Why are my test results far better than the paper results?

[hexiangdong2020]

I used Husformer to test it on the WESAD dataset and got the final result (Acc: 91.02%, F1: 87.38%). The results of this test are much higher than those tested on the WESAD dataset in the paper (Acc: 78.68%, F1: 79.51%). Why is this happening? Is it because the paper uses a different metric than the open source code? @zdz0086 @SMARTlab-Purdue @wonsu0513 @R7-Robot

[xnes]

Hello @hexiangdong2020 , how did you make to run the code? I am trying to use cogload but I'm lost with the bugs.

Did you have to do a lot of changes?

[Marfbin]

Hi, @hexiangdong2020 , I'm having trouble replicating the husformer code using the DEAP dataset. The definition of output_dim=1 in the main.py file causes the final output shape of the model to look like this (1024, 1), which causes problems in focalloss for calculating log_softmax for the first dimension, which results in logpt values being all zeros. How did you solve this problem？

[Shahtaj-Khalid]

I used Husformer to test it on the WESAD dataset and got the final result (Acc: 91.02%, F1: 87.38%). The results of this test are much higher than those tested on the WESAD dataset in the paper (Acc: 78.68%, F1: 79.51%). Why is this happening? Is it because the paper uses a different metric than the open source code? @zdz0086 @SMARTlab-Purdue @wonsu0513 @R7-Robot

Hi, is it possible for you to upload your test code? I'm trying run Husformer for WESAD too but facing issues.