I am trying to run distributed link prediction on ogbl-citation2 graph. I have partitioned the graph following the steps in the the official guide, but I could not make the training work. This seems different from the distributed node classification which works fine.
Is there a complete distributed link prediction training script like the distributed node classification example?
Also, how to run it using the dgl launch script? Are the arguments the same as node classification?
Does anyone has a complete distributed link prediction training script along train, val, test accuracy and the command to run it using the dgl launch script?
@Rhett-Ying could you please hlelp?
In the official guide:
num_labels = len(th.unique(g.ndata['labels'][0:g.num_nodes()]))
but the dataset does not contain any labels:
[rank0]: Traceback (most recent call last):
[rank0]: File "train.py", line 50, in <module>
[rank0]: num_labels = len(th.unique(g.ndata["labels"][0 : g.num_nodes()]))
[rank0]: File "dgl/python/dgl/distributed/dist_graph.py", line 386, in __getitem__
[rank0]: return self._data[key]
[rank0]: KeyError: 'labels'