[BUG] Machine learning crash entire DAW

[vanerio]

Hi!
I was trying to test the new machine learning function in MWaveShaperMB, with just a kick sample as dry and the same kick distorted as ideal. After all iterations, when the process is about to end, the entire DAW crashes, but not even freezes, it just close everything inmediately.
Used settings: all default, except sample paths

EDIT: Ok, the crash happens only when there are no multiparameters selected!

New thing: I noticed that Attempts number is not considered, as the process will just continue over it!

[Frankie.T]

Happens also with MP.
Not always though, i used it 5 times and crashed once

I used Reaper 7 if that matter

[vectorwarrior]

Well, it is highly experimental and in beta, definitely a work in progress.

I was getting consistent crashes on completion until I changed my pc from 'quiet' thermal mode to 'performance', so consider tweaking any power or performance modes to more high performance ones.

Also I was told to try to keep the number of MPs to as low as possible as the potential permutations increase exponentially with each MP you add.

Really excited to see this improve but it's obviously a long way from ready, so keep that in mind when testing it.

[SeeingInMidi]

How do you add parameters to learn?

Edit: Nevermind, figured it out. For anyone wondering, you have to assign the plugin parameters to an MP, and then refresh the machine learning window.

[Frankie.T]

Anyway i don't feel the way this tool works is the right approach for learning compression.

Ofc i have no clue on how it's written and i may be very wrong, but looking on the process, i see it just try to guess values randomly changing them, using (i don't know how) the dry and wet file of a processed reference.

I think a better approach should be to have well known input files that are made in a way the tool understand what they are and how the process was made.

I mean multiple audio input, where the first file should be something like a transient driven file (like a clap or a ride) then someting else with a first peak followed by a second one but lower in amplitude (than another one with an opposite amplitude) and so on.

Cause when i tried and analyzed some compressors i saw different behaviour made by the internal processing, and this data is what you need to understand how the compressor treat transients, not just about attack, release and ratio, but also how it react with different peaks (some have a very particular shape about attack and release that looks detached by what you select) and also the different ratio along the range.

I don't see how this informations can be guessed buy the tools without the needed data. It also looks weird the way he randomly try stuff, it looks like an algorithm that doesn't really try to learn from data (that doesn't have).


This is not meant to be a criticism, but just a feedback! We talked in another post about that feature with Vojtech and he released it! And that's insane! I love you man, just don't hate me

[vectorwarrior]

This random iterative approach is kind of the way it HAS to be done though. How do you propose it knows what values to try unless it tests them?

For what it's worth, I heard that the 'attempts' control DOES work. It does this many iterations, then uses the best from that batch and iterates on variations around the best of those results. So I guess it isn't total random, it's more like a two stage process.

The real challenge is how it analyses 'success'. That is what the controls mostly determine and where you will get good or useless results.

I personally like the user input test file. Depending on the type of effect you are trying to replicate you will be better off with a different input output file. You can also tailor the length to suit your needs, which massively effects processing time.

[MeldaProduction]

Yeah, it needs to "medical attention"

[jan-sandahl]

Looking forward to hear more on this feature. I don't know what to expect from a "Machine Learning" button among MSF or MPS utilities. Sounds exciting though.