Open CarloLucibello opened 3 years ago
Message-passing neural network relies on the connection between nodes. Messages are passed through edges. Therefore, the isolated node can be removed to avoid the error. I think this is not a bug and it is born this way.
Removing the isolated node is not always correct--the readout/pooling may easily be affected by this. There is nothing in the original paper or the mathematical form given in the docs that suggests this should fail, and the error message is very unhelpful. This should at least be officially documented, and probably also have an equivalent to @assert !any(issequal(0), nv.(graphs))
Exposed in #214