of thedata_list, then alist index out of rangeerror will result as the loop that saves the torchDataobjects, as it saves the filenames from the originalpdblist (list of uniprot ids) and is not aware which specific IDs had failed graph constructions.Expected behavior The
ProteinGraphDatasetshould be robust in handling failed graph constructions and store a list of failed / successful graphs, so that the returned object can be used with valid indexes still (and the user can be aware of which samples in the newly constructed dataset are not able to be used).