Please stop changing the codebase. Also, the LF component is different from the paper's #18
Description
Please, if you change the codebase, try to make the changes minimum or to not require us to download the code and modify, file by file all of the changes you have made. It is the third time that I find you have completely changed your code.
Furthermore, you have the low frequencies hard coded to just represent the first one (and repeating it for the rest of the frequency bands), which would correspond to the median of the signals, this is bad practice as your paper indicates otherwise.
Lastly, the third stage is simply adding the high frequencies/noise that your model could not capture, and all this does is add jittering to the signals, a traditional data augmentation approach that does not fit into the generative modeling paradigm, and in my opinion it is kind of cheating.
Hope you take these comments lightly. I have been using your code to generate 12-lead ECGs and have reached decent results, but I do find it a pain that you change your code at will, deviate from your paper and make older versions virtually useless.