NAM - Neural Amp Modeler: cracking the code! (How NAM works & how to increase the profile quality)

Подписаться 36 тыс.

Просмотров 9 тыс.

50% 1

In my last video about NAM, where I shared how to obtain the best digital copy of our amps, I was basically trying to increase the quality of my profiles just increasing the epochs. But then, I received some comments, and one actually from Steve Hackinson himself, suggesting that I can obtain much better performances not only increasing the epochs. Well, as I love challenges, I started studying the NAM python source code and well, I have been able to increase the quality of my profiles by more than 60% !!!
▶ Follow this link if you want to purchase gear and help the channel:
- amp modelling pedalboards: thmn.to/thocf/xkwpquisc9
- pedals: thmn.to/thocf/hckhu3ukgo
- tube amps: thmn.to/thocf/dq0y6nqi99
▶💽 Link to a playlist of my songs: • Leo Gibson Songs
▶ Check out my tuner: Cherub WST-905Li: bit.ly/3QkxO9B
▶🎸🔉 If you are interested in my IRs, please follow this link: www.leogibsonguitar.com/shop-...
▶🎸🔉 If you are interested in my ToneX Profiles, please follow this link: www.leogibsonguitar.com/shop-...
▶ if you want to support the channel: www.paypal.com/donate/?hosted...
▶ Check out my first LEO GIBSON BAND song here: • Leo GIbson Band - #1: ...
▶ Check out my second LEO GIBSON BAND song here: • Leo Gibson Band- #2: O...
-----------------------------------
OTHER VIDEOS:
▶ Pedalboards demo review:
G2 x Four: • Zoom G2x Four / G2 Fou...
NUX MG-400: • NUX MG 400: demo and r...
NUX MG 30: • NUX MG 30: Demo and Re...
NUX MG 300: • NUX MG 300: Demo & Rev... BOSS GX 100 demo and review: • BOSS GX 100: Detailed ...
Valeton GP 200 LT: • Valeton GP 200 LT: dem...
Valeton GP 200: • Valeton GP 200: demo a...
Valeton GP 100: • Valeton GP 100: demo a...
Ampero Stomp II: • Ampero Stomp II: demo ...
Ampero Silver Edition: • Ampero Silver Edition:...
Fractal FM3: • Fractal Audio FM3: Dem...
Neural DSP Quad Cortex: • Neural DSP Quad Cortex...
HeadRush MX5: • Headrush MX5: Demo and...
Pedalboards Latency: • The most extensive amp...
Flamma FX-200: • Flamma FX 200: demo an...
Kemper Stage: • Kemper Profiler Stage ...
POD GO: • LINE 6 POD GO: Demo & ...
Zoom G11: • Zoom G11: Demo and Rev...
Zoom G6: • ZOOM G6: Demo & Review...
NUX MG 300: • NUX MG 300: Demo & Rev...
Mooer GE 300 lite: • Mooer GE 300 Lite: Dem...
Mooer GE 250: • Mooer GE 250: Demo & R...
Mooer GE 150: • Mooer GE 150: Demo / R...
▶ Modeling pedalboards comparison videos:
NAM vs Quad Cortex vs ToneX vs Kemper: • NAM vs ToneX vs Quad C...
Quad Cortex vs Fractal FM3: • Neural DSP Quad Cortex...
Quad Cortex vs Helix: • Neural DSP Quad Cortex...
Quad Cortex vs Kemper: • Neural DSP Quad Cortex...
BOSS GX 100 vs POD GO: • BOSS GX 100 vs Line 6 ...
BOSS GX 100 vs GT 1000 tech spec compared: • BOSS GX 100 vs GT 1000...
BOSS GX 100 vs GT 1000 sound comparison: • BOSS GX 100 vs BOSS GT...
HeadRush MX-5 vs NUX MG-30: • HeadRush MX 5 vs NUX M...
HeadRush MX-5 vs POD GO: • Headrush MX 5 vs Line ...
HeadRush MX-5 vs Ampero: • HeadRush MX 5 vs Hoton...
HeadRush MX-5 vs Mooer GE 250: • HeadRush MX5 vs Mooer ...
NUX MG 300 vs Valeton GP100 vs Mooer GE 150 vs Harley Benton DNA: • NUX MG 300 vs Valeton ...
NUX MG30 vs Mooer GE 200 vs Ampero One: • NUX MG 30 vs Ampero On...
FM3 vs Kemper vs Helix vs Pod GO vs NUX vs Valeton vs Mooer vs Soldano: • FM3 vs Kemper vs POD G...
BOSS GT 1000 core vs HX Stomp: • BOSS GT 1000 core vs L...
Zoom G11 vs Mooer GE 300: • Zoom G11 vs Mooer GE 3...
Fractal FM 3 vs Kemper Stage: • Fractal audio FM 3 vs ...
POD GO vs GE 250: • Line 6 POD GO vs MOOER...
GT 1000 vs Line 6 Helix: • BOSS GT 1000 vs LINE 6...
Line 6 Helix vs Headrush: • Line 6 HELIX vs Headru...
VALETON GP 200 vs POD GO: • Valeton GP 200 vs Line...
VALETON GP 200 LT vs HX Stomp: • Valeton GP 200 LT vs L...
VALETON GP 200 (LT) vs MOOER GE 300 (Lite): • Valeton GP 200 and GP ...
GT 1000 vs GT 1000 core: • BOSS GT 1000 vs GT 100...
GT 1000 vs FM 3: • Fractal FM3 vs BOSS GT...
GT 1000 vs HEADRUSH: • BOSS GT 1000 vs HEADRU...
GT 1000 vs GE 300 vs HELIX vs HEADRUSH vs PLEXI: • GT 1000 VS GE 300 VS H...
GE 200 vs GE 250 vs GE 300 vs GE 150: • MOOER GE 150 vs GE 200...
GE 250 vs Hotone Ampero: • Mooer GE250 vs Hotone ...
GE 250 vs HX Stomp: • Mooer GE 250 vs Line 6...
GE 250 vs Gigboard: • MOOER GE 250 vs HEADRU...
GE 200 vs Ampero One: • Mooer GE 200 vs Ampero...
GE 250 vs GE 300: • MOOER GE 250 vs GE 300...
GE 150 vs ZOOM G1X Four: x • Mooer GE 150 vs Zoom G...
BIAS AMP 2 VS AMPLITUBE 4 VS SOFTUBE VS FORTIN vs PLEXI: • BIAS AMP 2 VS AMPLITUB...
--------------------------------------------
⏰Timecodes:
0:00 - Intro
0:54 - NAM Network parameters explained (dilation, epoch, learning rate, etc)
3:36 - Can we increase the quality of a profile?
11:02 - Simulations results changing many network parameters
#NAM #NeuralAmpModeler #NAMSetting

Видеоклипы

Опубликовано:

6 авг 2024

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 172

@LeoGibsonGtr Год назад

In this video I check out if the ESR has really on impact on the tone and feel, and you can also find the link to the profiles: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-oZpb3_klUpo.html

@IamMusicNerd Год назад

I am blown away by your attention to detail, and how well you explain everything in laymen’s terms. I learned a lot. Thank you!

@LeoGibsonGtr Год назад

Thank you sooo much for such a great compliment!!! I'm really happy to have been of some help!!!

@mCKENIC Год назад

Thank you for the extensive and thorough tests Leo,. Very, very interesting!

@LeoGibsonGtr Год назад

You are welcome and I'm happy that you have appreciated the info!!

@eduardhausauer3766 Год назад

Fantastic Video! Thanks a lot for the work you've put in to share your observations with us!

@LeoGibsonGtr Год назад

You are welcome and I'm really super happy that you like the video!!!

@amalgamaudioLV Год назад

Wow, extremely well done video, informative, to the point and accessible! Thank you so much for the time and effort invested into this. Subscribed!

@LeoGibsonGtr Год назад

Thank you so much, I'm really happy that you have appreciated the video...and your subscription really means a lot to me! Thanks!

@WholeLottaBulldog Год назад

This was a fantastic explanation, Leo. Thank you very much my friend.

@LeoGibsonGtr Год назад

You are welcome and I'm really happy that you have appreciated the video!!!!

@electrifried Год назад

Congratulations and many thanks Leo!!Awesome video!!

@LeoGibsonGtr Год назад

You are welcome and I'm really happy that you like the video!!!

@digitalchris6681 Год назад

Fantastic video! You're really pushing the future of modelling forwards. I was only going to use other people's models, but now can't resist having a try at modelling myself (using your parameters of course).....

@LeoGibsonGtr Год назад

Thank you so much, I'm happy that you like the video!!! Yes, when you start learning how it works....it becomes very fascinating ... I start deep diving and learning about neural networks...it's incredibly interesting...and fun....

@goodwill559 Год назад

I've been following your channel and can't believe you have time to produce content, delve into the tiniest detail of products, write new tunes and still have time to practice ! You are not just Leo, you are Leonardo da Vinci !

@LeoGibsonGtr Год назад

😀...thank you so much for such an amazing compliment!!! Really appreciated! Let me say, that the real Da Vinci is Steven Atckinson that has produced this amazing plugin and training procedure...kudos to Steve.... Thank you!

@goodwill559 Год назад

@@LeoGibsonGtr and always humble too. Yes, Steve, thank you too if you are reading for your very valuable contribution !

@LeoGibsonGtr Год назад

Thank you!

@robertodepetro1996 Год назад

Mostruoso per contenuti e per montaggio video! Sei veramente forte.

@LeoGibsonGtr Год назад

Grazie mille, sei veramente gentile!!! Super apprezzato!

@thecaveofthedead Год назад

Very impressive stuff. Looking forward to hearing about the optimum training settings and also looking forward to your Aida-X video.

@LeoGibsonGtr Год назад

Thank you!!1 I'm studying the AIDA solution...I hope to be able to train on my pc soon...so that I can have a clear overview of the whole solution...

@TheManfisman Год назад

That was super interesting! Thank you for sharing this

@LeoGibsonGtr Год назад

Thank you so much, I'm really happy that you have found the video interesting!!!

@andivax 3 месяца назад

You are The Legend, Leo. Nerd legend, but still 😊 Thank you! 💛💙

@LeoGibsonGtr 3 месяца назад

You are welcome!!! I'm really happy that you appreciate the videos!!!

@jwright8838 Год назад

Great video, Leo. Lots to think about.

@LeoGibsonGtr Год назад

Thank you so much, I'm really happy that you have appreciated the video!!!

@jcoulter43 Год назад

My brain hurts after listening to this You somehow made it interesting though. Can't wait to see how NAM progresses. God bless and rock on!

@LeoGibsonGtr Год назад

😀👍...pretty nerdy stuff here...but I think also very fascinating...I had a lot of fun trying to understand all this staff! Thank you and God bless you too!!

@picksalot1 Год назад

Fascinating deep technical dive into what is going on under the hood of NAM. Awesome that Steven Atkinson himself commented to you! 👍 It might be useful to start compiling some numbers about how much time average End Users are willing to spend creating NAM Profiles. My guess is a maximum of 15-20 minutes. Then compare that to best quality obtainable within that time frame. And finally compare the "audio quality" results to determine at what point the differences typically become inaudible. I have a pragmatic question. What "file format" does NAM produce for their IRs? The reason I'm asking is that if they are in ".wav" format, they could be imported into a hardware Modeler, like the HX Stomp. That might make them useful/portable outside of Computer Standalone, DAW, and VST environments. Thanks for all you hard work, and developing and sharing your impressive skill set. 👏

@LeoGibsonGtr Год назад

Thank you so much for sharing your thoughts and for the kind words!!! NAM store its profiles in what is called ".nam" format. The NAM plugin which can load and read these files is a VST plugin (or similar format), therefore so far it cannot run on our standard pedalboards...I hope that someone will implement it!

@dimitrisgakis9206 Год назад

In the Neural Amp Modeler (NAM), NY refers to the number of points used in the Fast Fourier Transform (FFT) algorithm. A higher NY value means that the audio signal being analyzed will be divided into more frequency components, allowing for a more detailed analysis of the spectral content. The default value of NY in NAM is 8192, which means that the audio signal is divided into 8192 frequency components. However, users can adjust the NY value to achieve more detailed or less detailed frequency analysis, depending on their specific needs and hardware limitations. For example, if an audio signal has a frequency range of 0 Hz to 24 kHz and a NY value of 24000 is used, the signal will be divided into 24000 frequency components, allowing for a very detailed analysis of the spectral content. However, a higher NY value also means more computational resources are required, so it's important to balance the desired level of detail with the practical limitations of the hardware being used. (source : ChatGPT)

@LeoGibsonGtr Год назад

Thank you so much!!! Actually, I thought that lower NY value increases the quality...at least this is what I have experienced. With lower values, the training procedure take longer and provides with better ESR....I will try to increase the value... Thank you once more!

@dimitrisgakis9206 Год назад

@@LeoGibsonGtr No my friend, I thank you for your amazing work!!! I would like to remind you that the information provided by ChatGPT may not always be accurate. Perhaps we could explore the idea of renting computational power through cloud computing and go crazy with the training settings. As a fellow nerd, I am very interested in this experiment and eager to see where it takes us. Keep up the great work!

@LeoGibsonGtr Год назад

Thank you soooo much Dimitri!

@dimitrisgakis9206 Год назад

@@LeoGibsonGtr Cheers! I wanted to let you know that I tried using your best resulting settings, but encountered an issue when attempting to adjust the "channels" parameter. I want you to know when I utilized the "Feather" train mode with the same settings as you suggested (minus the "Channels" parameter, with NY set to 24000 and used many Epochs about 10k), I achieved better results than the Standard training, I suggest exploring the lighter training modes as well. Could you please provide some guidance on how to modify the "Channels" parameter? Thank you!

@LeoGibsonGtr Год назад

Hi Dimitri, thank you, I will try with your suggestion. The trick is that the channels in the first layer has be equal to the input size of the second layer...it's kind of the first layer is an input for the second one.

@lirojhonson 6 месяцев назад

Respect 🙌, I followed your channel for GAS matters but now also for data science divulgation !

@LeoGibsonGtr 6 месяцев назад

Thank you so much for such great compliment!!!!

@rubenmedina5050 Год назад

You rock Leo, thank you for this information

@LeoGibsonGtr Год назад

You are welcome and I'm happy that you have appreciated the video!!!

@Fender308 Год назад

Very detailed explanation and your thoughts can improve tone capturing! Good work!

@LeoGibsonGtr Год назад

Thank you so much, I'm really happy that you have appreciated the video!

@MaxNMyles Год назад

I always seem to land on your videos. Great stuff!

@LeoGibsonGtr Год назад

Thank you so much, I'm happy that you like the video! ...and I'm also happy that the YT algorithm suggests my videos....😀

@zuider77 Год назад

This technology is astounding. It just about ranks with the discovery of the wheel. What will be next? And the fact it's offered for free is staggering. To think I almost bought a K*mper last year!

@LeoGibsonGtr Год назад

Yes, I do agree....it is really super fascinating!!!

@danthegeetarman Год назад

AMAZING video Leo! 🙏🙏🙏

@LeoGibsonGtr Год назад

Thank you soooo much, I'm really happy that you like the video!!!

@erickgarcia1262 Год назад

You are the best Leo, good job 👍

@LeoGibsonGtr Год назад

Thank you sooooooo much!

@topavelka Год назад

OMG . . . you crazy :-) i am watching all . . . good work Leo!

@LeoGibsonGtr Год назад

😀...thank you really appreciated!

@TheDenwww Год назад

Thanks Leo!

@LeoGibsonGtr Год назад

You are welcome!

@Schlumpf.Meister 7 месяцев назад

Leo, this is very interesting. A few comments in addition to your descriptions: - Batch size is not the size of the batches, but the number of batches computed in separately (and potentially in parallel). - Learning rate should be adjusted with respect to loss conversion (which unfortunately, NAM does not report continually) and not with regard to batch size. - Channels is the number of connections between two layers. By definition, a CNN will have one input into the 1st layer (it is reading a sequence sample by sample), the output of which is "channels" wide into the next layer. - This is where NY comes into play: From my superficial code analysis, this is the number of outputs generated per "datum" (it generates an output vector of samples). Larger NY requires less computing power but conflicts with batching. - The "gated" parameter adds gates (a concept from recurrent neural networks) to the layers thus increasing their size. Upon activation, the selected activation is combined with a sigmoid. - I also think that activation would be interesting to experiment with, in particular when generating models for slower computers, as the default activation by Tanh is computationally expensive, whereas ReLU would much less costly (my first quick test has led to good results even with ReLU) I am wondering if any of the hyperparameter adjustments had positive impact on the aliasing problem you had stated earlier? Training a pre-trained network is identical to running more epochs.

@LeoGibsonGtr 7 месяцев назад

Thank you so much....very interesting! As far as aliasing is concerned....I was not able to find any remedy...I read some post on line (I don't remember the links) from which it was evident that also Steve Atckison himself was searching for a solution. I'm not checking NAM since las march....so maybe now NAM has improved in terms of aliasing...I should check it out again. Thanks!

@Schlumpf.Meister 6 месяцев назад

@@LeoGibsonGtrAFAIK there was no relevant change in the core algo, most of that code seems to be from pre-2023. I am working in research, maybe I will issue a thesis topic on this to find root cause and remedies other than an anti-aliasing filter or oversampling. I am very sorry for your loss.

@LeoGibsonGtr 6 месяцев назад

Thank you!

@imcrazedandconfused Год назад

Excellent video! Very well investigated, nicely explained details, I love it! A very interesting thing to investigate would be, how indistinguishable a model is by the human ear in practice. I am not sure, how the null-test LUFS relates to this. I guess, that the k-weighting *should* relate to this directly, but human perception can focus on certain areas or be irritated by single frequencies quite easily, I guess. Same for transients! I am personally not deep enough into this, probably, but it would shed a light, how much precision is actually enough for whatever use of audio. I have the strong feeling, that well made models in Proteus, NAM and AIDA-X actually are already in the range of quality, where it would not matter in nearly every mix, and there could be rules of thumb, below which null-test LUFS further precision is not reasonable at all. (Things might only matter then, when several networks are chained, not paralleled). Just now I investigate AIDA-X, which just reached the surface and is a RTNeural based neural network amp modeler that is based on a similar model as Proteus (LSTM, with same codebase) but uses 48kHz and actually runs on the MOD Dwarf (commercially available ARM-based pedal running the MOD environment to make signal chains) and potentially on other hardware / RaspberryPis with soundcards. It is still in very early stage and just officially out, but has some nice features, like several pre-defined model configurations for predictable cpu load on the hardware pedal. And, since it is trained in 48kHz, you can use the amp capture audio that you use for NAM also directly for training of AIDA-X models. Exciting times we live in.

@LeoGibsonGtr Год назад

Hi thank you so much for your feedbacks and I'm happy that you have appreciated the video!! Yes, I'm also trying AIDA X, the problem here is that I'm not able to run it on my PC, but only on google...and, as you know, I would like to be able to try all the aspects of a solution before releasing a video. You are right about the difference between ESR values and actual human perception. Actually, I have tested to profiles with almost the same ESR, and it happened that I like more the one with worser ESR....I don't know if it is just "impressions", but sometimes I "feel" that a profile sounds/responds better, even if on paper it should be worse...I have to further investigate this aspect, as I don't have scientific proofs...but just sensations.. Thank you!

@imcrazedandconfused Год назад

@@LeoGibsonGtr I am working on an offline solution for PC for myself and got it basically working, although it misses still a few bits and bytes (analysis part of the colab script, but training and export of the model seems to work already). I will opensource when it is working.

@LeoGibsonGtr Год назад

Cool...let me know!

@GranulatedStuff Год назад

thanks Leo !

@LeoGibsonGtr Год назад

You are welcome!!

@marcelo_campitelli Год назад

Amazing video as usual Leo! Very nice understanding of eveything going on under the hood! How you could improve that number is a good question. The cleaner the captured signal you get is, the better the iterations would be, so lowering the noise floor of your reamp to the maximum would help. You said you captured with the cab, I think only capturing the amp will improve the ESR as well. Keep making great videos!

@LeoGibsonGtr Год назад

Thank you so much, I'm happy that you like the videos!!! You are right: I have profiled also the amp without the cab, obtaining better result in terms of ESR...but to my touch and ears, the profile of the entire rig "feels" better... Nice suggestion about the noise floor: do you know how I can do it? Thank you!

@marcelo_campitelli Год назад

@@LeoGibsonGtr using a good reamp box helps in that sense, but you probably use a good reamp box... just trying to cancel as much ground loop noise as possible, having the electric power outlets in your house/studio filtered so the 50/60hz hum/noise is as little as possible, things like that... another idea would be running every cable in your chain with balanced outputs (to reduce the noise to the maximum) but I'm being extremely picky here... anyways, you are already getting incredible results in my opinion

@LeoGibsonGtr Год назад

Ok thank you for your suggestions!!

@imcrazedandconfused Год назад

@@LeoGibsonGtr For this type of noise reduction, letting a plugin learn the actual noisefloor of the amp with an "empty" reamp signal and clean the captured audio of the modeler test signal with this noise profile you got would do the trick. It should work to reduce static noise and hum, ReaFir, for example, would be a tool that could be used for it. Izotope RX would offer more advanced options.

@LeoGibsonGtr Год назад

Very interesting...thank you!

@user-os3vt8lm6g Год назад

We all like challenges! Great video! You do really great job trying to get best model! The one important rule in training neural nets - try everything, nobody knows what will best in this concrete training data. You have very powerful computer if you able to run this models in realtime. My daw crashes in same complexity as in your 2-4 trial. I think if you will create only amp model without cab and then use this model + your captured cab IR then you will be able to get better LUFS. And can you feel the difference in quality between this trials just playing in real time? Is the quality and feel improving noticable? And please give us the opportunity to join this challenge, share your output.wav!

@LeoGibsonGtr Год назад

Hi, thank you so much, I'm happy you appreciate the video! Yes, I have a 4090, that allows to do these tests...otherwise it would have been impossible... Yes, I have trained also only amp...and generally I obtained better ESR...but to my hear and to my touch the profiles with the entire rig sounds better. Yes, I wanted to share the changed source code and the files...but finally I ended up out of time....maybe I will do a video showing the profile in action and the showing the source code...and the all the files...

@InTheSh8 Год назад

A big mistake that many make is to run their DAW project in 44.1kHz. I never made a profile, but I loaded them and my THR10 is not capable to run in 48kHz. So, the playback is slightly off. In fact so miniscule that many other users in the facebook group insisted that it is negligible. But it is a fact. The DAW project must be 48kHz, since NAM expects that as an input signal! I want to try to capture my amps and pedals, too. But I need a reamp-box.

@LeoGibsonGtr Год назад

Hi, thank you for sharing!

@michaelcreel106 Год назад

I have been working with nets similar to this, for a different purpose. The batch size is an important parameter, in my experience. A larger batch size can make more efficient use of the GPU, speeding up training. There is a limit though, too large a batch size will cause the GPU to run out of memory. Also, too large a batch size can cause the net to focus on a local minimum of the loss function. I think that taking one of your best configurations and trying batch sizes of 32, 64, 128, 256, for example, would have a good potential to improve results. One other thing, does NAM training use a training/testing split of the data? I haven't looked at the source code. If not, that would be another thing to experiment with, but considerably more work to implement. Thanks very much for doing this, it's very interesting, and NAM is really a great project.

@LeoGibsonGtr Год назад

Thank you so much for your message Michael. I have read that the batch size is also related to the Learning Rate: the small is the learning rate, the smaller should be the batch size and vice versa. As I have reduced the learning rate, I have also reduced the batch size...I will also check out your suggestions...Thank you!

@michaelcreel106 Год назад

@@LeoGibsonGtr A low learning rate and a small batch size are both ways to avoid converging to a local minimum of the loss function, but they both increase time needed to train. Perhaps this is needed for modeling sound well, I haven't been able to experiment with it myself. If it is possible to increase the batch size without encountering that problem and without running out of memory, then training will be significantly faster, making it easier to use more epochs or adjustments to the configuration of the net. Thanks again for your experiments.

@LeoGibsonGtr Год назад

You are welcome and thank you soooo much for these very interesting info!!!

@SHOWWHITE Год назад

Respect for your time and great results! I'm just wondering how you can achieve such a short period of training time. For the standard 1000 epochs training, it usually takes me more than an hour. Thank you!

@LeoGibsonGtr Год назад

You are welcome and I'm really happy that you appreciate the video! The epochs training time heavily depend on the graphic card and I have a 4090, which is very fast. I don't know your PC configuration...but the graphic card has a big impact on the epoch time. Thank you!

@jonathanarnold2225 Год назад

You are a wizard :)

@LeoGibsonGtr Год назад

😀 Thank you!

@notalkguitarampplug-insrev784 Год назад

You could change the wav training file, adding sine sweeps from 20Hz to 20Kz to make your model learn to avoid aliasing noises (I think it’s a big part of the null test deviation)…

@LeoGibsonGtr Год назад

Hi, thank you for sharing....I will try...

@imcrazedandconfused Год назад

This would be a great improvement for many applications of not only NAM! Great idea, since many plugins based on neural networks have been criticized in this regard!

@boshi9 Год назад

This won't help with aliasing as the model operates only within the digital domain.

@notalkguitarampplug-insrev784 Год назад

@@boshi9 neural networks can be trained to cancel aliasing noises, as it can be predicted

@boshi9 Год назад

@@notalkguitarampplug-insrev784 How?

@GranulatedStuff Год назад

I get crackles on 15-19. I can tell that they give me noticeable extra depth in the midrange. Number 9 seems the best compromise on my PC (12700k @ approx 3.8ghz 48 gigs ram) During my tests using Cantabile as the host I was getting HUGE memory leakage. After 30 minutes I was up to 9GB !

@LeoGibsonGtr Год назад

Thank you so much for sharing! I'm impressed thank such a powerful machine struggle with those profiles....interesting insight! May I ask you if you observed the memory leakage with all the profiles or only with some of them? Thank you!

@GranulatedStuff Год назад

@@LeoGibsonGtr oops sorry I missed your reply... The memory leakage only occurred when using an IR in the IR slot. Just checked using NAM v0.7.3 and it appears to have been fixed

@Orta-Studio Месяц назад

Hello Leo, I want to thank you for your remarkable work and the quality of your videos. I would like to know how to access the changes in the code lines and improve the quality and rendering of my Neural captures by experimenting to achieve a result close to your ESR. Is this a different platform, or do you do this on Google Colab or the Trainer software? I'm on "easy" mode right now but i really want to learn about that advanced code thing 🤩 Greetings from France !

@LeoGibsonGtr Месяц назад

Thank you so much, I'm really happy that you appreciate the video! I have installed ANACONDA on my PC and all the NAM file. You can simply access the source code, change it and launch again the NAM trainer. I dont' use the google collab. Thank you!

@Orta-Studio Месяц назад

@@LeoGibsonGtr Thank you very much, Leo ! I will receive my Nam player soon, and I am really excited !

@LeoGibsonGtr Месяц назад

You are welcome!!!

@fab672000 11 месяцев назад

Great video where did you find the documentation for these?

@LeoGibsonGtr 11 месяцев назад

Google....I found some articles about A I amp modelling, some other about A I technique more in general, and some other about WaveNet, and then I especially experiment myself. Thank you for your message!

@williammclemore5815 Год назад

Very interesting treatment of the digital side of NAM. What you did not address is how NAM uses the Analog to Digital algorithms to capture the analog characteristic of the amps/cabs. After all at the heart of NAM is 40 to 50 year old technology. That is the A/D, D/A conversion. Now NAM does a great job of capturing the characteristics of the amps and cabs and then doing a recreation from the digital (0s and 1s) data back to analog. We need to remember that all known life forms that have hearing, hears in analog, No known life form hears in digital. So my suggestion is to make a video addressing the A/D D/A conversion of NAM, which is at the heart of NAM.

@LeoGibsonGtr Год назад

Hi, thank you for sharing! I'm not sure whether I have properly understood your point or not. The A D A conversion is done by my audio interface and not by NAM, therefore what you mean by how NAM manages the A D A conversion? Thank you!

@williammclemore5815 Год назад

@@LeoGibsonGtr NAM can be either a stand alone or a VST3 plugin. If the A D A conversion is done outside of NAM then why is there a need for the 48K sample rate. The implication is that NAM is sampling the analog signal at 48k and then converting the analog signal to digital. Regardless of where it happens or from what hardware or software it is done by, it still is done. Both going into NAM and coming out of NAM. The point is the the amp captures, that we load up, are done from analog signals that have been converted to digital. As I said that technology is 40 to 50 years old, going back to the 1980's

@LeoGibsonGtr Год назад

Hi William, the conversion A D A is done for sure by our audio interface. As far as I have understood, NAM has to translate the already digital wave form of the input and target signal, to a data set that the neural network can process. But this translation all happened in the digital world. Still from what I understood, the problem to avoid here is aliasing, I mean the higher the freq rate is, the less should be the aliasing. In my opinion, a cool test to do is to work on a 192 kilo hertz data set...but this require to change NAM code more deeply....actually I don't know if I would be able to do it. Thank you!

@williammclemore5815 Год назад

@@LeoGibsonGtr My audio interface does not let me choose a sample rate. Now my DAW does, but my DAW is not a audio interface. My audio interface is a mixer with USB out and has no controls to set sample rates. Now both NAM and my DAW has the ability to set and change the sample rate. And NAM reports that it works bets at 48000 samples a minute (48K). Like I have said the analog signal has to be converted either by NAM, the DAW, and external A D box, or something else. And one would usually have the option of selecting the sample rate.

@williammclemore5815 Год назад

That should be 48000 samples a second not minute.

@slowblow Год назад

Leo, do you know if it is possible to extract the .nam model file to any universal format for pytorch, tensorflow or keras in instance for creating own vst plugin in matlab?

@LeoGibsonGtr Год назад

Hi, interesting point...Unfortunately I don't know... As far as I understood it is possible to translate json files into .nam...but not vice versa...but I'm not 100% sure. Thank you!

@gameshowfx3617 Год назад

Grande leo , i tuoi video sono molto interessanti soprattutto su queste nuove tecnologie . Ho visto che fai profili di tonex , li farai anche per nam ? Mentre per Aida x cosa ne pensi ?

@LeoGibsonGtr Год назад

Grazie mille! Si, ho un pacchetto di profili per ToneX e quindi mi sembrerebbe poco rispettoso per chi li ha acquistati rendere disponibili i profili NAM...che per la filosofia del prodotto dovrebbero essere free...sto pensando magari ad un pacchetto con alcuni IR...così da offrire un pacchetto completo. Su AIDA ho grandi aspettative, soprattutto perché lo puoi caricare su DWARF, quindi offerendo un pacchetto completo anche per suonare live. Tuttavia fino ad ora non sono riuscito a profilare in locare sul mio pc e quindi sto cercando di risolvere la cosa prima di fare un video dedicato...se non riesco....userò la procura on line e amen....ma mi piacerebbe poter studiare bene il codice, come ho fatto per NAM...

@gameshowfx3617 Год назад

Apprezzo quello che fai , e come lo fai . Complimenti

@LeoGibsonGtr Год назад

Grazie mille!!!!

@flmason 11 месяцев назад

So where is the part where the sound quality of the different parameter groups are compared?

@slowblow Год назад

I am confused with nr of channels mangling. If i changed 16 for 64 there shows an error telling output layer number of channels mismatch. Leo, how to calculate the correct values in each layer? I am using google colab "hard mode" version runtime environment.

@LeoGibsonGtr Год назад

Hi, the "channels" of the first layer has to equal to the "input_size" of the second layer. This is the architecture file: Architecture.STANDARD: { "layers_configs": [ { "input_size": 1, "condition_size": 1, "channels": 32, "head_size": 8, "kernel_size": 3, "dilations": [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512], "activation": "Tanh", "gated": False, "head_bias": False, }, { "condition_size": 1, "input_size": 32, "channels": 8, "head_size": 1, "kernel_size": 3, "dilations": [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512], "activation": "Tanh", "gated": False, "head_bias": True, }, ], "head_scale": 0.02, },

@Geeztown Год назад

Great video! I subscribed. I'm trying to model a fuzz pedal and getting an ESR of 0.25. Any advice on what settings I should change? Is there something special about fuzz pedals that makes them harder to train?

@LeoGibsonGtr Год назад

Hi, thank you so much I'm really happy that you like the video and your sub really mean a lot to me!!! Actually the more distortion you want to profile, the more difficult it is for the profiler and therefore I think that fuzz is pretty difficult to mimic. One thing to verify is that your input file and the target ones are perfectly aligned...this make some difference. Thank you and I hope you will be able to solve...

@Geeztown Год назад

@@LeoGibsonGtr Thanks for the reply. I double checked the alignment, and it was 1 sample off. However, it didn't make much difference, I'm still getting roughly 0.25 ESR. I'm going to try some of the adjustments you talk about in this video and see if I can make it any better.

@Geeztown Год назад

@@LeoGibsonGtr Update: I changed learning rate and decay to 0.001 and dilations to 2048. These changes got me to 0.07 ESR. I tried changing channels to 32, but I got an error. Thanks again for this video, these settings are making a big difference in my case!

@Geeztown Год назад

@@LeoGibsonGtr Update 2: Now I'm down to 0.0076!! And with only 100 epochs. I'm using learning rate and decay of 0.0001. Batch size of 32 although changing this didn't make a huge difference. Dilations of [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192]. I tried your 512 x3, but I got worse results. 8192 seemed to be the sweet spot for me that gave me the best result before getting worse at 16384. And I finally figured out the channels at 32. I had to change channels to 32, but also head size to 16, and the one below it I had to change input size to 32 and channels to 16. Honestly don't know what I was doing there, but that worked! Next i will try creating my own test tone file thanks to some info from Jason Zdora ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-CZL-aY2BEs8.html Thanks again for sharing the info in this video!

@LeoGibsonGtr Год назад

Thank you so much, really interesting info! Please check out also if the final profile will work in the plugin, as the profiles with large neural net could drain a lot of processing power when you load them into the plugin. If you want, please share your findings. Thanks!

@slowblow Год назад

Any chance for reverse engineering of .nam files to have an insight of inner wavenet layer structure?

@LeoGibsonGtr Год назад

Hi, it would be a huge task...and I even don't know if it is actually feasible... Thank you for your suggestion!

@Rhuggins 10 месяцев назад

Would it be theoretically possible to set this up to accurate capture compression behavior? I think it would have something to do with what NAM is told that the relevant time window of sampling would be

@LeoGibsonGtr 10 месяцев назад

Interesting....I think it is possible, but I'm not 100% sure...

@jtn191 7 месяцев назад

People have models of tape machines & tape sims so worth trying!

@firstnamesurname6550 Год назад

Excelent tests for Scott's NN model Just an hypothesis: It seems that 'ny' is related to 'train_val_split' ... then, The train_test_split function of the sklearn. model_selection package in Python splits arrays or matrices into random subsets for train and test data, respectively. but what specifically the acronym 'ny' means seems to be something that Scott would know ... by the code it appears that it refers to the size of the compiled information chunks to be randomized in the hidden layers of the network ... then, Scott set to 9 seconds the time for compiling the training data chunk once 'ny' get into '8192' it stops and set the data matrix subset to become randomized in NN layers ... for the validation of the income it seems that is not required a 'ny' buffer' ,,, then, once the chunk to randomize comes in, the model triggers for the next one ... but it is not - necessarily - something sequential, because, the GPUs are running this Python variable in parallel ... Just guessing ... If the parameter is too low, then, the learning becomes faster or slower according to the hidden layers of net but could be less or more accurate by the settled software architecture ... If the parameter is too high, the learning becomes a bit slow and could be less accurate, because the big chunk doesn't allow the randomized hidden layers to spot the patterns in the data chunks ... Again, just guessing ... Try X/2 or 2x values from 4096 The man to bring an accurate answer should be Scott himself ...

@LeoGibsonGtr Год назад

Hi, thank you so much! May I ask you if you mean Scott or Steven? In my experience NY is related to Epoch, I mean with half NY you get the same performances of 2/5 epochs. If it is lower, each epoch take much longer, but you get the same results in 2/5 of the epochs... Thank you for you interesting insights...

@firstnamesurname6550 Год назад

@@LeoGibsonGtr Ops, yep, Steven Atkinson ... The epoch steps determines the architecture for the hidden layers of nodes to randomize , it is supoused that more epoch steps, more time for the NN to randomize the nodes of each layer ... the size of the nodes ( maybe 'ny' ) determines the 'grain resolution' of the nodes to become randomized ... In theory, The goal is to find the desired optimal outcome with less epoch steps and the optimal size for the nodes 'compiled in less time' ( 'train_val_split' ) ... the thing that 'ny' are multiples of 2 is a sign that it is related to a compiler of bytes ... P.S. : Amazing hardcore test and post, but I missed your amazing and soulful playing in the results 🦆

@LeoGibsonGtr Год назад

Thank you!!! I prefer to stay on the tech side with this video...so no music...of course I'm really happy that you appreciate the music!! Thank you also for your interesting points! Basically if I train a NN with 5000 epochs and an NY = x, I obtain almost the same ESR, in almost the same time, of a NN trained with 2000 epoch and NY = x/2. So at the end, what I derived from that, is that I can avoid to test the NY, as I just need to increase or decrease the epochs....but I will try to do more testes. Thanks!

@JudgeFredd Год назад

Very nice Data Science skills...

@LeoGibsonGtr Год назад

Thank you, really super appreciated!!! Actually, I'm not a data scientist...this is the fruit of my personal studies and lectures in the last month. I love coding since I was a little child, and I have been graduated in computer science...but I was not coding since more than 20 years...so I can say that I have a pretty strong background in computer and coding, but I'm not a data scientist...Please take this into consideration as there could be mistakes in what I have done in this video...and I appreciate any suggestion to do things differently, if I made some mistake. Thank you!

@JudgeFredd Год назад

@@LeoGibsonGtr Music and Maths have common roots...

@LeoGibsonGtr Год назад

Actually....I strongly agree!!!

@digitalchris6681 Год назад

A quick question: for two models produced by just varying the parameters, if one esr is 0.01 and the other is, say half that at 0.005 - does the latter sound twice as good? ie do these esr numbers translate linearly into the perceived quality of the model when heard?

@LeoGibsonGtr Год назад

HI, I don't think so. Actually I have to do some more tests using the profiles I have created, but I feel slight differences in terms of "feel" even between two profiles very similar in terms of ESR...I'm still learning... but generally speaking lower ESR translates in better profiles. Thank you!

@boshi9 Год назад

@@LeoGibsonGtr I'm curious, did you try testing these models blindly? I.e. can you detect the differences in "feel" when you don't know whether you're playing a model with a lower or higher ESR?

@LeoGibsonGtr Год назад

Hi, I'm not 100% sure....as I said...this could be simply an "impression" / suggestion...Nevertheless, with my next video I will share almost 20 profiles of the exact same amps, done with different neural net architectures...so you can check out your own...There is one specific aspect related to ESR that I have noticed for sure...but this is a topic for even another video...😀

@boshi9 Год назад

@@LeoGibsonGtr Thanks, Leo! I'm looking forward to the new videos.

@digitalchris6681 Год назад

@@LeoGibsonGtr Me too !

@realbad3527 Год назад

Hi newbie here. how can i use delay & reverb on this vst? could someone help me 😅

@LeoGibsonGtr Год назад

Typically, in your Digital Audio Workstation, you have several slots where you can load VST plugins....in one of these slots you can load NAM and the other you can load whatever delay or reverb plugin you want.

@boshi9 Год назад

Hi Leo, perhaps I've missed this information in your video, but are you using a different recording for doing the null-test / calculating LUFS, or the same that was used for training? I'm asking because when doing ML there's always a risk of overfitting the model to your training data, and the model that performs best on the training dataset is typically not the model that will generalize well. You really don't want to overtrain for this reason.

@LeoGibsonGtr Год назад

Hi, the null test is done with a different DI track compared to the one used for training NAM. Thank you for your question!

@boshi9 Год назад

@@LeoGibsonGtr Cool, thanks for clarifying!

@LeoGibsonGtr Год назад

You are welcome!

@mandubien Год назад

I did’t get everything cause damn it’s complicated! However, by judging the results, the real question for me is: is it worth it to get 40-60% better ESR for so much more time? Will it sound 40-60% better than regular settings and 100 epochs ? Not sure at all…

@LeoGibsonGtr Год назад

Yes, exactly that's the point. In my opinion any improvement is worth the computational time, as the profile is done only once...and you keep it for many months / years. But it is definitely up to you (in my opinion) to evaluate if it is worth or not. Thank you!

@NickLeonard Год назад

great stuff, how would you implement these? can you share the best modified code?

@LeoGibsonGtr Год назад

Hi, as the software is evolving, I would suggest to make the changes checking out what I did in the video. The most complex one is related to the architecture that you can copy and paste from here: Architecture.STANDARD: { "layers_configs": [ { "input_size": 1, "condition_size": 1, "channels": 32, "head_size": 8, "kernel_size": 3, "dilations": [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512], "activation": "Tanh", "gated": False, "head_bias": False, }, { "condition_size": 1, "input_size": 32, "channels": 8, "head_size": 1, "kernel_size": 3, "dilations": [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512], "activation": "Tanh", "gated": False, "head_bias": True, }, ], "head_scale": 0.02, },

@NickLeonard Год назад

@@LeoGibsonGtr thank you, I'll try that out. Keep up the great work!

@LeoGibsonGtr Год назад

Thank you, really appreciated!!

@NickLeonard Год назад

@@LeoGibsonGtr I got excited to say it did improve the ESR, but actually I remembered wrong and it seems to have the same result, despite your code being different (I was careful to add it in the right place!) It didn't seem to run any slower, which is good, and it was already at .0043 with stock code. I wonder if Steve saw this and made some tweaks to the current version ;)

@GranulatedStuff Год назад

Hi Leo. Do you know how to change the 'num_workers' value ? I've searched through all the scripts and edited the CPU count wherever it crops up but it never works ? cheers

@LeoGibsonGtr Год назад

Hi, I have tried myself, but I ended up not doing it, as it basically provides no improvements. I don't remember where I changed the parameter. Thank you!

@GranulatedStuff Год назад

@@LeoGibsonGtr Thanks for replying. It's so annoying only using about 40% of one core when you've got 20 !

@davidelvis3901 Год назад

I think quad cotex, headrush, or Kemper are missing someone in their space. Hope they find Leo. ❤

@LeoGibsonGtr Год назад

😀😀....I would not be 100% that they appreciate me / my work....😀....

@ZiuaChitarelor Год назад

parafrasando dante..... me cojoni complimenti

@LeoGibsonGtr Год назад

Graze mille!!!!

@mirelchirila Год назад

how is leo’s channel so under viewed, at this point it’s my go to for gear reviews and I know electronics, honestly clearly so does he, his approach is too scientific for him not too.

@LeoGibsonGtr Год назад

Thank you soooo much for these amazing words and I'm really happy that you appreciate the channel!!!