Strange attributes of HGTConv

[yamaceay]

Hey, I recently noticed that two strange variables are defined in HGTConv.__init__() (torch_geometric/nn/conv/hgt_conv.py), causing low-level errors:

class HGTConv(...):
  def __init__(self, ...):
    ...
    # metadata looks like: [node_types, edge_types]
    # node_types is a list of all node types
    # edge_types is a list of tuples for each meta relation type (src_node_type, edge_type, dest_node_type)

    self.dst_node_types = list(set(metadata[1][1]))
    self.src_types = [edge_type[0] for edge_type in self.edge_types]

self.dst_node_types: It is assigned the unique values of second meta relation type.

For example:

metadata = [["X", "Y", "Z"], [["X", "a", "Z"], ["X", "b", "X"]]]
self.dst_node_types = list(set(metadata[1][1]))
assert(sorted(self.dst_node_types) == ["b", "X"]) # returns: True

The variable is used at one line, which causes an error at my code (torch_geometric/nn/conv/hgt_conv.py:212-215):

  def forward(self, ...):
    ...
    for node_type, out in out_dict.items():
        if node_type not in self.dst_node_types:
            out_dict[node_type] = None
            continue

self.src_types: It is assigned a non-unique list of all src nodes occurring at each edge type. The variable is unused in HGTConv as well as in its parent-classes.

Is there any possibility that it is a bug which hasn't been fixed yet? Or can you explain why and how these variables are used?
Thank you in advance!

[rusty1s]

Thanks for letting us know. This was mistakenly introduced by #7117, and I fixed it in #7183. Let me know if you have any doubts.