Advice on training a model

[Ryan Helsing]

I have large a set of 1 to N harmony to melody midi files that I've recorded myself playing. The melody midi files have pitch bend information and the harmony files are dont have much dynamics. How would you suggest I go about working on a model that when given a new harmony file, it generates several potential melody files as output? I have looked a little bit into magenta but I feel like I'm in over my head. Thanks!

[Ian Simon]

Hi Ryan, the closest model in the Magenta repo we have to what you're describing is probably Improv RNN, which is now several years old: https://github.com/magenta/magenta/tree/main/magenta/models/improv_rnn 

There are a number of caveats.  The model only accepts chord symbols as harmony, so if your harmony is arranged you'd have to somehow infer chord labels.  It also doesn't handle pitch bends; currently none of our symbolic models do.  And it "improvises"; it doesn't look at future chords when deciding what note to play.

But it might be a reasonable starting point.

-Ian

On Fri, Jul 23, 2021 at 12:34 PM Ryan Helsing <ry...@zevobenefits.com> wrote:

I have large a set of 1 to N harmony to melody midi files that I've recorded myself playing. The melody midi files have pitch bend information and the harmony files are dont have much dynamics. How would you suggest I go about working on a model that when given a new harmony file, it generates several potential melody files as output? I have looked a little bit into magenta but I feel like I'm in over my head. Thanks!

--
Magenta project: magenta.tensorflow.org
To post to this group, send email to magenta...@tensorflow.org
To unsubscribe from this group, send email to magenta-discu...@tensorflow.org
---
To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discu...@tensorflow.org.