PlyTitle_Generation

This is the official repository of Music Playlist Title Generation: A Machine-Translation Approach. The paper has been accepted by 2nd Workshop on NLP for Music and Spoken Audio co-located with ISMIR'2021.

• Pretrained Weight and Datasplit
• Presentation Video

Inference Results

This model use the track_id sequence as an input and return the playlist title sequence as an output.

• Melon dataset's results can be found here (test set).

    "25269": {
        "ground_truth": "취향저격 감성힙합+α 두번째",
        "prediction": "r&b soul introduction 버벌진트 
     
      "
    },
    "69588": {
        "ground_truth": "걸그룹의 대표적인 곡들",
        "prediction": "신나는 댄스곡 모음 
      
       "
    },
    "66941": {
        "ground_truth": "미리 메리 크리스마스",
        "prediction": "크리스마스 캐롤 크리스마스 캐롤 
       
        "
    },

       
      
     

• Spotify-million-playlist-dataset dataset's results can be found here (test set).

    "923035": {
        "ground_truth": "wedding dinner music",
        "prediction": "wedding - cocktail hour 
     
      "
    },
    "634077": {
        "ground_truth": "history of rap",
        "prediction": "old school hip hop 
      
       "
    }
    "540451": {
        "ground_truth": "metal up your ass",
        "prediction": "rock and roll 
       
        "
    },

       
      
     

Environment

1. Install python and PyTorch:

   • python==3.8.5
   • torch==1.9.0 (Please install it according to your CUDA version, in my case cu111)

2. Other requirements:

   • pip install -r requirements.txt

Training from scratch

1. Download the data files from spotify-million-playlist and Melon Kakao.

2. Run preprocessing code and split dataset

python preprocessing.py

or you can download pre-split dataset from here.

3. Training options (best pramas):

python train.py --dataset_type melon --model transfomer --shuffle True --e_pos False
python train.py --dataset_type mpd --model transfomer --shuffle True --e_pos False

4. Evaluation & Inference

python eval.py --dataset_type melon --model transfomer --shuffle True --e_pos False
python infer.py --dataset_type melon --model transfomer --shuffle True --e_pos False

Reference

https://github.com/bentrevett/pytorch-seq2seq