NAM: Neural Amp Modeler

Okay here's a little summary I've found when creating NAM captures;

1. The best captures are created when using real guitar parts as training signal.
2. Pushing the Input Gain in the Plugin results in a fuzzy blocking/farting type distortion unlike using the amp's gain control, so it's best to create captures for the MAXIMUM amount of gain you want to use when the Input is set to 0.0, yes, that means the captures has to be made for known audio interface dBu.

Here's a new capture of my modded high gain Peavey Valveking, it has more in common with a boosted Soldano/5150 after years of mods.
The capture is made for +12dBu inputs, that's 500mV=-15.8dBFS or 1.0v=-9.8dBFS.
Code: 
https://drive.google.com/file/d/1smium21Hm1P6ezdVRL7YwlmH1AdvPn6O

Settings are:
Bass: 7
Middle: 3
Treble: 7
Presence: 7
Resonance: 5
 
Welp, I'm out of time in Google Colab.
I've installed Anaconda3, almost 20GB.

Question, why does it barely use the GPU and instead the CPU is maxed? As far as I can see in the performance tab in Task Manager.

Anaconda3 sais it has found a GPU with cuda cores...

CPU load.png
 
For avoidance of doubt, my Anaconda install used CUDA out of the box, and the utilisation on my card hits 100% !!
 
James Freeman said:
Okay here's a little summary I've found when creating NAM captures;

1. The best captures are created when using real guitar parts as training signal.
2. Pushing the Input Gain in the Plugin results in a fuzzy blocking/farting type distortion unlike using the amp's gain control, so it's best to create captures for the MAXIMUM amount of gain you want to use when the Input is set to 0.0, yes, that means the captures has to be made for known audio interface dBu.

Here's a new capture of my modded high gain Peavey Valveking, it has more in common with a boosted Soldano/5150 after years of mods.
The capture is made for +12dBu inputs, that's 500mV=-15.8dBFS or 1.0v=-9.8dBFS.
Code: 
https://drive.google.com/file/d/1smium21Hm1P6ezdVRL7YwlmH1AdvPn6O

Settings are:
Bass: 7
Middle: 3
Treble: 7
Presence: 7
Resonance: 5
Click to expand...
All good @James Freeman ?
First forgive me for using a translator.
I really want to understand what you say about +12dBu = .5V = -15.8dBFS. What I've been doing: My interface is variable from +14dB to -18dB. I output a sinusoidal signal and measure the voltage until I get 3.08V which would be the +12dBu it hits the input clip. With this I calibrate the input gain. Until then OK. What I imagine you do now is keep the 0dBFS clipped signal coming from the interface to the DAW
(0dBFS(interface) = 0dBFS(DAW).
I don't understand where 15.8dBFS or 9.8dBFS comes in.
 
Orvillain said:
For avoidance of doubt, my Anaconda install used CUDA out of the box, and the utilisation on my card hits 100% !!
Click to expand...
Yeah it is definitely using the cuda cores, I just has to switch one of the graphs as suggested by Deadpan.

Frederico_Gomes said:
I output a sinusoidal signal and measure the voltage until I get 3.08V which would be the +12dBu it hits the input clip.
I don't understand where -15.8dBFS or -9.8dBFS comes in.
Click to expand...

Notice the negative sign: -dBFS in DAW.
When you feed the interface lower voltage, you will see minus dBFS in your DAW.

The 0.5v = -15.8dBFS is the result of 20log for +12dBu, ie. 20*log(0.5/3.084) <- google that line.
 
Last edited:
I've done more testing with higher gain on my amp.
When I turn the gain too high on my amp to the point I'm losing tightness and oomph with palm mutes, in other word "farting" or too much blocking distortion in the preamp, the capture will bake that in and no matter how I lower the input gain in the plugin the blocking distortion is present.
So, to strengthen my previous point of capturing with maximum amount of gain you would actually use with your amp before it starts to fart, and never go above the input dBu the capture was made for because the sound will also start to fall apart.

It seems that neural network based captures are sensitive to gain, preferably the capture is made for a specific dBu and your instrument input matches that dBu, then only attenuate from there.

In simple words:
1. Don't capture the amp when it sounds bad, it's unfixable later.
2. Don't push the capture beyond the real maximum gain it was captured with.
 
Last edited:
James Freeman said:
I've done more testing with higher gain on my amp.
When I turn the gain too high on my amp to the point I'm losing tightness and oomph with palm mutes, in other word "farting" or too much blocking distortion in the preamp, the capture will bake that in and no matter how I lower the input gain in the plugin the blocking distortion is present.
So, to strengthen my previous point of capturing with maximum amount of gain you would actually use with your amp before it starts to fart, and never go above the input dBu the capture was made for because the sound will also start to fall apart.

It seems that neural network based captures are sensitive to gain, preferably the capture is made for a specific dBu and your instrument input matches that dBu, then only attenuate from there.

In simple words:
1. Don't capture the amp when it sounds bad, it's unfixable later.
2. Don't push the capture beyond the real maximum gain it was captured with.
Click to expand...
Does this mean NAM doesn't clean up realistically at those extreme farty settings?

Hm... would it help if you used a training signal with a large dynamic range?
 
Ok, a question along these lines (unfortunately I can't test myself on this aging Mac Pro):
What happens when you throw rather uncommon stuff at NAM? Think along the lines of parallel amps (one clean, one driven). Or even some stuff impossible in the analog world, such as an amp with the overdrive going up (!) when you turn the input down (!), which (obviously just to a certain degree) is possible in, say, Guitar Rig.
Quite curious to know.
 
yeky83 said:
Does this mean NAM doesn't clean up realistically at those extreme farty settings?

Hm... would it help if you used a training signal with a large dynamic range?
Click to expand...

It already actually has pretty much the biggest dynamic range possible in that V1 signal - but one thing that I could see helping this (maybe) is by using the same signal in the middle again somewhere, only quieter. Also possibly some more rolled off guitar parts in the training?

One other is, when you train a model, your DI is actually comprised of two parts - DI signal is split at the end (hence why you want to put new data in the _middle_ of the v111 instead of at the end), and the last few seconds are used for testing at each iteration or epoch. The results of the test inform the model whether the new weights are good or not.

So I want to say logically that adding data to the "test" to reinforce this behavior as well as mid training might be helpful, but to be honest I'm not really that clear on how the ML training actually works - so I don't know if it would help. IIRC if you want to modify the test set you might have to run it manually or add a very small bit to the end?

Sascha Franck said:
Ok, a question along these lines (unfortunately I can't test myself on this aging Mac Pro):
What happens when you throw rather uncommon stuff at NAM? Think along the lines of parallel amps (one clean, one driven). Or even some stuff impossible in the analog world, such as an amp with the overdrive going up (!) when you turn the input down (!), which (obviously just to a certain degree) is possible in, say, Guitar Rig.
Quite curious to know.
Click to expand...

My guess is parallel captures won't work super well, just like most parallel amps IRL don't because of phasing and the overlapping frequencies. Blending seems to be a hard thing for ML to get right, but I'll admit I haven't tried it with NAM. The QC sucked at it though
 
northern_fox said:
just like most parallel amps IRL don't because of phasing and the overlapping frequencies.
Click to expand...

Hm, can't second that statement. I was using parallel amps for a while and it's been one of the more glorious setups of my "career". Only abandoned it because of the usual suspects (lots to lug around, lotsa cabling, etc.).
 
ToneX does a decent job of two amps at once. I did my VH4+JVM profile and it came out pretty good. Haven't tried NAM yet.
 
Orvillain said:
ToneX does a decent job of two amps at once. I did my VH4+JVM profile and it came out pretty good.
Click to expand...

Did run them with vastly different drive amounts? Think of almost clean and pretty distorted.
I could quite imagine that this would be something still throwing all capturing engines off.
 
Sascha Franck said:
Hm, can't second that statement. I was using parallel amps for a while and it's been one of the more glorious setups of my "career". Only abandoned it because of the usual suspects (lots to lug around, lotsa cabling, etc.).
Click to expand...

Mainly referring to digital stuff - with real amps I think it works better. You just have to be really careful about alignment with digital sources. It's a bit ironic to me that the QC could capture a whole amp pretty well, but could not get a simple tubescreamer as relatively correct because of the clean blend.
 
James Freeman said:
In simple words:

2. Don't push the capture beyond the real maximum gain it was captured with.
Click to expand...

Hey James !

Do you mean, leave the "Input" [or as I refer to it as the "Gain"] at its stock Captured Level as it doesn't respond well when increased beyond its Captures level (?)

Not sure if that's what you are saying (?)

Ben
 
The input gain is a simple flat volume increase for NAM and has the exact same effect as increasing the signal level going into the plugin (the same is also true for the tonex "gain"). However, just like with tonex - if you increase it a too much it definitely will not sound "correct". But I can usually use at least a few dB on it without pushing too far.

Do not turn the output up though - there is some form of limiting if you do that - I'm not actually sure why
 
Last edited:
northern_fox said:
Do not turn the output up though - there is some form of limiting if you do that - I'm not actually sure why
Click to expand...

The plugin output volume does not allow overflow, so it clips internally.
Most plugins will allow floating operation where they have basically infinite headroom until the final output bus in your DAW.

I think it is fixed but not released yet.
github.com

Don't clip output in DAW. (#168) · sdatkinson/NeuralAmpModelerPlugin@6432f52

Plugin for Neural Amp Modeler. Contribute to sdatkinson/NeuralAmpModelerPlugin development by creating an account on GitHub.
github.com github.com
 
BenIfin said:
Do you mean, leave the "Input" [or as I refer to it as the "Gain"] at its stock Captured Level as it doesn't respond well when increased beyond its Captures level (?)
Click to expand...

Until you know what dBu the capture is made for and your audio interface instrument input dBu, it means nothing.
When your capture and input dBu match, attenuate from there but never boost.
 
A little test I'll post results of tomorrow - turns out blending signals works quite well in NAM! this totally blew me away because I was expecting a "kind of" sounds like it result but it's substantially better than I had imagined. I did a darkglass blended pedal and a clean/high gain guitar amp for two tests. May also try fuzz + clean blend, or two high gain amps together.
 
You must log in or register to reply here.
Back
Top