aascode / DCASE2021

The official implementations of our DCASE 2021 task 1a

DCASE2021

This repository is the official implementations of our DCASE 2021 task 1a with technical report:

Soonshin Seo, Ji-Hwan Kim: "MobileNet using Coordinate Attention and Fusions for Low-Complexity Acoustic Scene Classification with Multiple Devices", submitted to task 1a of the 2021 DCASE Challenge

A technical report link at (http://dcase.community/documents/challenge2021/technical_reports/DCASE2021_Seo_52_t1.pdf)

Training

1. Download the development dataset form links at https://zenodo.org/record/3819968#.YLiqhfkzaUk
2. Use the script "feats.py" & data augmentation scripts
3. Use the script "train.py"

Quantization & Evauation

1. Use the script "run.sh"

Main methods

1. Normalization & data augmentations
2. MobileNet
3. Cooridnate attention
4. Early fusion & late fusion

Acknowledgement

We used the implementation presented in https://github.com/MihawkHu/DCASE2020_task1 as our baseline script.

Bibtex

@techreport{Seo_DCASE2021,
  author    = {Soonshin Seo, Ji-Hwan Kim},
  title     = {MobileNet using Coordinate Attention and Fusions for Low-Complexity Acoustic Scene Classification with Multiple Devices},
  institution = {DCASE2021 Challenge},
  year      = {2021},
}