/2022-2-NLP-music-BERT

Primary LanguageJupyter Notebook

Music XAI Improvement

Author: 강성현, 양현서, Richard Novenius

Fine Tuning while Freezing Bottom Encoder Layers

branch: ludrex
  • In musicbert/__init__.py, modify the build_model method of class MusicBERTSentencePredictionMultilabelTaskXAI(line 166), following the instruction given in the comment.
  • You may either
  1. freeze no parameters(default),
  2. freeze all parameters in the encoder and only update classifier parameters, or
  3. fix $k$ and freeze parameters in encoder layers $0,1,\ldots,k$, where $$0\le k\le \text{(number of encoder layers)}-1.$$ MusicBERT base and small models have 12 and 4 encoder layers, respectively.
  • To fine-tune,
bash scripts/classification/train_xai_base_small.sh
  • To evaluate fine-tuned models and get test accuracy,
bash scripts/eval_xai.sh

Data Augmentation 

branch: attempt/augmentation
  • method:
cd midi_augmentator
pip install -r requirements.txt # mido
python augmentation.py ../processed/segmented_midi/ --pitch_range=-6_12_1 --tempo_range 60_180_60 --velocity_range 50_90_40
  • Generated pitch = -6, -5, -4, -3, ..., +12
  • Generated tempo = 60, 120, 180
  • Generated velocity = 50, 90
  • Generated file name example = {original_filename}_{pitch}_-12_{tempo}_120_{velocity}_90.mid
  • Run python generate_total_csv.py segmented_midi/_augmented/ to generate new total.csv
  • Then run modified python map_midi_to_label.py to generate midi_label_map_apex_reg_cls.json file.
  • Then run modified python -u gen_xai.py xai to generate xai_data_raw_apex_reg_cls_augmented folder.
  • Then run modified bash scripts/binarize_xai.sh xai to generate xai_data_bin_apex_reg_cls_augmented folder.
  • Then run modified scripts/classification/train_xai_base_small.sh to train the model.