RandomLinkSplit' object has no attribute 'num_features'

❓ Questions & Help

I created my heterograph following steps from LOADING GRAPHS FROM CSV and now I’m trying to split the data data.train_mask = data.val_mask = data.test_mask = None data = T.RandomLinkSplit(data)

In the last version of PyG, train_test_split_edges is returning that its deprecated. So I used instead RandomLinkSplit and when I execute this line: encoder = VEncoder(data.num_features, out_channels=latent_size)

It returns: AttributeError: 'RandomLinkSplit' object has no attribute 'num_features'

What’s the equivalent of num_features ? I couldn’t find anything about this in the docs

[anniekmyatt]

Yes, you’re right, it is just like the behavior in deepsnap (I just double checked). I think your interface is a bit clearer, because in deepsnap you first call

dataset = GraphDataset([dg], task=task, edge_message_ratio=0.8 , edge_train_mode="disjoint")

and then

dataset_train, dataset_val, dataset_test = dataset.split(transductive=False, split_ratio=[0.8, 0.1, 0.1])

which could suggest to an unsuspecting user that 80% of edges are reserved for message passing, and 20% are supervision edges, and then you split those supervision edges using the [0.8, 0.1, 0.1] split ratio, yielding something like 80 message passing edges, 16 training supervision edges, 2 validation and 2 test edges. But this is not the case, the behaviour is as you implemented here.

I think the fact that you have it all handled in one function doesn’t give any counter-intuitive pointers like this and is much clearer.

Thanks again! 😃

[rusty1s]

This is correct. I think this is similar to how DeepSNAP determines the train ratio of message passing edges and supervision edges. Let me know if this is the correct behavior in your opinion 😃