Cab sim optimization and training speed question

[ThreepE0]

Hello again,

I'm not sure if this is the best place to put this (again, sorry) but I'm noticing after running quite a few trainings that some strange things happen when cab sim is on in the output wav file. I know that the simpler the better as far as input and output, but I did want to mention this in case it helps anyone else running the training, or in case a method that handles this better comes up.

I'm running some fairly long trainings (15000 epochs) as we speak. I notice that if I have the cab sim on, I get a prediction that is quite far from the original, both quieter and much much lower gain. I can share example files if that would help.

My suspicion is that the cab sim may be introducing some time-based distortion or delay, and this model might not care for that.

I'm finishing my 15000 epoch training now with the cab sim off, and will check back in once that's done. But I did run a 9000 epoch training and the results were pretty good with the cab sim off. In the included example wav files, I did notice that even with 9000 epochs, the string slide sound cut in and out a bit, like you'd expect from a bit crusher effect. It was an improvement over the default 1500 though.

All that being said, I wonder if there's a way to cut training speed down; I know the LSTM model is out there, but I haven't had great results with it yet. I do want to try it again with different parameters, but so far I've gotten better results with this model. LSTM is considered an older technology when compared to Wavenet, correct?

Without being very confident in what I'm talking about here, please bear with me: I'm thinking about learning rate; Is there a way to use a large learning rate at the beginning, and sort of pepper the field with guesses, then use the closes successful guess to train using a smaller learning rate? Maybe this is already built into how the model works, I don't know.

[GuitarML]

Wow, I have not tried that many epochs before. I've noticed that in direct comparisons between the SmartAmp and SmartAmpPro models, the SmartAmp (WaveNet) has a more natural sound. Although on higher distortion, WaveNet can't quite grab on to the signal, where LSTM seems to have an easier time with it up to a point. LSTM has been around longer.

I'm assuming the cab sim adds impulse response? That is like a time effect that would probably give pedalnet difficulty. The same as using an actual mic on an amp, but simulated. Feel free to share the wav files or resulting models!

Good guesses on the learning rate, what you are describing is called adaptive or dynamic learning rate, but it's not implemented here. That would be a good thing to look at, and I also suspect it would reduce the training times.

See this part of the pytorch documentation: https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate
and this: https://stackoverflow.com/questions/48324152/pytorch-how-to-change-the-learning-rate-of-an-optimizer-at-any-given-moment-no
one more: https://towardsdatascience.com/learning-rate-schedules-and-adaptive-learning-rate-methods-for-deep-learning-2c8f433990d1

[ThreepE0]

Hi Keith, Yes, right on the cab is an impulse response. Ok great, I think I'm starting to understand some of this stuff finally haha! Here is a dropbox link that includes a reaper session containing the inputs and outputs, as well as a couple frozen tracks with cabs added on afterwards. I also copied the .json and .ckpt files into the directory. That should be all that's needed to use the model right? https://www.dropbox.com/sh/f3bzctm9c47xx2t/AADKiJNuQifH4imS_hR_Bp12a?dl=0

…

On Fri, Feb 19, 2021 at 1:38 PM Keith Bloemer ***@***.***> wrote: Wow, I have not tried that many epochs before. I've noticed that in direct comparisons between the SmartAmp and SmartAmpPro models, the SmartAmp (WaveNet) has a more natural sound. Although on higher distortion, WaveNet can't quite grab on to the signal, where LSTM seems to have an easier time with it up to a point. LSTM has been around longer. I'm assuming the cab sim adds impulse response? That is like a time effect that would probably give pedalnet difficulty. The same as using an actual mic on an amp, but simulated. Feel free to share the wav files or resulting models! Good guesses on the learning rate, what you are describing is called adaptive or dynamic learning rate, but it's not implemented here. That would be a good thing to look at, and I also suspect it would reduce the training times. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#25 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA6OCQCAZTRGSPLGA7KULK3S72V2VANCNFSM4X4X4Z3Q> .

[GuitarML]

Nice! I tried out the model, I could tell it had difficulty training it, might have more luck on the high distortion with guitarLSTM/SmartAmpPro. Is that from Neural’s Granophyre plugin?

[ThreepE0]

Yessir that’s the plugin it came from 👍 I’ll try the LSTM tomorrow and check back in on how it goes. I just stumbled across this article and repo as well: https://musiclaboratorydotcom.wordpress.com/2019/01/05/a-deep-learning-approach-to-guitar-amp-modeling-part-1-intro-and-initial-attempts/ https://github.com/sdatkinson/neural-amp-modeler I ran a quick test and it seemed to hone in on the neighborhood of the sound relatively quickly. I'll be testing with that model as well, hopefully I can keep my GPU busy during work hours and check the results during breaks and in the evening. Stephen just updated the code base to use pytorch today. It's great to have such a sampling of different approaches, I think it's helping me absorb some of the concepts. What an exciting time to be alive haha!

…

On Sun, Feb 21, 2021 at 7:39 PM Keith Bloemer ***@***.***> wrote: Nice! I tried out the model, I could tell it had difficulty training it, might have more luck on the high distortion with guitarLSTM/SmartAmpPro. Is that from Neural’s Granophyre plugin? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#25 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA6OCQD67FULACC7AZ73OPTTAGRUZANCNFSM4X4X4Z3Q> .

[GuitarML]

Very cool, thanks for sharing! It’s such a fun application for ai, I expect lots of people will be trying out different variations on it.

[ThreepE0]

Hi Keith, I just wrapped up a training in PedalnetRT with 15000 epochs: https://www.dropbox.com/sh/nzfoweljztxba2b/AAAlFzLCPZ19QNFzmqI_Dt1-a?dl=0 ignore the misnamed 'grano' files, these are from the Neural DSP Gojira plugin. Also ignore my terrible skill in dialing in tone and playing haha Let me know what you think if you have a sec. I did notice that from about 10k to 15k (and most likely long before 10k actually,) the loss didn't really drop very much at all. I'm thinking this might be improved by using an adaptive learning rate, what do you think? I started looking briefly at pytorch_lightning and how to implement that, and it does look manageable at first glance, but I haven't dug into too much yet. Wondering if that would be a way to get the training time down as well. Listening to the output file vs the prediction, I can't hear any difference really until I add a cab. Once the cab is added, it becomes pretty evident that the bass/low end could be better.

…

On Sun, Feb 21, 2021 at 8:20 PM Keith Bloemer ***@***.***> wrote: Very cool, thanks for sharing! It’s such a fun application for ai, I expect lots of people will be trying out different variations on it. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#25 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA6OCQH5U6NMCSW2EZ5LVSDTAGWNNANCNFSM4X4X4Z3Q> .

[mishushakov]

@GuitarML can you enable discussions and convert this to a discussion?
https://docs.github.com/en/discussions/quickstart

would be helpful, thanks

[GuitarML]

That's a great idea, discussions are now enabled. I think the subject of this discussion has kind of evolved so I'll let @ThreepE0 start one and summarize where it will continue. I'm also going to enable discussions on the other repos.

[mishushakov]

cool, thanks
we could also use it to share models for now

also issue can be converted to discussions: https://docs.github.com/en/discussions/managing-discussions-for-your-community/moderating-discussions#converting-an-issue-to-a-discussion

[GuitarML]

Wow that’s easy, done! @ThreepE0 ill take a look at those training files as soon as I get a chance, thanks!

[ThreepE0]

Sounds good! Thanks again, and my apologies for blowing up your inbox

[GuitarML]

@ThreepE0 Hey I checked out your Gojira model, that's a fun one! 15000 epochs is a long time to train, and it's one of the heaviest distortion I've heard out of SmartAmp. The playing sounded really good to me, I don't know what you're talking about there. I think adaptive learning rate would reduce training time, and maybe improve accuracy some.