ZSL_music_tagging

Zero-shot learning for music auto-tagging and classification. This code works for both MSD and FMA dataset. All pre-filtering of tracks and tags were performed beforehand. We provide the list of tracks/tags and their necessary matrices.

Zero-shot Learning for Audio-based Music Classification and Tagging, Jeong Choi*, Jongpil Lee*, Jiyoung Park, Juhan Nam (* : equally contributed authors) Accepted at ISMIR 2019

Separately provided data : Link

├─ data_common
  ├─ msd
    ├─ tag_key_split_TGSPP.p 
        : tag split used in paper
    ├─ track_keys_AB_TRSPP_TGSPP.p 
    ├─ track_keys_A_TRSPP_TGSPP.p
    ├─ track_keys_B_TRSPP_TGSPP.p
    ├─ track_keys_C_TRSPP_TGSPP.p
        : track splits used in paper
          
    ├─ all_tag_to_track_bin_matrix.p
    ├─ tag_ids_in_key_order.p
    ├─ track_ids_in_key_order.p
    ├─ track_id_to_file_path_dict.p
    ├─ tag_key_to_id_dict.p 
       
  ├─ fma
    ├─ ... : same as above 

       
├─ data_tag_vector
  ├─ msd
    ├─ ttr_ont_tag_1126_to_glove_dict.p
        : GloVe vector data 
          (filtered using Tagtraum genre ontology) 
  ├─ fma
    ├─ genre_id_to_inst_posneg40_cnt_norm_dict.p
    ├─ genre_id_to_inst_posneg40_conf_norm_dict.p    
        : Instrument vector data

Audio (mel-spectogram) preparation (in 'scripts' folder)

python preprocess_audio_msd.py --dir_wav PATH_TO_MSD_AUDIO_WAV --dir_mel PATH_FOR_SAVING_MEL_FILES

Tag/track split data preparation (in 'scripts' folder)

First, prepare tag splits (train / test tags)

python data_split_tag.py --dataset msd --tag_split_name TGS01 

Using the tag split, prepare track splits (train / valid for AB, A, B)

python data_split_track.py --dataset msd --tag_split_name TGS01  --track_split_name TRS01 

Model training / inference / evaluation

Training

python train.py --dataset msd --track_split_name TRS01 --tag_split_name TGS01 --tag_vector_type glove --epochs 20 --track_split A

Inference

python extract_embeddings_multi.py --load_weights PATH_TO_WEIGHT_FILE

Evaluation

python eval.py --load_weights PATH_TO_WEIGHT_FILE