Skip to content

Latest commit

 

History

History
98 lines (68 loc) · 4.27 KB

README.md

File metadata and controls

98 lines (68 loc) · 4.27 KB

Prog-rock vs. Everything Else

An ML algorithm to distinguish Progressive Rock music from everything else.

Extract and save feature

Run save_feature.py.

  • Each subfolder in the tran_set\ has an .ods file stating the list of songs, and .mp3 files for songs.

    • Progressive_Rock_Songs\: {'songs': 142, 'others': ['prog_train.ods']}
    • Not_Progressive_Rock\Top_Of_The_Pops\: {'songs': 87, 'others': ['notprog_top_pops_train.ods']}
    • Not_Progressive_Rock\Other_Songs\: {'songs': 272, 'others': ['notprog_other_train.ods']}

  • Load music files using librosa.

    • It has features of
      • Spectral
        • Chromagram
        • Mel-scaled spectrogram
        • Mel-frequency cepstral coefficients (MFCCs)
      • Rythym
        • Tempo
      • Others
    • Feature plots saved at output/

Models

We build several baseline models based on the last years' best project, as well as several new models.

1. Baseline

conv1d(43,64) -> conv1d(64,64) -> linear(6848,600) -> linear(600,10) -> linear(10,2)

2. ModifiedBaseline

conv1d(43,64) -> conv1d(64,128) -> linear(13696,600) -> linear(600,30) -> linear(30,4) -> linear(4,2)

3. DeepBaseline

conv1d(43,64) -> conv1d(64,128) -> conv1d(128,256) -> conv1d(256,512) -> linear(13312,100) -> linear(100,10) -> linear(10,2)

4. DeepWiseBaseline

conv1d(43,86) -> conv1d(86,172) -> conv1d(172,344) -> linear(18232,2)

5. AcousticModel (O’Brien 2016)

conv1d(43,64) -> conv1d(64,128) -> conv1d(128,128) -> conv1d(128,64) -> conv1d(64,32)-> conv1d(32,32) -> linear(832,2)

  • O’Brien, Tim. "Musical Structure Segmentation with Convolutional Neural Networks." 17th International Society for Music Information Retrieval Conference. 2016.

6. GenreModel (End-to-end, )

conv1d(43,128) -> conv1d(128,128) -> conv1d(128,256) -> conv1d(256,256) -> conv1d(256,256) -> conv1d(256,256) -> conv1d(256,512)-> conv1d(512,10) -> linear(40,2)

  • Pons, O. Nieto, M. Prockup, E. Schmidt, A. Ehmann, and X. Serra, “End-to-end learning for music audio tagging at scale,” in Intl Society for Music Inf Retrieval Conf, 2018, pp. 1–8.

7. ResnetModel (Allamy and Alessandro 2021)

conv1d(43,128) -> res1d(128,128) -> res1d(128,256) -> … -> res1d(256,512)-> conv1d(512,10) -> linear(40,2)

  • Allamy, Safaa, and Alessandro Lameiras Koerich. "1D CNN architectures for music genre classification." 2021 IEEE symposium series on computational intelligence (SSCI). IEEE, 2021.

8. LSTM Model

Recurrent NN

Run algorithm

  • Place saved feature json files in relative path to this repo at ../data/[feature.json]. The default paths are set in train.py as:

    non_prog_other_path = "../data/Feature_Extraction_Other.json"
non_prog_pop_path = "../data/Feature_Extraction_Top_Pop.json"
prog_path = "../data/Feature_Extraction_Prog.json"

  • Run main.py.

  • For new CNN models, just import them into main.py and add corresponding (model_name, model()) to model_dict in main.py.

Output

  • Feature plots are saved in output folder as .pdf.

  • Model results are saved in output/model folder named after the model name. These include

    • Confusion matrices for train/test snippets/songs (all in one file)
    • Average confusion matrix after multiple runs
    • Model test labeling result
    • Model pickle file

  • Log file is generated at output/log_file.log

Post Analysis

  • Increasing the cutoff value of binary classification in the last layer of neural network
  • Increasing the cutoff value of the proportion of snippets to become a Prog Rock song

Conclusion

  • We try 8 different models (CNN Structure, Recurrent Structure, and ResNet structure) with two types of snippets (non-overlap and 50% overlap).
  • ResNet structure with normal snippets provides best prediction accuracy.
  • With post-processing techniques, we can further improve the model accuracy to 82.64%, which has 3.5% improvement compared to the baseline model.
  • We still find some type of musics that are hard to be classified under our criterion.
  • For future work, we suggest:
    1. More advanced post-processing classification techniques;
    2. More advanced image classification techniques Neural ODEs (Cui et al. 2023);
    3. Extract other features like lyrics.