1. The official repository for the project ESRelation:

Paper Title: Exploring Inter-Node Relations in CNNs for Environmental Sound Classification

2. ☞ Project Page for dynamic visualization examples.

3. Trained Models

• click here (pw: hack) to download the trained models

3.1 ESC-50 dataset

Model	5Fold-CV
ResNeXt-GMP (baseline)	90.3
ResNext-GMP + RBlock (our best)	91.9

3.2 US8K dataset

Model	10fold-CV
ResNeXt-GMP (baseline)	84.7
ResNeXt-GMP + RBlock (our best)	85.9

3.3 DCASE2018 Task1A dataset

Model	Fold1
ResNeXt-GAP (baseline)	77.24
ResNeXt-GAP + RBlock (our best)	79.15

4. To reproduce the classification results in the paper

• Install python environment, see requirements.yaml
  • The code is based on torch1.6.0 and python3.6.10

# needs to change the prefix in requirements.yaml file to your own directory
conda env create -f requirements.yaml

• Run with config files, under cfgs/ folders

python main.py --cfg <cfg-file>

5. To reproduce visualizations in the paper, see vis_paper

6. To reproduce visualizations in the project page, see vis_proj

7. Find this project interesting? ☞ See more papers on ESC.