6 ms·

Ah!, I think I understand what you are saying. I'm not sure if you used a NN before but you will notice that there is an exponential relationship with the numbe

Ah!, I think I understand what you are saying. I'm not sure if you used a NN before but you will notice that there is an exponential relationship with the number of inputs and the size of the training set required. Because I don't have millions of perfect training data I need to break down the problem and make ML models to deal with specific problems. For example predicting if a drum beat is compatible with a bassline etc.

p1esk 9y ago
What exponential relationship? Can you point me to any papers about it? It's actually quite easy to generate polyphonic (classical style) music with an RNN: just predict the entire note frequency range of one time step as a binary encoded vector. But yes, I can see how for multiple instruments, the sparsity of inputs might require more training samples. I'm wondering if the "layered" approach has some advantages, perhaps more control over some aspects of the music?
- stevehiehn 9y ago
  Ya, there is no official relationship. From my experiments the more inputs you have the larger the dataset required. I do not know what is best way to do this but the layered approach is how humans compose music so it made sense to try that. I'm not sure I agree that its easy to generate polyphonic music. I would like to be wrong but when you say "easy" I don't think you're referring to music that follows a harmonic structure or factors in timbre. Most music I hear follows structure with repeating themes etc. Not just a stream of conscious series of notes.
  - p1esk 9y ago
    Well, I just tried googling "rnn generated music": https://youtu.be/A2gyidoFsoI?t=1m44s https://youtu.be/A2gyidoFsoI?t=1m44s Sounds like a pretty good polyphony, but of course, it could be just overfitting on Bach training samples. Here are a couple of more sophisticated (and original) pieces: https://soundcloud.com/user-95265362/op-1-for-piano-solo-in-d-major https://soundcloud.com/user-95265362/op-1-for-piano-solo-in-... https://soundcloud.com/user-95265362/op-21-for-piano-solo-genesis https://soundcloud.com/user-95265362/op-21-for-piano-solo-ge...
    - stevehiehn 9y ago
      Those do sound amazing! My goal is to create interactive music. If you noticed in my UI it can change at any point. In order to do that I think I need to premeditate the harmonic structure. i.e chord progression. I suspect that those RNN are not conforming to a predetermined harmonic structure. I doubt if you ask it say generate a piece following Dmaj9/Amaj/Bmin/Bmin it will understand that but I dont know.
      - p1esk 9y ago
        if you ask it say generate a piece following Dmaj9/Amaj/Bmin/Bmin it will understand that Dmaj9/Amaj/Bmin/Bmin will be the seed sequence to the RNN. Of course, the network might ignore the seed, or if it's overfitting it will produce a chunk of the training sample which has this sequence in it, but you can definitely specify the starting point (and you usually do with RNNs).
    - stevehiehn 9y ago
      In order to use RNN in the way you suggest I think I would have to create a very large data set with annotated chord progressons. Which is something I would like to do but it would take a very long time to create the training sets.
      - p1esk 9y ago
        No need to annotate the chords. If you start from a midi file, you just encode the piece as a binary matrix (time range x frequency range). It gets trickier if different instruments are not separated because you would want to specify different instruments as separate channels (so your matrix becomes time range x frequency range x number of channels), but even this could be automated with ML (e.g. using image segmentation methods). The most time-consuming part is finding enough training compositions with labels (such as composer, genre, mood, style, etc), and data cleanup.

p1esk 9y ago

What exponential relationship? Can you point me to any papers about it? It's actually quite easy to generate polyphonic (classical style) music with an RNN: just predict the entire note frequency range of one time step as a binary encoded vector. But yes, I can see how for multiple instruments, the sparsity of inputs might require more training samples. I'm wondering if the "layered" approach has some advantages, perhaps more control over some aspects of the music?

stevehiehn 9y ago
Ya, there is no official relationship. From my experiments the more inputs you have the larger the dataset required. I do not know what is best way to do this but the layered approach is how humans compose music so it made sense to try that. I'm not sure I agree that its easy to generate polyphonic music. I would like to be wrong but when you say "easy" I don't think you're referring to music that follows a harmonic structure or factors in timbre. Most music I hear follows structure with repeating themes etc. Not just a stream of conscious series of notes.
- p1esk 9y ago
  Well, I just tried googling "rnn generated music": https://youtu.be/A2gyidoFsoI?t=1m44s https://youtu.be/A2gyidoFsoI?t=1m44s Sounds like a pretty good polyphony, but of course, it could be just overfitting on Bach training samples. Here are a couple of more sophisticated (and original) pieces: https://soundcloud.com/user-95265362/op-1-for-piano-solo-in-d-major https://soundcloud.com/user-95265362/op-1-for-piano-solo-in-... https://soundcloud.com/user-95265362/op-21-for-piano-solo-genesis https://soundcloud.com/user-95265362/op-21-for-piano-solo-ge...
  - stevehiehn 9y ago
    Those do sound amazing! My goal is to create interactive music. If you noticed in my UI it can change at any point. In order to do that I think I need to premeditate the harmonic structure. i.e chord progression. I suspect that those RNN are not conforming to a predetermined harmonic structure. I doubt if you ask it say generate a piece following Dmaj9/Amaj/Bmin/Bmin it will understand that but I dont know.
    - p1esk 9y ago
      if you ask it say generate a piece following Dmaj9/Amaj/Bmin/Bmin it will understand that Dmaj9/Amaj/Bmin/Bmin will be the seed sequence to the RNN. Of course, the network might ignore the seed, or if it's overfitting it will produce a chunk of the training sample which has this sequence in it, but you can definitely specify the starting point (and you usually do with RNNs).
  - stevehiehn 9y ago
    In order to use RNN in the way you suggest I think I would have to create a very large data set with annotated chord progressons. Which is something I would like to do but it would take a very long time to create the training sets.
    - p1esk 9y ago
      No need to annotate the chords. If you start from a midi file, you just encode the piece as a binary matrix (time range x frequency range). It gets trickier if different instruments are not separated because you would want to specify different instruments as separate channels (so your matrix becomes time range x frequency range x number of channels), but even this could be automated with ML (e.g. using image segmentation methods). The most time-consuming part is finding enough training compositions with labels (such as composer, genre, mood, style, etc), and data cleanup.