Mapping edge_attr back to original node indices after sampling

[jasperhyp]

I sampled a huge graph and got a small Data object for each batch, but I cannot find an efficient way to map edge_attr back to the original node indices after I map edge_index back to the original node indices with the following code:

ori_map = batch.n_id
batch_edge_index = ori_map[batch.edge_index]

where n_id is initially just torch.arange(num_of_nodes).

The only way I can think of is to directly search for each new edge in the mapped edge indices... But that is way too inefficient.

[rusty1s]

Can't you do

data.e_id = torch.arange(data.num_edges)
...
data.edge_attr[batch.e_id] = batch.edge_attr

[jasperhyp]

Ah sorry my bad, that works. I tried this before but I came across other issues, then I forgot this orz

[jasperhyp]

I think e_id won't work because we'll still need to call the original-length edge_index to get mapping.

But actually this is a dumb question... We don't need to map edge_attr back because once we've mapped edge_index back edge_attr is just working as usual orz Sorry for taking your time!