COMI

Codes are tested on:

• python 3.7
• pytorch 1.4.0
• pretty_midi 0.2.9
• mido 1.2.9

Data Preparation

data processing-related files:

• data_config.py
• data_cleaning.py
• tokenization_by_time.py
• preprocess.py

First, the midi data is put under ./data/midicn/midi/

Then, do some data filtering and cleaning. The cleaned midi files are put under ./data/midicn/cleaned_midi/

python data_cleaning.py midicn

Then, pre-process each midi file and save as a txt file; at the same time, the ./data/vocab.txt is created. Saved txt files are separated as three sub-sets (train, val, test) and are put under ./data/midicn/txt/. This step is to represent a midi as a token sequence using the vocabulary that we've defined.

python preprocess.py midicn