Regarding visualization and router

[yinangit]

Great work !

1. How were the visualizations of Figures 4 and 5 made? Taking Figure 4 as an example, how is the missing modality bank with a shape of [$2^{|M|}-1$, num_modality, num_patch, hidden_dim] used to visualize cosine similarity?
2. Is the difference between G-Router and S-Router only based on whether the proposed loss $L_{ce}$ is used or not? Does G-Router only use $L_{balance}$, while S-Router uses $L_{ce}$ and $L_{balance}$? And their structures are the same, just using different loss functions during different training periods?

Looking forward to your answer !