Good questions, modeling the parameter space for different knobs will be a focus of future work. It will probably be more practical to learn one parameter space, like gain, and leave EQ for traditional algorithms.
No, the training is not accomplished on the Raspberry Pi, I used Google Colab, which has GPU runtimes. I had similar training times with local GPU. I haven't tried it yet, but I suspect any practical training will take too long on the Pi. I'll probably look into recording samples with the Pi/HiFiBerry hat, and then upload the samples to another computer (or the cloud/Google Colab), and then send the trained model back to the Pi. I believe I'll be able to automate that process at some point.
A question comes to mind: If the original pedal has N knobs, how can ML hope to learn the whole N-dimension parameter space? Are there any shortcuts?
Also, when you say PyTorch learning took 40 minutes, is this also running on RPi?