emadeldeen24/AttnSleep

Regarding application of model to SHHS dataset

Closed this issue · 1 comments

Div12345 commented

Hi, thanks for sharing the code open-source!

I was part of trying to integrate this algorithm into braindecode (braindecode/braindecode#341 (comment)) and while trying a test for the SHHS dataset with the sfreq = 125Hz and 30sec windows and d_model = 100, I ran into the problem where the size of data entering the layer norm was 99 and not 100. I tried tweaking the sfreq to 126 and it worked.
Any reason for this discrepancy?

Thanks in advance!

emadeldeen24 commented

I guess it is a sample length issue that leads to kernel size and padding issues in MRCNN. That's why I included MRCNN_SHHS to overcome this issue in SHHS, where I just modified the kernel size of the last Conv1d layer. You should use that and modify d_model to 100.

Hopefully this helps!

  • 👍1