yzz673 / brant-2 Goto Github PK

Hi, thank you very much for open-sourcing your code for this paper.
I am trying to figure out how to use the model for a new dataset.

It seems that the inputs to the model is in the format of data, psd and labels.
Could anyone share what is the expected format of the data and psd?

I've noticed that it is commented that the data should be of shape [batch size, channel number, 16, patch length] and psd should be of shape [batch size, channel number, 16, psd length].

I have a few questions on this:
(1) Does channel number refer to the number of channels, and can it be variable for a single dataset?
(2) What does the 16 refer to? My new dataset only has 2s recording of 200Hz so there's only 400 samples per observation and it seems that the model runs through too many conv blocks and hits an error. What is the minimum size of the input data?
(3) What is the recommended patch length?
(4) What is the psd length referring to? Is it the 256 frequencies calculated from the fft by default?

Thank you for your time!