Unexpected behavior from DataLoader on Data with customized attributes

[jipq6175]

Hi, first of all, thanks for the amazing lib.

I encountered an unexpected behavior on loader.DataLoader when mini batching the data.Data with customized attributes. I have a set of graphs and each graph contains a node mask (tensor.bool). When I use DataLoader to create batch data, the node mask was transformed into torch.int. This behavior happens only for certain attribute names.

Reproduce (pyG version 2.0.3)

import networkx as nx

import torch
from torch_geometric.data import Data
from torch_geometric.loader import DataLoader
from torch_geometric.utils import from_networkx

# creating a cyclic graph
g = nx.Graph()
g.add_nodes_from([0, 1, 2])
g.add_edges_from([(0, 1), (1, 2), (2, 0)])

# nx -> pyg data
data = from_networkx(g)

# assign dummy node attributes
data.x = torch.zeros(3, 2)

# different node masking names
data.node_mask = torch.tensor([0, 0, 1], dtype=torch.bool)
data.some_mask = torch.tensor([0, 0, 1], dtype=torch.bool)
data.int_mask = torch.tensor([0, 0, 1], dtype=torch.bool)
data.interfac_mask = torch.tensor([0, 0, 1], dtype=torch.bool)
data.interface_mask = torch.tensor([0, 0, 1], dtype=torch.bool)
data.interface_masks = torch.tensor([0, 0, 1], dtype=torch.bool)
data.face_mask = torch.tensor([0, 0, 1], dtype=torch.bool)

dataloader = DataLoader([data, data, data], batch_size=3, shuffle=False)
for batch_data in dataloader: print(batch_data)
print()

print(batch_data.node_mask)
print(batch_data.some_mask)
print(batch_data.int_mask)
print(batch_data.interfac_mask)
print(batch_data.interface_mask) # ???
print(batch_data.interface_masks) # ???
print(batch_data.face_mask) # ???

The output from above is

DataBatch(edge_index=[2, 18], num_nodes=9, x=[9, 2], node_mask=[9], some_mask=[9], int_mask=[9], interfac_mask=[9], interface_mask=[9], interface_masks=[9], face_mask=[9], batch=[9], ptr=[4])

tensor([False, False,  True, False, False,  True, False, False,  True])
tensor([False, False,  True, False, False,  True, False, False,  True])
tensor([False, False,  True, False, False,  True, False, False,  True])
tensor([False, False,  True, False, False,  True, False, False,  True])
tensor([0, 0, 1, 3, 3, 4, 6, 6, 7])
tensor([0, 0, 1, 3, 3, 4, 6, 6, 7])
tensor([0, 0, 1, 3, 3, 4, 6, 6, 7])

Shouldn't the last three be the same as the others? It seems weird to me some attribute names work and some don't. I wonder what the cause is. Thanks!

[rusty1s]

Interesting. By default, we auto-increment any attribute names that contain the word index or face. For face, this assumption is way too strong, and I fixed it in master, see 16f236b.

[jipq6175]

Thank you