Improvements of Discriminative Feature Space Training for Anomalous Sound Detection in Unlabeled Conditions

venv

• Make venv (unlabledasd/venv)
  • python version was 3.10.10
  • python3 -m venv venv
  • source venv/bin/activate
  • pip install -r requirements.txt

Data

• Download DCASE2023 and DCASE2024 data
• Modify the data structure as follows: (You can also use symbolic links)
  • <data_dir>/dcase2023/all/raw/bandsaw/train
  • ...
  • <data_dir>/dcase2023/all/raw/valve/test
  • <data_dir>/dcase2024/all/raw/3DPrinter/train
  • ...
  • <data_dir>/dcase2024/all/raw/valve/test
• Add ground-truth information
  • This process renames the filenames of the test set using ground-truth labels.
  • Set data_dir to <data_dir> in the preprocess/add_gt_info.sh.
  • cd preprocess and execute ./add_gt_info.sh

Training and Testing

• The full shell script of the experiments is provided in the jobs/exp

• exp1.sh, exp2.sh, and exp3_???.sh correspond to Tables 1, 2, and 3, respectively.

• Set data_dir in each shell script and execute it

• These shell scripts run experiments with five different seeds and several methods using a simple for loop. If you want to parallelize the process or run it selectively, you can split the for loop.

• When you execute the shell scripts, the training will strat, and the results will be stored in results/exp/<dcase>/<method>/<machine>/infer/version_epoch<epoch>/*_test_result.csv

  • infer/version_epoch_12_14_16 is the result of an ensemble of anomaly scores for the 12, 14, and 16 epochs
  • hmean_official of infer/versoin_epoch<epoch>/*_test_result.csv is the official score of DCASE Task2 Challenge
  • hmean_official is the harmonic mean of the AUC of the source domain (0_source_auc_all), the AUC of the target domain (0_target_auc_all), and the pAUC of both domains (0_all_pauc_all)

• Before executing exp3.sh, pseudo labels should be generated (see next section).

Pseudo-labeling

• pseudoattr/pseudo_attr.sh generates several types of pseudo-labels using a simple for loop. If you want to parallelize the process or run it selectively, you can split the for loop.

• Some pseudo-labeling methods require some preparation. See pseudoattr/README.md for more details.

• Here is a correspondence table between the names in the experiment and the names in the paper

  Name@paper	2023	2024
  N/A	pattr_macdom_subloss_0256_4096 (exp3_NA23.sh)	subloss_0256_4096 (exp1.sh)
  Class	pattr_bic_pre_subloss_0256_4096 (exp3_PL.sh)	pattr_bic_pre_subloss_0256_4096 (exp3_PL.sh)
  Triplet	pattr_bic_triplet_subloss_0256_4096 (exp3_PL.sh)	pattr_bic_triplet_subloss_0256_4096 (exp3_PL.sh)
  PANNs	pattr_bic_panns_subloss_0256_4096 (exp3_PL.sh)	pattr_bic_panns_subloss_0256_4096 (exp3_PL.sh)
  OpenL3	pattr_bic_openl3_subloss_0256_4096 (exp3_PL.sh)	pattr_bic_openl3_subloss_0256_4096 (exp3_PL.sh)
  GT	subloss_0256_4096 (exp1.sh)	pattr_gt_subloss_0256_4096 (exp3_GT24.sh)

Citation

@article{fujimura2024improvements,
  title={Improvements of Discriminative Feature Space Training for Anomalous Sound Detection in Unlabeled Conditions},
  author={Fujimura, Takuya and Kuroyanagi, Ibuki and Toda, Tomoki},
  journal={arXiv preprint arXiv:2409.09332},
  year={2024}
}