Magenta Performance-RNN Etudes

[Circu Virtu]

Hi,

I did some modifications to the performance-rnn generator and got some interesting
music I wanna share with you.
Don't expect too much. This are the first runs and more ideas an work is ahead to hopefully
improve the results.

https://soundcloud.com/circu-virtu/sets/magenta-performance-rnn-etudes

Comments, critics welcome.

Cheers,
cv

[Adam Roberts]

Very nice! What changes did you make?

--
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-discuss+unsubscribe@tensorflow.org
---
You received this message because you are subscribed to the Google Groups "Magenta Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discuss+unsubscribe@tensorflow.org.

[Circu Virtu]

tnx, I'm using 'parameter modulation', changing generation parameters over time.
In this examples it was mainly modulatiing temperature over time.
Just the first tests which could be improved a lot to change other parameters as well and adapt
the parameter changes to note statistics etc.
Btw. is there any detailed description of how the parameters like beam-size, branch-factor in performance-rnn work.

--cv

Am Freitag, 4. August 2017 18:27:52 UTC+2 schrieb Adam Roberts:

Very nice! What changes did you make?

On Fri, Aug 4, 2017 at 8:38 AM, Circu Virtu <circu...@gmail.com> wrote:

Hi,

I did some modifications to the performance-rnn generator and got some interesting
music I wanna share with you.
Don't expect too much. This are the first runs and more ideas an work is ahead to hopefully
improve the results.

https://soundcloud.com/circu-virtu/sets/magenta-performance-rnn-etudes

Comments, critics welcome.

Cheers,
cv

--
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

---
You received this message because you are subscribed to the Google Groups "Magenta Discuss" group.

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

[Ian Simon]

We've been working on similar ideas.  #810 adds the ability to provide Performance RNN with a desired note density and pitch class histogram, e.g. you can ask the model to generate a performance with ~8 notes per second using a C major scale.  #819 lets you change these parameters over the course of a performance, e.g. you can start in C major and then switch to G major.

The beam search parameters work the same in all of our RNN models; probably the best description we have is here: https://github.com/tensorflow/magenta/blob/master/magenta/common/beam_search.py#L77

-Ian

[Circu Virtu]

Thank you for the infos. Great to have some additional parameters to play with.
Is there anything special  to consider when training those models?
I see that there are several pretrained .mag files (although most if them just link to performance_with_dynamics.mag)

Is there any chance to get those parameters controlled in real-time my the MIDI interface as well?

--cv

[Ian Simon]

The Performance RNN .mag files should be fixed now with #822.  Yes, real-time control would be pretty cool and it's something we're thinking about.

-Ian

[Circu Virtu]

I have not used the new parameters yet. Still struggling to get a magenta development environment up and running.
But at least made some improvements to my code and managed to create some music again.
Enjoy (hopefully)

https://soundcloud.com/circu-virtu/performance-rnn-etude-5

--cv

[Circu Virtu]

Finally  got the develoment environment with the new parameters installed an ready to produce some music
This time I tried some different sounds other than piano as with my other etudes.
Enjoy.

https://soundcloud.com/circu-virtu/performance-rnn-pad-1

..cv

Am Montag, 14. August 2017 18:42:20 UTC+2 schrieb Ian Simon:

[Jesse Engel]

Very cool. Did you slow down the output at all or is this the straight output from the model? The model was trained on a lot of classical music performances, so when it shifts keys it makes a very interesting dissonant texture with all the reverb.

To unsubscribe from this group, send email to magenta-discuss+unsubscribe@tensorflow.org

---
You received this message because you are subscribed to the Google Groups "Magenta Discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discuss+unsubscribe@tensorflow.org.

[Circu Virtu]

Tnx, yes you got it right. The tempo is slowed down (80 instead of default 120).
I usually change the tempo when rendering MIDI from the model to music to fit the 'sound'.
And I used the additional parameters:
--notes_per_second="[3,5,10,5,3]" \
--pitch_class_histogram="[[2, 0, 1, 0, 1, 1, 0, 1, 0, 1, 0, 1],[0, 0, 1, 2, 0, 1, 0, 0, 0, 0, 0, 0],[2, 0, 1, 0, 1, 1, 0, 1, 0, 1, 0, 1]] " \


--cv