MuseNet • Hacker News Board

Re: MuseNet

#161

Post by namibj » Fri, Apr 26, 2019, 12:06 PM UTC

whatshisface wrote: post #149

lawfulcactus wrote: post #133

Earlier quoted context omitted.

Can you explain? I'm not an expert on ML by any stretch of the imagination, but you'd think with the sort of stringent logical coherence required to construct useful programs, it'd be a pretty subpar use case. Or do you mean smaller-scope tools to aid programming, like linters and autocompleters?

I wonder if you could find a representation for computer programs that eliminated all of the degrees of freedom that were syntax errors, leaving only valid programs. In a sense that's what an AST is but you can still have invalid ASTs. I bet it would be a lot easier to generate interesting programs in a representation like that.

There is cartesian genetic programming and some lisp-like models to encode a program as a tree where all combination are valid. Combined with recent work on convolutional graph DNNs, this might be a good approach.

Re: MuseNet

#162

Post by matz1 » Fri, Apr 26, 2019, 12:32 PM UTC

quixoticaxolotl wrote: post #156

Permit wrote: post #66

Earlier quoted context omitted.

>This isn't chess or go, where you either win or you don't. Chopin composed for a reason, and it wasn't just an excuse to throw a lot of notes at the page. I disagree based on your following quote: > Aside from that, this still sounds like aimless noodling. It's far more polished noodling with some awareness of genre cliches, but it's still essentially aimless - and so meaningless. This sort of problem can easily be…

I think the core difference of opinion here seems to stem from a different interpretation of "meaningless". You're arguing an academic definition of meaningful: can you design an AI that solves a hard problem? gambler (and I think TheOtherHobbes, who you are replying to) are arguing a practical definition of meaningful: does it solve a problem people actually have ? It's neat that you can make artificial music, but a…

One thing I can think of is making it faster and easier, maybe even for people who are less skilled in music to generate music that sound good.

That is valuable to me.

Re: MuseNet

#163

Post by hatsunearu » Fri, Apr 26, 2019, 1:29 PM UTC

jefftk wrote: post #144

throwaway66666 wrote: post #53

Earlier quoted context omitted.

It's midi. It's meant to be notation, not a realistic composition. I suppose the background "machine gun" piano notes would be a bassy synth combined with 90s electronic music drums. So please do not focus that much on the audio or performance part of it, but on the composition part of it. I do agree they could have trained it about the importance of velocity though. (That neural net, and most young music students ou…

>It's midi. It's meant to be notation, not a realistic composition. MIDI certainly can represent realistic compositions, when used with good synths, though I agree it's not what it's known for! > I suppose the background "machine gun" piano notes would be a bassy synth combined with 90s electronic music drums. Uh, what sort of bluegrass music do you listen to and where can I find some of it?

I think they meant the gaga-beatles one, there's a lot of "machine gun" piano sounds.

https://soundcloud.com/openai_audio/gaga-beatles

Re: MuseNet

#164

Post by p1esk » Fri, Apr 26, 2019, 2:18 PM UTC

pdxww wrote: post #142

Impressive. Would this model benefit from something like "dilated attention"? Instead of feeding it raw sound samples, we could split the input into 16 sec, 8 sec, 4 sec and so on slices, assign each slice a "sound vector" serving as a short description of that slice and let the generator take those sound vectors as input. This should supposedly let it gain global consistency in output. Now an unpopular opinion. I'm…

Instead of feeding it raw sound samples, we could split the input into 16 sec, 8 sec, 4 sec and so on slices, assign each slice a "sound vector" serving as a short description of that slice and let the generator take those sound vectors as input.

I didn’t quite get it. How would you feed this variable sized input?

Re: MuseNet

#165

Post by p1esk » Fri, Apr 26, 2019, 2:20 PM UTC

davesque wrote: post #158

ML models that generate music continue to improve. However, to me, the result still just seems like a novelty and nothing more. Perhaps it will take much more iteration and improvement in compute power before I hear a piece of music that I find truly inspiring that was generated by an AI. On the other hand, I've been much more impacted by visual "art" already being generated by AIs. Perhaps the musical medium is iron…

Have you seen a “truly inspiring” piece of visual art generated by AI?

Re: MuseNet

#166

Post by gwern » Fri, Apr 26, 2019, 3:59 PM UTC

dharma1 wrote: post #90

gdb wrote: post #16

Earlier quoted context omitted.

Not yet, but we plan to release the code & pre-trained model!

What kind of hardware (and what size data set) would you need to train with new types of music? Doable with a single 1080ti and a couple of hundred midi files? Also, can you do supervised learning with this - say melody input and chords (with good voice leading) output?

> Doable with a single 1080ti and a couple of hundred midi files?

A 1080ti would probably require something like several days or a week. It depends on how big the model is... Probably not a big deal. However, a few hundred MIDI files would be pushing it in terms of sample size. If you look at experiments like my GPT-2-small finetuning to make it generate poetry instead of random English text ( https://www.gwern.net/GPT-2 ), it really works best if you are into at least the megabyte range of text. Similarly with StyleGAN, if you want to retrain my anime face StyleGAN on a specific character ( https://www.gwern.net/Faces#transfer-learning ), you want at least a few hundred faces. Below that, you're going to need architectures designed specifically for transfer learning/few-shot learning, which are designed to work in the low _n_ regime. (They exist, but StyleGAN and GPT-2 are not them.)

Re: MuseNet

#167

Post by ptrenko » Fri, Apr 26, 2019, 4:37 PM UTC

gdb wrote: post #16

ipsum2 wrote: post #9

Neat output, the project is very similar compared to the Google Doodle a few weeks back, where they generated music based off of user-submitted notes. https://www.google.com/doodles/celebrating-johann-sebastian-... Since this is OpenAI, is MuseNet open source?

Not yet, but we plan to release the code & pre-trained model!

Hi please provide an ETA?

Re: MuseNet

#168

Post by uuwp » Fri, Apr 26, 2019, 7:24 PM UTC

p1esk wrote: post #164

pdxww wrote: post #142

Impressive. Would this model benefit from something like "dilated attention"? Instead of feeding it raw sound samples, we could split the input into 16 sec, 8 sec, 4 sec and so on slices, assign each slice a "sound vector" serving as a short description of that slice and let the generator take those sound vectors as input. This should supposedly let it gain global consistency in output. Now an unpopular opinion. I'm…

Instead of feeding it raw sound samples, we could split the input into 16 sec, 8 sec, 4 sec and so on slices, assign each slice a "sound vector" serving as a short description of that slice and let the generator take those sound vectors as input. I didn’t quite get it. How would you feed this variable sized input?

The same way we feed the variable size sequence of characters or sound samples into this RNN. Instead of raw samples at the 16 kHz rate, we'll have one sequence of 1 sample per second, another sequence of 1 sample per 0.5 seconds and so on. We can go as far as 1 sample per 1/48000 sec, but I don't think it's practical (but this is what these music generators do).

Re: MuseNet

#169

Post by raindropm » Sat, Apr 27, 2019, 4:10 AM UTC

BowBun wrote: post #71

codelitt wrote: post #64

While this may not be perfect composition, this is a surreal (and almost sad) moment in my life to hear music that is passable created by a computer under it's own volition. I work at a company that works with a lot of machine learning so I generally understand its limitations and haven't ever been an alarmist. That being said, I generally always thought of it being applied to automate work. For some reason I had alw…

Fear not, my friend, because the execution is only a part of why music and art as a whole carries meaning for us. Consider two tracks that are identical (forget copywrite for a minute). Between one that an AI generated and a human composed, I would personally grant the human-generated version more credit and enjoy it more. The story of how art is created and the stories of the artist are as substantial to appreciatin…

I kinda thank you for your comment, it gives me more...hope? :) (I'm an artist) The same way when we see other human did something extraordinary that we think human can't do(based mostly on ourselves) For example, an artist who can draw something so life-like or sculptor who shape hard marble into soft, flowing clothes — with just hammer and chisel.

Human potential always intrigues us "What? Human can do THAT?" kind of way.

Yes, the machine and AI can do the same thing at the fraction of time, from the practicality standpoint, but it's not and never be the same — it's empty. It's just lifeless product and we never feel related to it.

Re: MuseNet

#170

Post by pdxww » Sat, Apr 27, 2019, 4:12 AM UTC

p1esk wrote: post #164

pdxww wrote: post #142

Impressive. Would this model benefit from something like "dilated attention"? Instead of feeding it raw sound samples, we could split the input into 16 sec, 8 sec, 4 sec and so on slices, assign each slice a "sound vector" serving as a short description of that slice and let the generator take those sound vectors as input. This should supposedly let it gain global consistency in output. Now an unpopular opinion. I'm…

Instead of feeding it raw sound samples, we could split the input into 16 sec, 8 sec, 4 sec and so on slices, assign each slice a "sound vector" serving as a short description of that slice and let the generator take those sound vectors as input. I didn’t quite get it. How would you feed this variable sized input?

To illustrate more this idea, let's use soundtrack v=negh-3hi1vE on youtube. Such soundtracks consist of multiple more or less repeating patterns. The period of each pattern is different: some background pattern that sets the mood of the music may have a long period of tens of seconds. The primary pattern that's playing right now has a short period of 0.25 seconds, plays for a few seconds and then fades off. The idea is to split the soundtrack into 10 second chunks and map each chunk to a vector of a fixed size, say 128. The same thing we do with words. Now we have a sequence of shape (?, 128) that can be theoretically fed into a music generator and as long as we can map such vectors back to 10 second sound chunks, we can generate music. Then we introduce a similar sequence that splits the soundtrack into 5 second chunks. Then another sequence for 2.5 seconds chunks and so on. Now we have multiple sequences that we can feed to the generator. Currently we take 1/48000th second slices and map them to vectors, but that's about as good as trying to generate meaningful text by drawing it pixel by pixel (which we can surely do and the model will have 250 billion weights and take 2 million years to train on commodity hardware).