Proof of concept TensorFlow Neural Net to classify midi files. Based on the fully connected MNIST example from TensorFlow.
Reads MIDI files in midicsv format.
The conversion process from MIDI to an array suitable for machine learning is:
- Split the MIDI event sequence into "windows" of time, 100ms by default.
- Create a 100-element array for every window (should be enough note values to cover most keyboard music).
- For every "note on" event that occurs within a window, increment the position in its array that corresponds to the note value.
- Group the windows into samples, 10 seconds by default.
The system also artifically increases the amount of data available by repeatedly shifting the data by 1ms.
Here's a test run that classifies some Bach preludes and fugues from the The Well-Tempered Clavier against some childrens' songs. I got the midi files from http://www.mfiles.co.uk/ and then ran them through midicsv.
After ~1300 steps, it achieves 100% accuracy with the training, validation, and test sets!
Extracting ../musicid/midis/train
Read 31400 windows from ../musicid/midis/train/bach/book1-fugue02.mid.csv
Read 76900 windows from ../musicid/midis/train/bach/book1-fugue24.mid.csv
Read 35000 windows from ../musicid/midis/train/bach/book1-prelude01.mid.csv
Read 38800 windows from ../musicid/midis/train/bach/book1-prelude02.mid.csv
Read 39700 windows from ../musicid/midis/train/bach/book1-prelude03.mid.csv
Read 26300 windows from ../musicid/midis/train/bach/book1-prelude06.mid.csv
Read 96100 windows from ../musicid/midis/train/bach/book1-prelude24.mid.csv
Read 27800 windows from ../musicid/midis/train/bach/book2-fugue02.mid.csv
Read 57200 windows from ../musicid/midis/train/bach/book2-prelude02.mid.csv
Read 72000 windows from ../musicid/midis/train/bach/book2-prelude12.mid.csv
Read 5900 windows from ../musicid/midis/train/children/bye-baby-bunting.mid.csv
Read 12100 windows from ../musicid/midis/train/children/here-we-go-round-the-mulberry-bush.mid.csv
Read 5500 windows from ../musicid/midis/train/children/hey-diddle-diddle.mid.csv
Read 5500 windows from ../musicid/midis/train/children/hickory-dickory-dock.mid.csv
Read 7800 windows from ../musicid/midis/train/children/hot-cross-buns.mid.csv
Read 5900 windows from ../musicid/midis/train/children/lavenders-blue.mid.csv
Read 3700 windows from ../musicid/midis/train/children/mary-mary-quite-contrary.mid.csv
Read 11700 windows from ../musicid/midis/train/children/ring-a-ring-o-roses.mid.csv
Read 12100 windows from ../musicid/midis/train/children/see-saw-margery-daw.mid.csv
Read 24000 windows from ../musicid/midis/train/children/sing-a-song-of-sixpence.mid.csv
Read 11900 windows from ../musicid/midis/train/children/twinkle-twinkle-little-star.mid.csv
Extracting ../musicid/midis/validation
Read 60100 windows from ../musicid/midis/validation/bach/book1-fugue14.mid.csv
Read 24200 windows from ../musicid/midis/validation/bach/book1-prelude14.mid.csv
Read 7800 windows from ../musicid/midis/validation/children/goosey-goosey-gander.mid.csv
Read 12100 windows from ../musicid/midis/validation/children/rock-a-bye-baby-tune-a.mid.csv
Read 12100 windows from ../musicid/midis/validation/children/rock-a-bye-baby-tune-b.mid.csv
Read 11700 windows from ../musicid/midis/validation/children/rub-a-dub-dub.mid.csv
Extracting ../musicid/midis/test
Read 70800 windows from ../musicid/midis/test/bach/book2-fugue07.mid.csv
Read 81200 windows from ../musicid/midis/test/bach/book2-prelude07.mid.csv
Read 5900 windows from ../musicid/midis/test/children/humpty-dumpty.mid.csv
Read 5500 windows from ../musicid/midis/test/children/little-jack-horner.mid.csv
Read 5500 windows from ../musicid/midis/test/children/little-miss-muffet.mid.csv
Read 7800 windows from ../musicid/midis/test/children/the-muffin-man.mid.csv
Step 0: loss = 0.70 (0.228 sec)
Step 100: loss = 0.30 (0.214 sec)
Training Data Eval:
Num examples: 6000 Num correct: 4958 Precision @ 1: 0.8263
Validation Data Eval:
Num examples: 1200 Num correct: 843 Precision @ 1: 0.7025
Test Data Eval:
Num examples: 1700 Num correct: 1520 Precision @ 1: 0.8941
Step 200: loss = 0.23 (0.256 sec)
Step 300: loss = 0.19 (0.527 sec)
Training Data Eval:
Num examples: 6000 Num correct: 5523 Precision @ 1: 0.9205
Validation Data Eval:
Num examples: 1200 Num correct: 804 Precision @ 1: 0.6700
Test Data Eval:
Num examples: 1700 Num correct: 1483 Precision @ 1: 0.8724
Step 400: loss = 0.14 (0.262 sec)
Step 500: loss = 0.16 (0.217 sec)
Training Data Eval:
Num examples: 6000 Num correct: 5917 Precision @ 1: 0.9862
Validation Data Eval:
Num examples: 1200 Num correct: 1176 Precision @ 1: 0.9800
Test Data Eval:
Num examples: 1700 Num correct: 1641 Precision @ 1: 0.9653
Step 600: loss = 0.10 (0.615 sec)
Step 700: loss = 0.10 (0.216 sec)
Training Data Eval:
Num examples: 6000 Num correct: 5935 Precision @ 1: 0.9892
Validation Data Eval:
Num examples: 1200 Num correct: 1200 Precision @ 1: 1.0000
Test Data Eval:
Num examples: 1700 Num correct: 1691 Precision @ 1: 0.9947
Step 800: loss = 0.05 (0.277 sec)
Step 900: loss = 0.06 (0.664 sec)
Training Data Eval:
Num examples: 6000 Num correct: 5980 Precision @ 1: 0.9967
Validation Data Eval:
Num examples: 1200 Num correct: 1200 Precision @ 1: 1.0000
Test Data Eval:
Num examples: 1700 Num correct: 1699 Precision @ 1: 0.9994
Step 1000: loss = 0.05 (0.334 sec)
Step 1100: loss = 0.03 (0.232 sec)
Training Data Eval:
Num examples: 6000 Num correct: 5998 Precision @ 1: 0.9997
Validation Data Eval:
Num examples: 1200 Num correct: 1200 Precision @ 1: 1.0000
Test Data Eval:
Num examples: 1700 Num correct: 1700 Precision @ 1: 1.0000
Step 1200: loss = 0.04 (0.620 sec)
Step 1300: loss = 0.04 (0.234 sec)
Training Data Eval:
Num examples: 6000 Num correct: 6000 Precision @ 1: 1.0000
Validation Data Eval:
Num examples: 1200 Num correct: 1200 Precision @ 1: 1.0000
Test Data Eval:
Num examples: 1700 Num correct: 1700 Precision @ 1: 1.0000