How to deal with different number of channels within one dataset?

[ddicee]

Hi Weibang,

Thanks for sharing your excellent work! There's one thing that I'm not super clear about. When preprocessing a dataset like TUEP, in which the number of channels ranges from 19 to 23, how did you deal with this variation so that different samples could be batched and dumped into hdf5 files for later use?

Thanks in advance.

[935963004]

Thanks for your interest. In my implementation, I generate multiple hdf5 files for different channel numbers. For example, all EEG data with 19 channels are transformed into one hdf5 file while EEG data with 23 channels are transformed into another hdf5 file.

[ddicee]

Thanks a lot for your answer. I'm training the VQNSP using the default configurations as described in the readme, but noticing that the quant_loss keeps increasing while the total loss is decreasing. Just wonder if you have ever seen something like this before. Would you consider this as normal?

[935963004]

This is a normal phenomenon. The quant_loss will first increase and then converge to a stable value (about 0.02 in my case). However, the total loss will be always decreasing. So, don't worry about it.