So I am going through your paper and in Section 3.1 Edge Convolution, equation #8 suggests some learnable weights theta and phi.
Looking at the graph feature function, it does not use any weights when constructing a graph (feature = concat(xi - xj, xi)). Is that right or subsequent convolutional layers in partseg or semseg models allow learning of weights for this graph feature?
By the way, I am using this implementation of your network in pytorch
So I am going through your paper and in Section 3.1 Edge Convolution, equation #8 suggests some learnable weights theta and phi.
Looking at the graph feature function, it does not use any weights when constructing a graph (feature = concat(xi - xj, xi)). Is that right or subsequent convolutional layers in partseg or semseg models allow learning of weights for this graph feature?
By the way, I am using this implementation of your network in pytorch
i.e