conformersed's Introduction

ConformerSED

Implementation of the paper CONFORMER-BASED SOUND EVENT DETECTION WITH SEMI-SUPERVISED LEARNING AND DATA AUGMENTATION

Dataset

To run this repository's script, please make sure the dataset is cofigured as follows:

└── dcase20_task4
     └── dataset
         └── audio
             ├── train
             |   ├── synthetic20
             |   ├── weak
             |   └── unlabel_in_domain
             ├── validation
             └── public

To acquire the dataset, run stage 1 in bin/run.sh or download directly.

Quick start

$ conda env create -f environment.yaml
$ conda activate dcase20t4
$ wandb login <YOUR_API_KEY>
$ bash ./bin/run.sh

Results

	Event-based macro F1	PSDS
Validation	47.7%	0.637
Public	49.0%	0.681

To reproduce the result on public eval set, download pretrained model and config file here.

Then, put model and config as follows:

├── data
|    └── feat
|        └── sr16000_n_mels64_n_fft1024_n_shift323
|            └── public
└── exp
    └── conformer_sed                                  
        ├── model
        │   └── model_best_score.pth
        ├── model_config.yaml
        ├── trainer_config.yaml
        ├── post_process_params.pickle
        └── stats.npz

After that, you can run test script

$ python ./src/test.py

Disclaimer

Some audio samples in the original dataset may be missing for some reason (e.g., removed on the internet). For audio samples used in this experiment, see metadata.
This repository results are slightly better than reported in the paper.