Skip to content

nttcslab/dcase2023_task2_baseline_ae

BranchesTags

Repository files navigation

Anomalous Sound Detection

DCASE 2025 Challenge Task 2 and DCASE 2023 Challenge Task 2 Baseline Auto Encoder: dcase2023_task2_baseline_ae

This is an autoencoder-based baseline for the DCASE2025 Challenge Task 2 (DCASE2025T2), the DCASE2024 Challenge Task 2 (DCASE2024T2) and the DCASE2023 Challenge Task 2 (DCASE2023T2).

This source code is an example implementation of the baseline Auto Encoder of DCASE2025T2, DCASE2024T2 and DCASE2023T2: First-Shot Unsupervised Anomalous Sound Detection for Machine Condition Monitoring. This baseline implementation is based on the previous baseline, dcase2022_baseline_ae. The model parameter settings of this baseline AE are almost equivalent to those of the dcase2022_task2_baseline_ae.

Differences between the previous dcase2022_baseline_ae and this version are as follows:

  • The dcase2022_baseline_ae was implemented with Keras; however, this version is written in PyTorch.
  • Data folder structure is updated to support DCASE2025T2, DCASE2024T2 and DCASE2023T2 data sets.
  • The system uses the MSE loss as a loss function for training, but for testing, two score functions depend on the testing modes (i.e., MSE for the Simple Autoencoder mode and Mahalanobis distance for the Selective Mahalanobis mode).

Description

This system consists of three main scripts (01_train.sh, 02a_test.sh, and 02b_test.sh) with some helper scripts for DCASE2025T2 (For DCASE2024T2 and DCASE2023T2, see README_legacy):

  • Helper scripts for DCASE2025T2

    • data_download_2025dev.sh Updated on (2025/04/01)
      • "Development dataset":
        • This script downloads development data files and puts them into data/dcase2025t2/dev\_data/raw/train/ and data/dcase2025t2/dev\_data/raw/test/.
    • data_download_2025add.sh Newly added!! (2025/05/15)
      • "Additional train dataset for Evaluation":
        • This script downloads Addition data files and puts them into data/dcase2025t2/eval\_data/raw/train/.

  • 01_train_2025t2.sh

    • "Development" mode: Updated on (2025/04/01)
      • This script trains a model for each machine type for each section ID by using the directory data/dcase2025t2/dev_data/raw/<machine_type>/train/<section_id>.
    • "Evaluation" mode: Newly added!! (2025/05/15)
      • This script trains a model for each machine type for each section ID by using the directory data/dcase2025t2/eval_data/raw/<machine_type>/train/<section_id>.

  • 02a_test_2025t2.sh (Use MSE as a score function for the Simple Autoencoder mode) Updated on (2025/04/01)

    • "Development" mode:
      • This script makes a CSV file for each section, including the anomaly scores for each WAV file in the directories data/dcase2025t2/dev_data/raw/<machine_type>/test/.
      • The CSV files will be stored in the directory results/.
      • It also makes a csv file including AUC, pAUC, precision, recall, and F1-score for each section.

  • 02b_test_2025t2.sh (Use Mahalanobis distance as a score function for the Selective Mahalanobis mode) Updated on (2025/04/01)

    • "Development" mode:
      • This script makes a CSV file for each section, including the anomaly scores for each wav file in the directories data/dcase2025t2/dev_data/raw/<machine_type>/test/.
      • The CSV files will be stored in the directory results/.
      • It also makes a csv file including AUC, pAUC, precision, recall, and F1-score for each section.

  • 03_summarize_results.sh

    • This script summarizes results into a CSV file.

C.f., for DCASE2023T2, see README_legacy.

Usage

1. Clone repository

Clone this repository from GitHub.

2. Download datasets

We will launch the datasets in three stages. Therefore, please download the datasets in each stage:

3. Unzip the downloaded files and make the directory structure as follows:

  • dcase2023_task2_baseline_ae
    • data/dcase2024t2/dev_data/raw/
      • <machine_type0_of_additional_dataset>/
        • train (only normal clips)
          • section_00_source_train_normal_0000_.wav
          • ...
          • section_00_source_train_normal_0989_.wav
          • section_00_target_train_normal_0000_.wav
          • ...
          • section_00_target_train_normal_0009_.wav
        • test/
          • section_00_source_test_normal_0000_.wav
          • ...
          • section_00_source_test_normal_0049_.wav
          • section_00_source_test_anomaly_0000_.wav
          • ...
          • section_00_source_test_anomaly_0049_.wav
          • section_00_target_test_normal_0000_.wav
          • ...
          • section_00_target_test_normal_0049_.wav
          • section_00_target_test_anomaly_0000_.wav
          • ...
          • section_00_target_test_anomaly_0049_.wav
        • attributes_00.csv (attributes CSV for section 00)
      • <machine_type1_of_additional_dataset> (The other machine types have the same directory structure as <machine_type0_of_additional_dataset>/.)
  • data/dcase2025t2/eval_data/raw/
    • <machine_type0_of_additional_dataset>/
      • supplemental/ (after launch of the additional training dataset)
        • section_00_machine_0001_.wav
        • ...
        • section_00_machine_0100_.wav
      • train/ (after launch of the additional training dataset)
        • section_00_source_train_normal_0000_.wav
        • ...
        • section_00_source_train_normal_0989_.wav
        • section_00_target_train_normal_0000_.wav
        • ...
        • section_00_target_train_normal_0009_.wav
      • attributes_00.csv (attributes CSV for section 00)
    • <machine_type1_of_additional_dataset> (The other machine types have the same directory structure as <machine_type0_of_additional_dataset>/.)

4. Change parameters

You can change parameters for feature extraction and model definition by editing baseline.yaml. Note that if values are specified in the command line option, it will overwrite the parameter settings in baseline.yaml.

4.1. Enable Auto-download dataset

If you haven't yet downloaded the dataset yourself nor you have not run the download script (example, data_download_2025dev.sh) then you may want to use the auto download. To enable the auto-downloading, set the parameter --is_auto_download (default: False) True in baseline.yaml. If --is_auto_download is True, then auto-download is executed.

5. Run the training script (for the development dataset)

Run the training script 01_train_2025t2.sh. Use the option -d for the development dataset data/dcase2025t2/dev_data/<machine_type>/raw/train/.

# using DCASE2025 Task 2 Datasets
$ 01_train_2025t2.sh -d

The two operating modes of this baseline implementation, the simple Autoencoder, and the Selective Mahalanobis AE modes, share the common training process. By running the script 01_train_2025t2.sh, all the model parameters for the simple Autoencoder and the selective Mahalanobis AE will be trained at the same time. After the parameter update of the Autoencoder at the last epoch specified by either the yaml file or the command line option, the covariance matrixes for the Mahalanobis distance calculation will be set.

6. Run the test script (for the development dataset)

6.1. Testing with the Simple Autoencoder mode

Run the test script 02a_test_2025t2.sh. Use the option -d for the development dataset data/dcase2025t2/dev_data/<machine_type>/raw/test/.

# using DCASE2025 Task 2 Datasets
$ 02a_test_2025t2.sh -d

The 02a_test_2025t2.sh options are the same as those for 01_train_2025t2.sh. 02a_test_2025t2.sh calculates an anomaly score for each wav file in the directories data/dcase2025t2/dev_data/raw/<machine_type>/test/ or data/dcase2025t2/dev_data/raw/<machine_type>/source_test/ and data/dcase2025t2/dev_data/raw/<machine_type>/target_test/. A CSV file for each section, including the anomaly scores, will be stored in the directory results/. If the mode is "development", the script also outputs another CSV file, including AUC, pAUC, precision, recall, and F1-score for each section.

6.2. Testing with the Selective Mahalanobis mode

Run the test script 02b_test_2025t2.sh. Use the option -d for the development dataset data/dcase2025t2/dev_data/<machine_type>/raw/test/.

# using DCASE2025 Task 2 Datasets
$ 02b_test_2025t2.sh -d

The 02b_test_2025t2.sh options are the same as those for 01_train_2025t2.sh. 02b_test_2025t2.sh calculates an anomaly score for each wav file in the directories data/dcase2025t2/dev_data/raw/<machine_type>/test/ or data/dcase2025t2/dev_data/raw/<machine_type>/source_test/ and data/dcase2025t2/dev_data/raw/<machine_type>/target_test/. A CSV file for each section, including the anomaly scores, will be stored in the directory results/. If the mode is "development", the script also outputs another CSV file, including AUC, pAUC, precision, recall, and F1-score for each section.

7. Check results

You can check the anomaly scores in the CSV files anomaly_score_<machine_type>_section_<section_index>_test.csv in the directory results/. Each anomaly score corresponds to a wav file in the directories data/dcase2025t2/dev_data/<machine_type>/test/.

anomaly_score_<machine_type>_section_00_test.csv

section_00_source_test_normal_0000_car_A2_spd_28V_mic_1_noise_1.wav,0.3084583878517151
section_00_source_test_normal_0001_car_A2_spd_28V_mic_1_noise_1.wav,0.31289517879486084
section_00_source_test_normal_0002_car_A2_spd_28V_mic_1_noise_1.wav,0.4160425364971161
section_00_source_test_normal_0003_car_A2_spd_28V_mic_1_noise_1.wav,0.25631701946258545
...

Also, anomaly detection results based on the corresponding threshold can be checked in the CSV files decision_result_<machine_type>_section_<section_index>_test.csv:

decision_result_<machine_type>_section_<section_index>_test.csv

section_00_source_test_normal_0000_car_A2_spd_28V_mic_1_noise_1.wav,0
section_00_source_test_normal_0001_car_A2_spd_28V_mic_1_noise_1.wav,0
section_00_source_test_normal_0002_car_A2_spd_28V_mic_1_noise_1.wav,0
section_00_source_test_normal_0003_car_A2_spd_28V_mic_1_noise_1.wav,0
...

In addition, you can check performance indicators such as AUC, pAUC, precision, recall, and F1-score:

result.csv

section, AUC (source), AUC (target), pAUC, pAUC (source), pAUC (target), precision (source), precision (target), recall (source), recall (target), F1 score (source), F1 score (target)
00,0.88,0.5078,0.5063157894736842,0.5536842105263158,0.4926315789473684,0.0,0.0,0.0,0.0,0.0,0.
arithmetic mean,00,0.88,0.5078,0.5063157894736842,0.5536842105263158,0.4926315789473684,0.0,0.0,0.0,0.0,0.0,0.
harmonic mean,00,0.88,0.5078,0.5063157894736842,0.5536842105263158,0.4926315789473684,0.0,0.0,0.0,0.0,0.0,0.

8. Run training script for the additional training dataset Newly added!! (2025/05/15)

After the additional training dataset is launched, download and unzip it. Move it to data/dcase2025t2/eval_data/raw/<machine_type>/train/. Run the training script 01_train_2025t2.sh with the option -e.

$ 01_train_2025t2.sh -e

Models are trained by using the additional training dataset data/dcase2025t2/raw/eval_data/<machine_type>/train/.

9. Run the test script for the evaluation dataset (after June 1, 2025)

10. Summarize results

After the executed 02a_test_2025t2.sh, 02b_test_2025t2.sh, or both. Run the summarize script 03_summarize_results.sh with the option DCASE2025T2 -d.

# Summarize development dataset 2025
$ 03_summarize_results.sh DCASE2025T2 -d

After the summary, the results are exported in CSV format to results/dev_data/baseline/summarize/DCASE2025T2.

If you want to change, summarize results directory or export directory, edit 03_summarize_results.sh.

Legacy support

This version takes the legacy datasets provided in DCASE2020 task2, DCASE2021 task2, DCASE2022 task2, DCASE2023 task2, and DCASE2024 task2 dataset for inputs. The Legacy support scripts are similar to the main scripts. These are in tools directory.

learn more

Dependency

We developed and tested the source code on Ubuntu 22.04.5 LTS.

Software package

  • Python == 3.11.11
  • cuda == 11.8
  • libsndfile1

Python packages

  • Pytorch == 2.6.0+cu118
  • torchvision == 0.21.0+cu118
  • numpy >= 2.1.3
  • pyYAML == 6.0.2
  • scipy >= 1.15.2
  • librosa >= 0.10.2.post1
  • matplotlib >= 3.10.1
  • tqdm >= 4.67.1
  • seaborn >= 0.13.2
  • fasteners >= 0.19

Change Log

4.1.1

Fixed the function to download datasets

  • Fixed an error that occurred when using the automatic download function.
  • The path to unzip the DCASE2025T2 Additional training dataset was incorrect. Fixed the issue in data_download_2025add.sh.

4.1.0

Added

  • Provides support for the additional training datasets to be used in DCASE2025T2.

4.0.1

Previous DCASE Task2 scripts have been moved into legacy

  • The DCASE2023 and DCASE2024 scripts have been removed from the top directory.
  • You can continue to use the DCASE2020 through DCASE2024 scripts by using the scripts in the legacy directory.

4.0.0

Added

  • Provides support for the development datasets used in DCASE2025T2.

3.3.0

Made DCASE2024 Task2 legacy

3.2.3

Updated citation in README

  • Updated the README and README legacy files to reflect the latest citations.

3.2.2

Fixed test script

  • Updated to fix issues in 02a_test_2024t2.sh and 02b_test_2024t2.sh.

3.2.1

Fixed download shell script

  • The URL to download DCASE2024T2 Evaluation dataset was incorrect. Updated to fix issues in data_download_2024eval.sh.

3.2.0

Added

  • Provides support for the evaluation dataset to be used in DCASE2024T2.

3.1.1

Updated the Zenodo link for the DCASE2024T2 additional training dataset

  • We have corrected the Additional training datasets data used in DCASE2024T2.
    • eval_data_3DPrinter_train.zip and eval_data_RoboticArm_train.zip were replaced with eval_data_3DPrinter_train_r2.zip and eval_data_RoboticArm_train_r2.zip
    • For the other machine types, data files are identical.
    • This version reflects the update on the additional training dataset.

3.1.0

Added

  • Provides support for the additional training datasets to be used in DCASE2024T2.

3.0.2

Added

  • Added information about ground truth and citations for each year's task in README.md and README_legacy.md.

3.0.1

Added

  • Provides support for the legacy datasets used in DCASE2020, 2021, 2022, and 2023.

Fixed

  • Fixed a typo in README.md in the previous release, v3.0.0.

3.0.0

Added

  • Provides support for the development datasets used in DCASE2024.

2.0.1

Fixed

  • Fix anomaly score distribution.
  • Decision threshold has changed, but AUC, pAUC, etc. have not changed.

2.0.0

Added

  • Provides support for the legacy datasets used in DCASE2020, 2021, 2022, and 2023.

Ground truth attribute

Ground truth attributes

Citation

If you use this system, please cite all the following four papers:

  • Tomoya Nishida, Noboru Harada, Daisuke Niizumi, Davide Albertini, Roberto Sannino, Simone Pradolini, Filippo Augusti, Keisuke Imoto, Kota Dohi, Harsh Purohit, Takashi Endo, and Yohei Kawaguchi. Description and discussion on DCASE 2024 challenge task 2: first-shot unsupervised anomalous sound detection for machine condition monitoring. In arXiv e-prints: 2406.07250, 2024. URL
  • Noboru Harada, Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Masahiro Yasuda, and Shoichiro Saito. ToyADMOS2: another dataset of miniature-machine operating sounds for anomalous sound detection under domain shift conditions. In Proceedings of the Detection and Classification of Acoustic Scenes and Events Workshop (DCASE), 1–5. Barcelona, Spain, November 2021. URL
  • Kota Dohi, Tomoya Nishida, Harsh Purohit, Ryo Tanabe, Takashi Endo, Masaaki Yamamoto, Yuki Nikaido, and Yohei Kawaguchi. MIMII DG: sound dataset for malfunctioning industrial machine investigation and inspection for domain generalization task. In Proceedings of the 7th Detection and Classification of Acoustic Scenes and Events 2022 Workshop (DCASE2022). Nancy, France, November 2022. URL
  • Noboru Harada, Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, and Masahiro Yasuda. First-shot anomaly detection for machine condition monitoring: a domain generalization baseline. Proceedings of 31st European Signal Processing Conference (EUSIPCO), pages 191–195, 2023. URL

About

No description, website, or topics provided.

Resources

Readme
Activity
Custom properties

Stars

72 stars

Watchers

5 watching

Forks

18 forks
Report repository

Releases 18

v4.1.1 Latest
May 19, 2025
+ 17 releases

Packages

No packages published

Contributors 2

Languages