NAM: Neural Amp Modeler

[northern_fox]

I've been getting better results too - there was a tweak to one of the sine sweeps also, and there are some updated calculations for delay and the added checks for whether something has changed from start to finish or the file is very noisy. I did a some testing comparing my modified v1 file and original v1 results with v2 and the new version, and have found the newer one to do just as good of a job ESR wise but easier/faster since it's shorter than my extended version was. There have been some minor code changes along the way including a learning rate typo since the last release that might have helped too.

I've also been running the LR and LR decay at half the default which has been pretty helpful results wise but takes a lot longer to train

 

[MirrorProfiles]

Made a pack of 1985 Mesa Mark III Black Dot models at various settings. I dialled them in with various guitars and different speaker types so hopefully they translate nicely no matter what hat you’re using. Mostly rock/metal/higher gain tones, its what I love this amp for most.

Audio Demo:

Free one here:

https://tonehunt.org/MirrorProfiles/db9d876a-7b0d-413c-a822-7968f9bc3dbf

50% until July 1st: https://mirrorprofiles.myshopify.com/discount/BX2JKX67FTDX?redirect=/products/nam-amp-pack-1-mark-iii

 

[EOengineer]

Made a pack of 1985 Mesa Mark III Black Dot models at various settings. I dialled them in with various guitars and different speaker types so hopefully they translate nicely no matter what hat you’re using. Mostly rock/metal/higher gain tones, its what I love this amp for most.

View attachment 8220

Free one here:

MirrorProfiles Mesa Boogie Mark III FREE | ToneHunt

MirrorProfiles Mesa Boogie Mark III FREE is a model by MirrorProfiles. These are famously the first production amps to follow the infamous IIC+. My particular amplifier has the export X105 power transformer, simulclass 6L6/EL34 power amp, reverb and EQ (and the EVM12L speaker). Captured at...

tonehunt.org

50% until July 1st: https://mirrorprofiles.myshopify.com/discount/BX2JKX67FTDX?redirect=/products/nam-amp-pack-1-mark-iii

I keep thinking of shooting some Tonex captures of my Blue Stripe (now modded to red stripe). The Mark III owns so hard.

 

[MirrorProfiles]

I keep thinking of shooting some Tonex captures of my Blue Stripe (now modded to red stripe). The Mark III owns so hard.

ToneX models take about 2 hours on advanced for me. With NAM, I can reamp them all first and the training takes 20 minutes per model (using Colab pro). Especially nice for when I’m using a cab as a load, and it’s loud in the room, I can do the reamping in the same session rather than spread out every 2 hours. For an amp like a Mark series where there are so many possibilities, using NAM really appealed more. I still want to make more ToneX packs of amps but for the time being, I’ll probably focus on the caveman amps until I can get the training times lower.

Would love to hear your Mark III too, nice to have some of their tones archived digitally in some way

 

[EOengineer]

ToneX models take about 2 hours on advanced for me. With NAM, I can reamp them all first and the training takes 20 minutes per model (using Colab pro). Especially nice for when I’m using a cab as a load, and it’s loud in the room, I can do the reamping in the same session rather than spread out every 2 hours. For an amp like a Mark series where there are so many possibilities, using NAM really appealed more. I still want to make more ToneX packs of amps but for the time being, I’ll probably focus on the caveman amps until I can get the training times lower.

Would love to hear your Mark III too, nice to have some of their tones archived digitally in some way

The training time is the only thing that’s stopping me from profiling my amps right now.



At some point I’ll stop being lazy and dig into NAM but Tonex is what I have on hand at the moment.

It sounds like your profiling strategy is similar to what I’d planned. I was going to hook up my various heads through a 1960A in my basement with my Captor in line just to grab the direct signal while getting the authentic impedance interaction from the cab. I might also isolate and mic the cab to get the whole rig too. Like you said - very time consuming.

Really looking forward to checking out your captures.

 

[quadgear]

The training time is the only thing that’s stopping me from profiling my amps right now.

Why does that matter? Can you set it to run overnight and get through a bunch in a week or so?

 

[EOengineer]

Why does that matter? Can you set it to run overnight and get through a bunch in a week or so?

yeah I’m thinking cranked amps blasting in my house all night is going to interfere with some of the other night time activities I enjoy…like sleeping.

 

[quadgear]

yeah I’m thinking cranked amps blasting in my house all night is going to interfere with some of the other night time activities I enjoy…like sleeping.

I was thinking just the computational part of training, not the inject sound and record part....

 

[Foxmeister]

I was thinking just the computational part of training, not the inject sound and record part....

When it comes to ToneX, they are sort of the same thing - you can't do a boatload of re-amps and then train all the results in the same way you can with NAM.

 

[CoreToneCapture]

Normalizing the track before export will only really change the final capture volume, but I normalize all of mine to -3 dB true peak for consistency. The normalization I mentioned earlier is just for plugin playback volume between captures, and has nothing to do with actually capturing

What are you normalizing with? I'm not sure Studio One can Normalize to a db level...is there better tools to batch normalize a bunch of captures i already have? Thanks.

 

[James Freeman]

 

[northern_fox]

He's pretty out of date and a little wrong on some of what he says here and definitely has his personal preference towards modeling vs capturing, but overall an okay review. He also seems a little annoyed that he's doing it at all haha. And kind of missing the point that it's not a commercial thing and certainly was never an expectation that your "average user" knows how to navigate colab and github. In the beginning you basically had to know python to use it at all, but it was fine because it was really meant for people that were interested as opposed to "try a new plugin!" for general audience. It's grown from there a lot, but the main goal still isn't to create a product that replaces other products. It exists for fun and as a tool, but the hype does its thing.

I'm really happy to see what it's become - but it's still largely a one man free time job that he works on for fun

 

[NightWalker]

Made a pack of 1985 Mesa Mark III Black Dot models at various settings. I dialled them in with various guitars and different speaker types so hopefully they translate nicely no matter what hat you’re using. Mostly rock/metal/higher gain tones, its what I love this amp for most.

View attachment 8220

Audio Demo:

Free one here:

https://tonehunt.org/MirrorProfiles/db9d876a-7b0d-413c-a822-7968f9bc3dbf

50% until July 1st: https://mirrorprofiles.myshopify.com/discount/BX2JKX67FTDX?redirect=/products/nam-amp-pack-1-mark-iii

thank you for the free capture,sound pretty good,is there any more discount coupons soon?

 

[2112]

@polyeffects just released a beta firmware for Beebo that's running NAM captures. Currently just a few captures they've made themselves but apparently next firmware will allow users to load their own captures. Nicely organised with pictures of the amps and categories for each capture like amp channels/boost switches.

Pretty CPU intensive though - NAM block, cab and 2 delays hits 95%.

 

[James Freeman]

Pretty CPU intensive though - NAM block, cab and 2 delays hits 95%.

Very interesting.
I wonder if they are running the light (feather) captures or the standard?

 

[BenIfin]

Very interesting.
I wonder if they are running the light (feather) captures or the standard?

From the Beta FW Page:-

' ..... The NAM module only play Nano captures and they are still quite heavy. There should be enough CPU for a mono cab and a delay or two, but running reverb at the same time won’t work ..... '

Ben

 

[yeky83]

What are nano captures, is it lighter than feather? How’s the tonal difference of all the capture types, quite noticeable?

 

[James Freeman]

"lite" models will run approximately 1.5 times faster than "standard".
"feather" models will run more than 2 times faster than "standard".
"nano" models will run about 2.5 times faster than "standard".

I don't know how it affects the accuracy of the capture though.

Add "nano" preset for WaveNet and LSTM by mikeoliphant · Pull Request #279 · sdatkinson/neural-amp-modeler

Also skew the LSTM presets toward simpler architectures. Based on my testing I think the set of LSTM presets is pretty good, but I'm happy to discuss alternatives. In my opinion, it really isn't wo...

github.com

 

[Orvillain]

@northern_fox

When training with NAM I see this message:

You are using a CUDA device ('NVIDIA GeForce RTX 4090') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/gen...ision.html#torch.set_float32_matmul_precision

Where does that variable get set, and does changing it speed up training times?

 

[James Freeman]

That's normal, it is already using the Tensor cores.
I got the same message with a RTX 3070 and been told it's normal.

As a side note, the training process is really power hungry, the entire system is almost maxed.