Significant audio quality drop (timbre_transfer notebook) in 0.14.0

The recent push to 0.14.0 includes a change that seems to have messed with the performance of the pretrained models. Unfortunately, there were a lot of commits in that version push, so I’m not exactly sure what the cause is yet, but I suspect it may be something regarding loading the checkpoints improperly.

I’m looking into this right now, but for now, if you want to avoid this I recommend installing version 0.13.0, and I will make that the default for the notebook for now. !pip install ddsp==0.13.0

[jesseengel]

Sorry you had the trouble. I believe the DAGLayer was having trouble claiming the variables of its sub components. Looks like it came from using __dict__[] instead of setattr() to create modules in DAGLayer, which was skipping __setattr__ where variable inheritance occurs, so the processor_group variables weren’t collecting the IR variable from Reverb.

As ddsp.training is actively used for research purposes, we unfortunately don’t have the bandwidth to support complete backwards compatibility, but hopefully versions before 1.0 should work for old models and training new models with 1.0 should still work.

On Tue, Feb 2, 2021 at 7:59 AM lauraibnz [e-mail] wrote:

I checked further and the FiltereNoise output signal was actually fine, but the Reverb was acting differently (something to do with the latest changes in the dags?), outputting a very low signal. Finally using a copy of the old RnnFcDecoder with the old nn functions and a copy of the old ProcessorGroup (even using the new synths and effects) made it possible to use the old models as before. Still haven’t figured out why this is happening though.

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/magenta/ddsp/issues/253#issuecomment-771739235, or unsubscribe https://github.com/notifications/unsubscribe-auth/AANFCCM5Y6NJ4EJIBIQ2JCTS5AOOXANCNFSM4TKY3CCQ .

[lauraibnz]

I checked further and the FiltereNoise output signal was actually fine, but the Reverb was acting differently (something to do with the latest changes in the dags?), outputting a very low signal. Finally using a copy of the old RnnFcDecoder with the old nn functions and a copy of the old ProcessorGroup (even using the new synths and effects) made it possible to use the old models as before. Still haven’t figured out why this is happening though.