Unexpected behaviour of label tensors in save/load_graphs

Questions
#1

Hello!

I have noticed a strange behaviour when saving/loading graphs with incorrect label dimension:

import torch
import dgl

g = dgl.graph((torch.tensor([1,2,3]), torch.tensor([2,1,3]))
dgl.save_graphs(
  filename = 'test.bin',
  g_list=[g], labels = {"test": torch.tensor(42)}
)

if we load it, the result is not expected (label tensor value is different):

print(dgl.load_graphs(filename = 'test.bin'))
([Graph(num_nodes=4, num_edges=3,
      ndata_schemes={}
      edata_schemes={})], {'test': tensor(0)})

“Correct” way to save is to add matching dimension to the label tensor

dgl.save_graphs(
   filename = 'test.bin',
   g_list=[g], labels = {"test": torch.tensor([42])}

dgl.load_graphs(filename = 'test.bin')
([Graph(num_nodes=4, num_edges=3,
      ndata_schemes={}
      edata_schemes={})], {'test': tensor([42])})

It would perhaps be beneficial to add a dimension check during save_graphs, to stop this mistake.
Typical pattern when this happens is when one wishes to save each graph separately for future data-loading purposes.

Thank you!