Multi-tasks for Lyrics Classifcation

Unofficial code for paper: Transformer-based approach towards music emotion recognition from lyrics

• Paper link

Requirements

• Pytorch (1.10.1+cu113)
• Numpy
• Pandas
• Tensorboard
• Transformers (3.0.0)

Datasets

We used MoodyLyrcis dataset for this repo. Due to potential copyright issues, cannot share the training data.

1. You have to put train.csv and test.csv in folder ./multitask/data/.
2. The csv columns we used be like:

index,artist,song,mood,lyric

3. Only need two columns: mood and lyric in the training process. You can adjust the costom dataset multitasks/utils/dataset.py to fit your own dataset.

Enviroment Setup

1. PyTorch with CUDA-enabled GPU.
2. Change hyper-parameters at file: ./multitask/config.py (Optional)

Run Training

Follow these stepes to run:

cd multitasks
python main.py

Results

Run command:

tensorboard --logdir='{log_path_you_define_at_config.py}'

Tensorboard screen shot