Datasets:
audio audioduration (s) 10 11 | label class label 0
classes |
|---|---|
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null | |
null |
DCASE 2020 Task 2 — Development Dataset (STgram-MFN redistribution)
Redistribution of the DCASE 2020 Challenge Task 2 Development Dataset (Zenodo record 3678171) in the exact on-disk layout expected by the STgram-MFN reference runs. It contains MIMII and ToyADMOS normal/anomalous machine sounds for unsupervised anomalous-sound detection (ASD).
Original authors: Yuma Koizumi, Yohei Kawaguchi, Keisuke Imoto. License: CC BY-NC-SA 4.0 (non-commercial) — inherited from the source; attribute the original authors and keep any redistribution under the same terms.
Layout
<machine>/{train,test}/<label>_id_<XX>_<index>.wav
<machine>∈fan, pump, slider, valve, ToyCar, ToyConveyortrain/— normal onlytest/— normal + anomaly- audio: mono, 16 kHz, 10 s (
.wav, 16-bit PCM)
Labels
There is no separate label file — everything is encoded in the path:
| Field | Source | Example |
|---|---|---|
| machine type | parent directory | fan |
| machine id (0–7) | id_XX token |
id_00 |
| normal / anomaly | filename prefix | anomaly → 1 |
STgram-MFN trains self-supervised by machine id (machine-id,
8 ids × 6 machines = 42 classes); the normal/anomaly split is used only for
AUC/pAUC evaluation, never as a training target.
Statistics
- 6 machine types, 30,987 clips, ~9.7 GB
train: normal only ·test: normal + anomaly
Intended use
Reference backbone training + evaluation for the backbone-swap / INT8
quantization study (STgram-MFN vs. EfficientAT mn01) in
LakoreAI/Research_AnomalySoundDetection. Evaluate with the official DCASE
protocol (AUC / pAUC / mAUC).
Usage
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="LakoreAI/stgram-mfn-dcase2020-dev",
repo_type="dataset", local_dir="data/raw",
)
uv run python scripts/data/prepare_data.py --root data/raw --check
Related
- Evaluation / additional-training set:
LakoreAI/stgram-mfn-dcase2020-eval - Source: https://zenodo.org/records/3678171
Citation
@inproceedings{koizumi2020dcase,
title = {Description and Discussion on DCASE2020 Challenge Task 2:
Unsupervised Anomalous Sound Detection for Machine Condition Monitoring},
author = {Koizumi, Yuma and Kawaguchi, Yohei and Imoto, Keisuke and others},
booktitle = {DCASE Workshop},
year = {2020}
}
- Downloads last month
- 58