Quick study of the effects different layers and training runs have on magenta Google's Tensorflow music generation Neural Network Program.
july 28, 2016
temperature change the randomness of the magenta output
temperature 1 https://clyp.it/fs40wzuf
temperature 3 https://clyp.it/0jwhnnhx
temperature 5 https://clyp.it/u5mabdyy
temperature 7 https://clyp.it/pf0mhrnx
Jul22, 2016
Sorry, have not been keeping this readme.md file up to date.
I just re-organized the folder structure to include other builds with different number of midi files. I got magenta working on making it possible to re-train a run with a diffrerent set of midi files and have tested it.
https://groups.google.com/a/tensorflow.org/forum/#!topic/magenta-discuss/riVSW-dKJ3k
I also got magenta to set up a variable called temperature which allows some adjustment to the ability for a song to remember an exact song or something slightly different.
https://groups.google.com/a/tensorflow.org/forum/#!topic/magenta-discuss/SCZR9qrQUGM
I am presently testing this "temperature" variable.
Clyp.it links for anyone who can't play online the github midi or mp3 files
Rocksetta-Magenta-Research-input-1-a-few-of-my-favorite-things
Rocksetta-Magenta-Research-input-2-god-rest-ye-merry-gentlemen
Rocksetta-Magenta-Research-primer-fur-elise
midi = number of inputs = 2 for this entire github repository
Layer = Tensorflow layers in the statement
--hparams='{"rnn_layer_sizes":[50]}'
Could also be a 2 dimensional training neural network
--hparams='{"rnn_layer_sizes":[30, 20]}'
Train = Number of training loops, note the defualt is to make a checkpoint and print data every 10 loops.
Rocksetta-Magenta-Research-midi2-layer-2d-25-25-train300
Rocksetta-Magenta-Research-midi2-layer-2d-5-5-train300
Rocksetta-Magenta-Research-midi2-layer-2d-5-5-train1000
Rocksetta-Magenta-Research-midi2-layer-2d-5-50-train300
Rocksetta-Magenta-Research-midi2-layer-2d-50-5-train1000
Rocksetta-Magenta-Research-midi2-layer3-train20000
Rocksetta-Magenta-Research-midi2-layer5-train300
Rocksetta-Magenta-Research-midi2-layer5-train1000
Rocksetta-Magenta-Research-midi2-layer-200-train1000
Rocksetta-Magenta-Research-midi2-layer5-train1000
Rocksetta-Magenta-Research-midi2-layer50-train300
Rocksetta-Magenta-Research-midi2-layer50-train1000
Rocksetta-Magenta-Research-midi2-layer-50-train3000
Rocksetta-Magenta-Research-midi2-layer500-train300
.
.
.
#June 27, 2016
When everything is done every training run should end up zipped in the named folders so that you can setup my training run with your own primer.mid file. Those folders should also include the .mid file output as well as a converted output to .mp3
The All-inputs folder has the 2 midi file inputs (a-few-of-my-favorite-things.mid and god-rest-ye-merry-gentlemen.mid) and fur-elise.mid as the primer.mid file
Please make Pull Requests with the same folder labelling scheme
midi2.LAYERS.TRAINING
Obviously we are going to do runs with more than two midi files but I needed to make things simple here to look for patterns.
By Jeremy Ellis
Maker of http://www.rocksetta.com
Twitter @rocksetta