How to train the lightweight scoring network to pickup import patches to be VQ-ed.

[yexiguafuqihao]

Hi, I have a question as described in the issue, the way/rules to separate all feature vectors in the grid feature map from the encoder into important and unimportant samples confused me. As depicted in Sec 3.2: The larger the score s_l is, the more important the region feature z_l is. However, I don't understand the label assignment strategy to distinguish the important samples from the unimportant ones to train the lightweight scoring function. Would you please kindly specify it in this issue? I appreciate it in advance.