midi-track-identification's Introduction

Midi-Track-Identification

Categorize tracks of midi files into four classes:

0: Melody
1: Drum
2: Bass
3: Other

There are two main goals of this task:

Data Cleansing:
- The track names in midi files are usually less meaningful.
- The program number and channel number are not in GM Midi format.
Separate perceptually melodic tracks based on hand-labeling data

Note that it's possible that there are more than one melodic tracks in a song.

Result

Accuracy

Train: 1.0
Test: 0.98

Dataset

Jazz Realbook
Total: 1094 songs (7:3)
- Train: 765
- Test: 329

Conception

Feature Extraction
Classification (Random Forest, n_estimation=100)

Figures

Testing Stage
Feature Distribution
For further understanding, please refer to here

Usage

python compile_dataset.py
python compile_feature.py
python analysis.py

To run testing on self-prepared data, please check out the test.py

midi-track-identification's People

Contributors

Stargazers

Watchers

midi-track-identification's Issues

“Midi-Track-Idenfication” -> "Midi-Track-Identification"

As title.
Please revise it.

Can any body fork and update this repo?

I ran these codes in python 3.8 and got a number of errors. I tried to fix some of them due to deprecated modules (like change "np.int" to "int") but still could not get script.py to work (it threw errors such as "midi_file object has no attribute of track", "track object has no attribute of pianoroll", etc.)

Maybe this repo is based on some particular old-time versions of pypianoroll and miditoolkit. Can any body fork and update it? Better to include the package requirements into the requirements.txt.

Thanks in advance, you gurus

wayne391 / midi-track-identification Goto Github PK