Pytorch_geometric: issues on dense_diff_pool function

[PandaGuang]

I was trying to run pytorch_geometric/examples/enzymes_diff_pool.py , but I came across an issue, which is below,

x, adj, l1, e1 = dense_diff_pool(x, adj, s, mask)
ValueError: not enough values to unpack (expected 4, got 3)

I read the description of this function on Pytorch Geometric Documentation and it states that the function dense_diff_pool is supposed to return 4 output, but running the code, it only gives 3.

[rusty1s]

Ah I see, it looks like you are running an older version of PyG. The current master version should fix this problem.

[PandaGuang]

Get it! It works! Thank you very much!

[PandaGuang]

When I change the number of epochs from 151 to a larger number, the training loss becomes NAN after 150 epoch, what is the problem?

[rusty1s]

I will look into it.

[rusty1s]

The issue is fixed in current master. There was a bug in DenseSAGEConv where neighborhood normalization was performed even if in case there were no neighbors (which could happen for DiffPool). Thank you.