search

microsoft / PhysioPro

A deep learning framework for physiological data processing and understanding.
MIT License
67 stars 12 forks source link

customized dataset #29

Open Chic-J opened 1 month ago

Chic-J commented 1 month ago

Hi! Thanks for this great job！I read the MMM paper，and i want use it in UNM Dataset（EEG datatype as well. I wonder how to get the attention mask for the UNM Dataset？Or how can i train without attention mask？

victorywys commented 1 month ago

Hi, thanks for your interest in our paper! The setting of the attention mask is demonstrated as Figure 1(d) in our paper. Generally speaking, to design your own attention mask, you need to:

  1. Divide channels into regions (groups)
  2. Add a hypernode for each region
  3. Allow attention among channels of the same group
  4. Allow attention among hypernodes.
  5. Allow attention among between channel and the hypernode of its region.
  6. Disallow all other connections.

To modify it from the code, you can refer to the code here. By setting the func_areas, you can build your own dataset. And don't forget to modify the number 62 in the following lines to your number of channels.

If you have further questions, please let us know.

Chic-J commented 1 month ago

I got it. are there any suggestions for region splitting and hypernodes?

victorywys commented 1 month ago

Generally, we split channels to groups according to the brain regions. Appendix C.1 in our paper includes our default setting of region division as well as some ablated ones. You can map your channels to such regions or customize your own setting accordingly.