Update README.md
0d9c7cf
verified
metadata
license: apache-2.0
Quad Music Transformer
SOTA quality fast music transformer with symmetrical quad MIDI notes encoding
Original Version
Features demonstration
Composer Version
MuseNet-style workflow for endless supervised continuation generation
Bulk Generator Version
Bulk improvs and continuations generation
Technical notes
SOTA quality was achieved by using the following specific techniques:
1) Quality source MIDI dataset (quality over quantity)
2) MIDI dataset augmentation by time (x2) and pitches (x3)
3) Timings normalization, quantization and compression (128)
4) Larger model embed size (2048) with less layers (16) and heads (16)
5) Training longer since the MIDI dataset is small (2 full epochs)
6) Using MIDI instruments families (16) instead of full MIDI instruments range (128)
7) Using symmetrical quad MIDI notes encoding
8) 8k sequence length so that the model can learn long-term music scructure
9) Using fp16 precision so that the model is sufficiently fast with low memory footprint
10) Hex (16) MIDI velocity range to avoid velocity overfitting while preserving velocity details
11) Chords sorting by instruments families (L-H) and by pitch (H-L)
Project Los Angeles
Tegridy Code 2024