Open wwerkk opened 1 year ago
seems like even using the first maxlen
tokens from original sequence on loop does not prevent the rotation - might be inherent to the way this model architecture predicts.
this problem could use some proper debugging for sure.
generations in notebooks works flawlessly, which would suggest the issue is related generation loop in online.py
specifically. some cross referencing between the generate function in the notebooks and in the script could potentially solve this.
The drumloop2 model generates rotations of the sequence which tend to move by token or two with each generation, even with temperature <=1.
Offline generation works properly, as it results in
ABCB
pattern which corresponds exactly to the drum pattern in the data.Might be worth to test offline generation with longer sequences to determine if the pattern shifts in that case as well.