midihum - GitLite

KO | EN

gitlite — search

#javascript #hacktoberfest #react #go #android #docker #nodejs #macos #windows #awesome #awesome-list #pytorch #ios #openai

midihum

(★ 139)

MIDI humanisation with machine learning

파일 탐색기

최종 버전 다운로드 (.zip)

- midihum.json
- midihum_scaler.json
.gitignore
.pylintrc
chord_identifier.py
find_duplicate_midi_files.py
LICENSE
main.py
midi_scraper.py
midi_to_df_conversion.py
midi_utility.py
midihum_model.py
midihum_velocities_over_time.png
midihum_velocity_scatter.png
plotter.py
prepare_midi.py
README.md
requirements.txt
start_jupyter.sh
test_prepare_midi.py
time_displacer.py
xgboost_train_model.ipynb
xgboost_tune_hyperparams.ipynb

# 설치 가이드

1. 코드 내려받기

git clone https://github.com/erwald/midihum

깃허브에서 프로젝트 코드 전체를 내 컴퓨터로 내려받습니다.

cd midihum

방금 내려받은 프로젝트 폴더 안으로 이동합니다.

2. Python

쉬움 추천

사전 준비물

Git GitHub에서 프로젝트 코드를 내려받으려면 필요합니다.
Python 3 설치 시 'Add Python to PATH' 옵션을 꼭 체크하세요 (Windows).

pip install -r requirements.txt

requirements.txt 등에 명시된 파이썬 라이브러리를 설치합니다.

✅ 에러 메시지 없이 실행되고 터미널에 안내 문구가 출력되면 정상입니다.

이 레포의 README에 적힌 실제 명령어를 그대로 가져왔습니다.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# midihum

midihum (the tool formerly known as rachel) is a command-line tool for humanizing MIDI -- that is, for taking as input MIDI compositions and producing as output those same compositions with new velocity (loudness/dynamics) values for each of the contained notes. It uses gradient boosted trees, with ~400 engineered features, and is trained on 2.6K competition piano performances. For more information, see [this blog post](https://www.erichgrunewald.com/posts/introducing-midihum-an-ml-based-midi-humanizing-tool/).

midihum requires Python 3. It has been tested on macOS Ventura 13.0.1 and Debian 11.

## Usage

Using midihum is easy. First clone the repository, navigate to the midihum/ directory, and install dependencies:

```shell
pip install -r requirements.txt
```

Then -- making sure you're still in the midihum/ directory -- simply:

```shell
python main.py humanize /path/to/file.mid /path/to/humanized_file.mid
```

(Or use `python3` instead of `python` if that is your Python 3 binary.)

## Performance

The midihum program performs surprisingly well, at least for solo piano works of roughly the type it was trained on, i.e., from the Baroque, Classical, and especially Romantic periods of Western art music. Here are the true velocities and the predicted velocities for nine randomly chosen, not-cherry-picked performances from the validation set:

![True versus predicted velocity scatter plot](https://raw.githubusercontent.com/erwald/midihum/master/midihum_velocity_scatter.png)

In the above plot, each dot is a note on event, randomly sampled from the piece. For the same nine pieces, the model also captures the dynamics over the course of the composition well:

![True versus predicted velocities over time](https://raw.githubusercontent.com/erwald/midihum/master/midihum_velocities_over_time.png)

## Model

The program uses [XGBoost](https://xgboost.readthedocs.io/en/stable/) gradient boosted trees for its model, where each observation is one MIDI "note on" event (see `midihum_model.py`), with a large (~400) set of derived features (see `midi_to_df_conversion.py`).

## Acknowledgements

The model is trained on 2,579 performances from the International Piano-e-Competition for pianists aged 35 and under. The midihum tool is dedicated to those talented young performers.