You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In calling the model multiple times to create audio longer than 4 seconds, I ran into a GPU OOM issue caused by crepe.
ddsp.spectral_ops.reset_crepe() is used in the notebooks, but is not really documented.
It turns out that if you call ddsp.spectral_ops.reset_crepe() repeatedly (once before every model call) that leaks memory. This method should only be called once.
PROPOSED RESOLUTION:
Add code comment to ddsp.spectral_ops.reset_crepe() describing that it should only be called once per session AND that it leaks memory if called repeatedly (instead of just freeing its memory). Each time a colab uses this command, also document there this peculiarity.
Lastly, but not required, it would be very helpful that the timbe transfer colab actually works on the full audio file, instead of 4 seconds.
The text was updated successfully, but these errors were encountered:
What also should be described is, if I load a new DDSP model, do I need to re-initialize crepe or not?
It would also just be great to fix the underlying leak so repeatedly calling crepe wouldn't be so fraught with peril
In calling the model multiple times to create audio longer than 4 seconds, I ran into a GPU OOM issue caused by crepe.
ddsp.spectral_ops.reset_crepe()
is used in the notebooks, but is not really documented.It turns out that if you call
ddsp.spectral_ops.reset_crepe()
repeatedly (once before every model call) that leaks memory. This method should only be called once.PROPOSED RESOLUTION:
Add code comment to
ddsp.spectral_ops.reset_crepe()
describing that it should only be called once per session AND that it leaks memory if called repeatedly (instead of just freeing its memory). Each time a colab uses this command, also document there this peculiarity.Lastly, but not required, it would be very helpful that the timbe transfer colab actually works on the full audio file, instead of 4 seconds.
The text was updated successfully, but these errors were encountered: