Some questions about the eval metrics of this paper.

[lajihaonange]

After reading your paper and code, some questions confused me. I would be very appreciate if someone can explain my question.
First, it was argued swin-T is used in this paper, but actually, in your model definition code, it seems to be swin-L.
Second, your evaluation code calculates top-5 in function _average_top_k_result(eval.py)，using all output of modules, such as select, drop, FPN layers, combiner, original output, and cat those output to get the final metrics, is this reasonable?
Thirdly, the highest-5 acc in your code are not equal to any layer, printed as this picture. So, what is the meaning of highest 1-5? And how to get them?

[lajihaonange]

Well, now I know swin-t == swin-transformer. But the last two question still confuse me a lot now.

[lajihaonange]

And in my environment, my result is as this picture:

So it confuses me that highest results is different with layer 1-4 and the combiner layer.

[archerli1]

Hello, I have read the code of this PIM program and there is a part that confuses me. The author seems reluctant to answer questions. I wonder if you could help me with it? In the paper, the author used four types of errors for classification training, but it seems that only the selected points were used for prediction through graph convolution in the end. I'm not sure if my understanding is correct. If it is, why use four types of errors for training instead of just using the error generated by the prediction of graph convolution?