G1 and G2 parameters

[ghost]

There are G1 and G2 parameters in 'predict' and 'update' layer formulas. It is initially said that they represent just number of groups in conlvolutional layers. Later, in 'Parameter Efficiency' section, you tell that G2 also equals the number of internal channels, so we should specify out_channels = G2 for the first convolution and in_channels = G2 for the second one. Than, in this answer you claim that you have C * K + C parameters for C channels and kernel size K. So, did I understand correctly, that you've set G1 = G2 = C? Seems like it means that LiftPool layer doesn't take into account inter-channel dependencies.