Proto Init

[YcZ76]

Hi, thanks for such a promising work on graph-level task. Here i just confront some problems. According your paper, In E-step , you adopt a categorical distribution to init your prototypes, and this is consistent to the code. While i notice that you also mention using k-means to init this in Section 3.2 and 5.1, the code publised here seems you did not include that? Or, what am i missing? thanks~

[HelloWorldLTY]

Hi, I think you should run k-means with n_cluster=50 to initiliaze it. In their version of ICLR submission, they use PRCL with k-means to initiliaze the cluster.