[FEATURE] Compositional models - general Directed Acyclic Graphs

[sdatkinson]

Is your feature request related to a problem? Please describe.
Counterpart to #594 but for arbitrary compositions that can be expressed as directed acyclic graphs (DAGs). This is more complex but far more powerful (e.g. allows for parallel processing).

Describe the solution you'd like
Similar to #594, possibly superseding it eventually.

Describe alternatives you've considered
N/A

Additional context
More complex than sequential, more edge cases to handle, harder to infer things like receptive field but still fun :)

[webprofusion-chrisc]

Hi Steven, I'm currently build a new plugin https://guitar.soundshed.com and I'm trying to tackle the problem of blending multiple nam models to represent for instance gain knob (and other parameter) position/audio blending. Is this proposed feature something that works towards that kind of goal or am I reading it wrong?

[sdatkinson]

That's not what I have in mind, no.

One might use this to make a NAM that e.g. combines a NAM of an amp with a NAM of a cab ("full rig"), or perhaps a NAM of a drive pedal into an amp (serial) into two cabs (parallel), perhaps into a preamp model (sum-in), etc.

[harritaylor]

Hey @sdatkinson 👋🏻 this sounds fun.
Is the idea something along the lines of:

1. describe each individual NAM processing block in a DAG
2. pass input.wav through this DAG
3. capture the end result as a new NAM

input.wav-->preamp.nam+-->4x10-cab.nam+-->my-rig.nam
                      |               |             
                      +-->1x12-cab.nam|             

Or do you have something in mind that doesn't require another training step?