how do I train a model from the beginning?a music model (actually more like a text model because it is notes only, like C#4, A-3, etc you get the ideaI need something that understands that A is an example of music and B is not-musicI would feed it with note data that is certainly not music as human would understand it. And also with note data that is considered music (some works of a well known artist, only for a test run, once everything works I will begin from zero again using only my own melodies)The non-music would be literally two hands smashing at piano in a incoherent cacophony so that the AI would understand what is certainly should not create and mix.
Where do you draw the line, do you consider rap to be music?
>>108191320So you want to make a music classifier? Or do you want to make a music generator?Also, to check, you plan to feed in something like a sibelius file (tempo+score(string of chords or rests)?
there are a few different things you can do. google has a rnn suite under the magenta project that is somewhat similar to what youre looking for; it gets trained on midi data. i have trained two models, one to generate drum loops i can practice guitar over and one that generates midi in the style of old nes games. i did exactly what you would expect for these, i fed in a shitload of drum loops for the first one and a shitload of nes midi loops for the second and ran them for a week on my cpu. the results are good.i recommend poking around the magenta framework. its pretty easy to get your head around the concepts since its an rnn and i think they are one of the more simple types of ml.things like acestep the local suno clone are diffusion based-ish. these are quite a bit more difficult to understand than the fairly simple maths behind how magentas rnns work and unless you want to really move into research for researches sake, i would recommend sticking with 'older' ml concepts like rnns and maybe even markov chains etc for your described purpose. you can most def acheive what youre after with a lot of string and glue. start with the google magenta project.
step 0: gather datastep 1: gather more datastep 2: gather more datastep 3: gather more datastep 4: gather more datastep 5: gather more datastep 6: gather more datastep 7: gather more datastep 8: gather more datastep 9: gather more datastep 10: gather more datastep 11: gather more datastep 12: gather more datastep 13: gather more datastep 14: gather more datastep 15: gather more datastep 16: gather more datastep 17: gather more datastep 18: gather more datastep 19: gather more datastep 20: gather more datastep 21: gather more datastep 22: gather more datastep 23: gather more datastep 24: gather more datastep 25: gather more datastep 26: gather more datastep 27: gather more datastep 28: gather more datastep 29: gather more datastep 30: gather more datastep 31: gather more data
>>108191320maybe start with something simpler, like tic tac toegrasp the basics of tensors and whatnot, then the answer will come to you when you have a better understanding of the tools
>>108191320you'll probably end up with a noise detector if you try fitting a classifier this way
>>108192103I'm a musician lol, not going to learn computing stuff too deeply
>>108192152Nta, but unless you're loaded you need to learn some programming to create an LLM.
>>108192152>not going to learn computing stuff too deeplythen youre not going to create any models from scratch. it is really that simple
