No Softmax in GCN example

[utkarsh0902311047]

Hi,
I don't understand why there is no softmax layer in the GCN example mentioned in the original semi-supervised learning paper.
Also, please explain to me a little about where multi-layer perceptron is used in two layers of GCN.

Thank you

[akihironitta]

I don't understand why there is no softmax layer in the GCN example mentioned in the original semi-supervised learning paper.

@utkarsh0902311047 I believe that's because the loss used in the example is torch.nn.functional.cross_entropy that doesn't require inputs to be normalised as documented in the PyTorch docs:

The input is expected to contain the unnormalized logits for each class (which do not need to be positive or sum to 1, in general).

[utkarsh0902311047]

@akihironitta
Thank you for the explanation.
Please also help with my second question, "where multi-layer perceptron is used in two layers of GCN?"

[rusty1s]

Each GNN layer performs a linear transformation, so stacking these layers + non-linearities will resemble MLPs (in case of empty edge indices).