Closed SeongJinAhn closed 4 years ago
θ is indeed a vector of K polynomial coefficients describing a filter. For computational efficiency, MxN filters (which compute N output feature maps from M input feature maps) are applied at the same time to the input. Hence θ is implemented as an array of shape KxMxN.
Thanks. I understand that theta can be extended for computational efficiency. However, then there is another question.
Then is it right that ChebConvs assume that coefficients of L^k (theta) depend on the features. rather then the same thetas do work to every input feature and output feature. Is it right??
Yes. Filters are learned to compute output feature maps from input feature maps, hence are specific to a pair of input-output features. Say the input is an RGB image (3 features) and the output a vegetation segmentation mask (1 feature). You'd want the 3x1=3 filters to be different, because the input features have a different meaning.
In the paper, the theta is said to be a vector of polynomial coefficients. Hence, I understood that theta_k should be a constant.
However, it seems that theta_k is a matrix in complemented codes. Is there any reason to complement in this way? Or theta should be revised to matrices in R^KxMxN?