Hi, when I try to reimplement your paper “Learning Hierarchical Graph Neural Networks for Image Clustering”
Tabel 2 and Tabel 3 all obtained very high perfermance compared to your benchmark methods. But whether these performance on those dataset used save parameter of (k,l and Pτ)？
Thank you
Hi, thanks for your interests in DGL and HiLANDER. The hyperparameters for Table 2 and Table 3 are specified in Section 4.2, i.e. Implementation Details. In terms of the number of layers in training, if not described in paper, I’d suggest you to try L=3. When doing inference, the number of layers is determined by early-stopping in Section 4.3.