Martin

Hey guys, I found Pytorch has a function `torch.linalg.eigh` which could achieve good results but is much faster than cupy or scipy. I use it to replace `scipy.sparse.linalg.eigsh` in `normalized_cut`....

You may need to tune the weights of different modals.

Hi, can you share your training log and other configurations? What the batchsize and gpu number do you use for training?

The training log seems fine to me. Have you tried CTR-GCN on NW-UCLA? Can you achieve the performance reported in CTR-GCN?

Currently I am working on a new project, the pretrained weight would be released in Nov.