However, the downside of relying on Flax for its standard Dense module is that there's less hope that the trained model could be re-exported into some target language like C++.
On Thu, Oct 20, 2022 at 1:42 PM David Braun <br...@ccrma.stanford.edu> wrote: > Thanks! > > Yes, referring to your DNN: > https://github.com/grame-cncm/faust/blob/master-dev/examples/filtering/DNN.dsp > > If you look at the DDSP diagram (https://magenta.tensorflow.org/ddsp), > there are encoders and decoders in red. The DNN.dsp could achieve what > those encoders/decoders do. > > Also, the DNN.dsp hard-codes the weights. One would need slightly > different faust code so that the weights are stored in either a > soundfile or sliders. > > Another takeaway is that you don't necessarily need to do this DNN in > Faust. You can use Faust to JAX to get a Flax module, but then combine it > with standard Flax modules such as Dense ( > https://flax.readthedocs.io/en/latest/api_reference/_autosummary/flax.linen.Dense.html) > Get the best of both worlds! > > > > On Thu, Oct 20, 2022 at 1:24 PM Julius Smith <julius.sm...@gmail.com> > wrote: > >> Yes, kudos! >> >> I can't wait to try optimizing some Faust models using gradient descent! >> >> In principle, this solves the training problem for >> <faust>/examples/filtering/DNN.dsp, right? >> >> Cheers, >> - Julius >> >> On Thu, Oct 20, 2022 at 11:16 AM Stéphane Letz <l...@grame.fr> wrote: >> > >> > Hi David, >> > >> > Thanks for this awesome contribution ! Hopefully opening an entirely >> new world in the Machine Learning domain for the Faust ecosystem! >> > >> > Cheers, >> > >> > Stéphane >> > >> > > Le 20 oct. 2022 à 12:48, David Braun <br...@ccrma.stanford.edu> a >> écrit : >> > > >> > > Faust now supports the machine learning framework JAX as a backend. >> DawDreamer is a python module where you can conveniently convert Faust code >> to JAX. DawDreamer has two new Colabs demonstrating this Faust to JAX >> pipeline: >> > > • Faust to JAX: Optimize parameters using gradient descent. >> > > • Faust to QDax: Optimize with Quality-Diversity algorithms. >> > > There's an additional notebook covering the Box API in Python. You >> can access the intermediate representation of Faust, programmatically >> compose boxes, then render audio with DawDreamer, or transpile to a target >> language (C++, JAX, Rust, etc.). >> > > >> > > More information in the tweet: >> https://twitter.com/DoItRealTime/status/1583042473227452416 >> > > >> > > Please check out the Colabs, and feel free to contact me. >> > > >> > > Best, >> > > >> > > David >> > > _______________________________________________ >> > > Faudiostream-users mailing list >> > > Faudiostream-users@lists.sourceforge.net >> > > https://lists.sourceforge.net/lists/listinfo/faudiostream-users >> > >> > >> > >> > _______________________________________________ >> > Faudiostream-users mailing list >> > Faudiostream-users@lists.sourceforge.net >> > https://lists.sourceforge.net/lists/listinfo/faudiostream-users >> >> >> >> -- >> "Anybody who knows all about nothing knows everything" -- Leonard Susskind >> >>
_______________________________________________ Faudiostream-users mailing list Faudiostream-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/faudiostream-users
