EMOPIA_cls

This is the official repository of EMOPIA: A Multi-Modal Pop Piano Dataset For Emotion Recognition and Emotion-based Music Generation. The paper has been accepted by International Society for Music Information Retrieval Conference 2021. This repository is the Emotion Recognition part (Audio and MIDI domain).

News!

2021-10-29 update matlab feature (key, tempo, note density)

2021-07-21 update dataset

2021-07-20 Upload all pretrained weight

you can check ML performance in notebook

Environment

Install python and PyTorch:
- python==3.8.5
- torch==1.8.0 (Please install it according to your CUDA version.)
Other requirements:
- pip install -r requirements.txt
git clone MIDI processor (already done)
- MIDI-like(magenta)
- REMI
- If you want to bulid new REMI corpus, vocab from other dataset, plz check official repo of compund-word-transfomer and EMOPIA_cls/midi_cls/midi_helper/remi/src

Usage

Inference

download model weight in Here, unzip in project dir.

MIDI domain inference

 python inference.py --types {midi_like or remi} --task ar_va --file_path {your_midi} --cuda {cuda}
 python inference.py --types {midi_like or remi} --task arousal --file_path {your_midi} --cuda {cuda}
 python inference.py --types {midi_like or remi} --task valence --file_path {your_midi} --cuda {cuda}

Audio domain inference

 python inference.py --types wav --task ar_va --file_path {your_mp3} --cuda {cuda}
 python inference.py --types wav --task arousal --file_path {your_mp3} --cuda {cuda}
 python inference.py --types wav --task valence --file_path {your_mp3} --cuda {cuda}

Inference results

    python inference.py --types wav --task ar_va --file_path ./dataset/sample_data/Sakamoto_MerryChristmasMr_Lawrence.mp3

    ./dataset/sample_data/Sakamoto_MerryChristmasMr_Lawrence.mp3  is emotion Q3
    Inference values:  [0.33273646 0.17223473 0.63210356 0.07314324]

    python inference.py --types midi_like --task ar_va --file_path ./dataset/sample_data/Sakamoto_MerryChristmasMr_Lawrence.mid

    ./dataset/sample_data/Sakamoto_MerryChristmasMr_Lawrence.mid  is emotion Q3
    Inference values:  [-1.3685153 -1.3001229  2.2495744 -0.873877 ]

Training from scratch

Download the data files from HERE.
Preprocessing

a. audio: resampling to 22050

b. midi: magenta feature extraction, remi feature extraction
```
 python preprocessing.py
```

training options:

a. MIDI domain classification

 cd midi_cls
 python train_test.py --midi {midi_like or remi} --task ar_va
 python train_test.py --midi {midi_like or remi} --task arousal
 python train_test.py --midi {midi_like or remi} --task valence

b. Wav domain clasfficiation

 cd audio_cls
 python train_test.py --wav sr22k --task ar_va
 python train_test.py --wav sr22k --task arousal
 python train_test.py --wav sr22k --task valence

Authors

The paper is a co-working project with Anna, Joann and Nabin. This repository is mentained by me.

License

The EMOPIA dataset is released under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). It is provided primarily for research purposes and is prohibited to be used for commercial purposes. When sharing your result based on EMOPIA, any act that defames the original music owner is strictly prohibited.

Cite the dataset

@inproceedings{{EMOPIA},
         author = {Hung, Hsiao-Tzu and Ching, Joann and Doh, Seungheon and Kim, Nabin and Nam, Juhan and Yang, Yi-Hsuan},
         title = {{MOPIA}: A Multi-Modal Pop Piano Dataset For Emotion Recognition and Emotion-based Music Generation},
         booktitle = {Proc. Int. Society for Music Information Retrieval Conf.},
         year = {2021}
}

ChungHaLee/EmoPIA