NAM: Neural Amp Modeler

BenIfin said:
^^^ In my view, you have perfectly summarized the key point about NAM and every other emerging potentially open-sourced AI modeler.

These are the very reasons why Tonex is such an awesome sounding and feeling and supported unit and eco-system.

Then think about when it was released and made available to buy and use to "the public" - in short, its still in its absolute infancy and will doubtlessly get better and better and better etc.

I know this is not a popular view from those that are all-in on the "open sourced" approach, but i.m.h.o, the only real chance that NAM-like products have to compete at a "pro" level, is for a "big-player" to adopt and incorporate and develop them into their own software and hardware products - unless/until that happens, its a free and fun tech to play with.

Ben
Click to expand...

The funny part is, NAM isn't really a product in that sense and hasn't really been developed towards being one. I know I've already said it here, but this is *still* a project focused on the tech more than anything. The notoriety it's picked up recently has put it on a lot of people's radar as a "competing solution" to big companies and commercial offerings, but it wasn't really meant to be that - that's more of the expectation of the greater community now that they've discovered it and started asking for many features / use cases / etc. I think to a certain extent that expectation has now shaped how it will grow going forward in some ways now and it might become more user oriented.

There was never a "plan" to build a whole ecosystem like tonex, ndsp, etc have - all community borne ideas which are truly amazing to see coming to life, but you'll also have to take the whole thing from the wider point of view that this kind of project differs from what most users are used to seeing.
 
northern_fox said:
Your GPU is 2x as fast as mine :(
Click to expand...
mwahahahahahahaha! Sweeeet.

I'm running an NVIDIA 3060ti. You can pick those up for reaaaaaaasonable prices, depending on budget.

I just jumped straight to 1000 epochs.

I'm kinda interested in looking at the guts of this thing. I've done a bit of TensorFlow in the past myself, training a model to recognise certain features of audio files. Would be quite educational I think!
 
Orvillain said:
mwahahahahahahaha! Sweeeet.

I'm running an NVIDIA 3060ti. You can pick those up for reaaaaaaasonable prices, depending on budget.

I just jumped straight to 1000 epochs.

I'm kinda interested in looking at the guts of this thing. I've done a bit of TensorFlow in the past myself, training a model to recognise certain features of audio files. Would be quite educational I think!
Click to expand...

You can get as deep as you like by tweaking to configuration files as well - but often the better the result the heavier the model in CPU. Adding the the middle of the v1_1_1 wav is a good way to improve ESR without adding weight. There are also other training models besides wavenet that can be played with if you know how (I don't). I generally get 10-13 IT/s on my 1080ti which is actually pretty good considering its age. Taking me about 7 days of nonstop training to do the lite/feather versions of my GP1000 set (67 captures)
 
339614782_237469952127709_3707733840473350344_n.png


This is.... surprisingly good.
 
Not bad for being a total fucking idiot.

I tell you what else as well.... I fucked up. I had been doing the calibration stuff and had left my send track set to -17dB. So the amp was receiving a lower signal than it should have been.

So I turned up the input knob in the NAM plugin to 17dB to account for it. But I figured it would be total bullshit compared to the real amp. And it isn't. It's still very close. That means the transfer function of the input circuit part of the model (layman understanding!) is actually really accurate.
 
Orvillain said:
Even 500 epochs is giving me 0.008 ESR.
Click to expand...

If I'm doing large sets I only run to about 600 - you can measure the difference with a null test and find that somewhere between 400-1000 it often doesn't matter, realistically speaking. More won't hurt though. I've run a few up to like 10,000 just to check. If you fire up the tensorboard you can watch the results. This is a feather mode with a bit higher ESR running to 600

1681149372712.png
 
You must log in or register to reply here.
Back
Top