Issue on setting knn about learn-to-cluster

[yl-1993]

Hi @rose-jinyang, it seems that the issue is related to this question on stackoverflow, which may be caused by corrupted faiss_k_160.npz. Could you please check the file to see whether the size is normal or whether it can be read by np.load?

For the kNN setting, we found two empirical rules in our experiments:

• The k for GCN-V and the k for GCN-E are not necessarily to be the same. As GCN-E is capable of predicting robust linkage, increasing the k for GCN-E usually brings performance gain, especially when the neighborhood structure is complex. But as you said, they can share the same k value. In the first version of our arxiv paper, we adopt k=80 for both GCN-V and GCN-E. In current version, we adopt k=80 for GCN-V and k=160 for GCN-E.
• The k for GCN-E training and the k for GCN-E testing are not necessarily to be the same. We can choose to train GCN-E with complex graphs with large k value, and test it with easy local graph with small k value. However, there is a trade-off that if we use a small k value, the candidate set will be small and may not contain the correct one for connection. For simplicity, we usually keep them the same.

[rose-jinyang]

Thanks for your reply.
I trained with k=80 for GCN-V and k=160 for GCN-E.
I tried with k=80 for GCN-V and k=160 for GCN-E even when testing.
GCN-V test was successful but GCN-E test was not.
So I changed k for GCN-V test to 160 and the GCN-E test was successful.
But I am strange that the class_num after GCN-E test is greater that one after GCN-V test.
I tested with 8000 classes.
The number of the predicted classes after GCN-V test was 9383.
But the number of the predicted classes after GCN-E test was 9416.
How can I understand this?
and I am going to use k=160 for testing both GCN-V and GCN-E.
Please let me know if it is okay.
Thanks.

[yl-1993]

(1) As the error is raised when loading knn, I think it may not be caused by the inconsistent k value between GCN-V and GCN-E. But I am not very sure why this happened according to current message. If you can provide a minimum example to reproduce the error, I am happy to help diagnose the problem.
(2) There is no guarantee between the number of predicted classes of GCN-V and GCN-E. If the number of predicted classes of GCN-E increases, it means GCN-E cuts down some linkages according to predictions, partitioning a class into several sub-classes, and verse vice.
(3) Sure. It is okay to use k=160 for both GCN-V and GCN-E, but it will increase the testing time as the built affinity graph is more dense.

[yjhuasheng]

Thanks for your reply.
I trained with k=80 for GCN-V and k=160 for GCN-E.
I tried with k=80 for GCN-V and k=160 for GCN-E even when testing.
GCN-V test was successful but GCN-E test was not.
So I changed k for GCN-V test to 160 and the GCN-E test was successful.
But I am strange that the class_num after GCN-E test is greater that one after GCN-V test.
I tested with 8000 classes.
The number of the predicted classes after GCN-V test was 9383.
But the number of the predicted classes after GCN-E test was 9416.
How can I understand this?
and I am going to use k=160 for testing both GCN-V and GCN-E.
Please let me know if it is okay.
Thanks.

您好，请问你聚类的结果如何呢，我使用自己的数据进行在vgcn上训练，效果不是很好，可以和您交流一下吗