Strange behaviour with 'dataset.num_classes'

[CaueEvang]

Hi, dear community.

I'm fair new to the usage of PyG, and I'm trying to create a custom dataset based on a CSV file. I've read the oficial documentation and this other tutorial and come up with the following code:

#First the train dataset
graph_list_train = []
for ijet in range(0,n_train):
    jet = ak_array_new[ijet]
        
    node_features, label   = get_node_label_features(jet)
    edge_features, n_edges = get_edge_features(jet)
    edge_index             = get_edge_index(jet)
    
    label = torch.tensor(label, dtype=torch.float)
        
    graph = Data(x = node_features,
                 edge_index = edge_index,
                 edge_attr  = edge_features,
                 y          = label)
    graph_list_train.append(graph)
        

#Now the test dataset
graph_list_test = []
for ijet in range(n_train,n_jets):
    jet = ak_array_new[ijet]
    
    node_features, label   = get_node_label_features(jet)
    edge_features, n_edges = get_edge_features(jet)
    edge_index             = get_edge_index(jet)
        
    label = torch.tensor(label, dtype=torch.float)
    
    graph = Data(x = node_features,
                 edge_index = edge_index,
                 edge_attr  = edge_features,
                 y          = label)
    graph_list_test.append(graph)

where:

• get_node_label_features() returns a matrix of shape [n_nodes, 2 features] and the label of the graph as a scalar(to be further converted to a torch tensor)
• get_edge_features() return as matrix of shape [n_edges, 1]
• get_edge_index returns the COO connectivity matrix

After this, I save the lists as a .pt file:

torch.save(graph_list_train, 'graph_dataset_train.pt')
torch.save(graph_list_test, 'graph_dataset_test.pt')

The issue appears when I load these files and try to access information about the graphs: num_nodes, num_edges, and other properties seem to be okay, but num_classes returns an error message.

Without num_classes:

dataset_train = torch.load('graph_dataset_train.pt')
# Gather some statistics about the first graph.
print(f'Number of nodes: {data_train.num_nodes}')
print(f'Number of edges: {data_train.num_edges}')
print(f'Average node degree: {data_train.num_edges / data_train.num_nodes:.2f}')
print(f'Has isolated nodes: {data_train.has_isolated_nodes()}')
print(f'Has self-loops: {data_train.has_self_loops()}')
print(f'Is undirected: {data_train.is_undirected()}')

output:

Number of nodes: 28
Number of edges: 378
Average node degree: 13.50
Has isolated nodes: False
Has self-loops: False
Is undirected: False

With num_classes:

# Gather some statistics about the first graph.
print(f'Number of nodes: {data_train.num_nodes}')
print(f'Number of edges: {data_train.num_edges}')
print(f'Number of classes: {data_train.num_classes}')
print(f'Average node degree: {data_train.num_edges / data_train.num_nodes:.2f}')
print(f'Has isolated nodes: {data_train.has_isolated_nodes()}')
print(f'Has self-loops: {data_train.has_self_loops()}')
print(f'Is undirected: {data_train.is_undirected()}')

output:

Number of nodes: 28
Number of edges: 378
KeyError: 'num_classes'

During handling of the above exception, another exception occurred:

AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_410/283485623.py in <module>
      8 print(f'Number of nodes: {data_train.num_nodes}')
      9 print(f'Number of edges: {data_train.num_edges}')
---> 10 print(f'Number of classes: {data_train.num_classes}')
     11 print(f'Average node degree: {data_train.num_edges / data_train.num_nodes:.2f}')
     12 print(f'Has isolated nodes: {data_train.has_isolated_nodes()}')

/cvmfs/sft.cern.ch/lcg/views/LCG_104a_cuda/x86_64-centos7-gcc11-opt/lib/python3.9/site-packages/torch_geometric/data/data.py in __getattr__(self, key)
    439                 "dataset, remove the 'processed/' directory in the dataset's "
    440                 "root folder and try again.")
--> 441         return getattr(self._store, key)
    442 
    443     def __setattr__(self, key: str, value: Any):

/cvmfs/sft.cern.ch/lcg/views/LCG_104a_cuda/x86_64-centos7-gcc11-opt/lib/python3.9/site-packages/torch_geometric/data/storage.py in __getattr__(self, key)
     78             return self[key]
     79         except KeyError:
---> 80             raise AttributeError(
     81                 f"'{self.__class__.__name__}' object has no attribute '{key}'")
     82 

AttributeError: 'GlobalStorage' object has no attribute 'num_classes'

I appreciate any help on this matter.

[rusty1s]

The num_classes property is only available at the InMemoryDataset level. Since you are working here with a list of data objects (which is totally fine), this property is not available.