MusicGeneration

Sequence Extraction

this porject provide following representation for midi format files:

• MIDI-like
• REMI
• MuMIDI
• CP(to do)

you can transform midi into sequence data as following steps

cd mg/model/utils
python preprocess_{MIDI_like,REMI,MuMIDI,CP}.py {midi_folder} {result_output_foler} {int:num_workers, 10 is recommand}

the structure of representation

the API of representation(like MIDI-like, REMI, MuMIDI)

it should be encapsulate as a EventSeq(or REMI_EventSeq or MuMIDI_EventSeq in sequence.py/REMI.py/MuMIDI.py)

• extract_events(midi_path)->event_seq
• to_array(event_seq)->np.array()
• from_array(np.array)->event_seq() for decoding
• write_midi(event_seq,output_path)
• feats_range()->dict=<feature_name, index>
• dims_feats()->dict=<index , (feat_name, feat_val)>

the structure of model

the hyper parameter should be load in cmd rather than only yaml file.

you can consider mg/model/MuiscTransformer/train.py or mg/model/Event_Melody/train.py at the begining

---

Please see Installation Instructions