GenRepASD

Pytorch implementation of Deep Generic Representations for Domain-Generalized Anomalous Sound Detection: https://arxiv.org/abs/2409.05035

Setting up

1. Install the requirements pip install -r requirements.txt

2. Download the DCASE2020T2 and DCASE2023T2 datasets and place them according to the given directory structure specified in data_config.yaml.

3. Download pre-trained weights of BEATs from https://github.com/microsoft/unilm/tree/master/beats and place them in a pre-trained directory.

Domain-shift experiment on DCASE2023T2 Eval set.

Run GenRep using MemMixup with Ks=990 (best performance).

python run_genrep_dcase2023.py \
    --dataset_name dcase2023 \
    --model_name beats_ft1 \
    --pretrained_model_dir <path_to_pretrained_model> \
    --temporal_pooling \
    --n_mix_support 990 \
    --alpha 0.9 \
    --save_official

After run the above script, you can also evaluate the performance using DCASE2023T2 official evaluator (you should get the same result as above):

bash dcase2023_task2_evaluator/03_evaluation_eval_data.sh

Low-shot experiment on DCASE2020T2 Dev set.

with 200-shot

python run_genrep_dcase2020.py \
    --dataset_name dcase2020 \
    --model_name beats_ft1 \
    --pretrained_model_dir <path_to_pretrained_model> \
    --temporal_pooling \
    --num_samples 200

Acknowledgement

• We thanks the authors of BEATs for providing the pre-trained weights.
• We thanks SPADE, ssl4asd, STgram-MFN, DCASE2023 for providing the codebase and evaluation scripts.

Citation

If you find this work useful, please consider citing:

@misc{saengthong2024deep,
    title={Deep Generic Representations for Domain-Generalized Anomalous Sound Detection},
    author={Phurich Saengthong and Takahiro Shinozaki},
    year={2024},
    eprint={2409.05035},
    archivePrefix={arXiv},
    primaryClass={cs.SD}
}