I'm contemplating simplifying the UI of the trainers. Currently, there's a checkbox that users should check if the gear they're modeling includes a cabinet at the end of its signal chain:
Browser trainer
GUI trainer on Windows
What it does under the hood is penalize the model for getting the high-frequency part of its predictions incorrect. There's not actually a change to the model like adding an IR "layer" to the end of the model or anything (see also: why you can't "remove" the cabinet from a full rig model).
The proposal is to remove the checkbox from the UI and have it always be "on".
Pros:
Simpler UI, closer to "It just works".
You can always use bin/train/main.py to customize training if you want.
I can roll it back in a future version if something else goes wrong.
Cons:
This would make it harder to tell if the model includes a cab (though the metadata should be used for this, ideally).
Slower training, especially on macOS devices with apple Silicon versions that don't support some of the operations needed (Is this all of them? I don't own any MacBooks with Apple Silicon to test with...)
Slightly higher ESRs, usually (though it's often very minuscule--and sometimes better!)
I'm seeking input about this. Feel free to let me know your thoughts at https://github.com/sdatkinson/neural-amp-modeler/issues/436 or shoot me an email.