This repo contains code for our paper: Urban Sound Classification : striving towards a fair comparison.

It provides a fair comparison by using the same input representation, metrics and optimizer to assess performances. We preserve data augmentation used by the original papers. We hope this framework could help evaluate new architectures in this field.

Python version recquired : 3.6 (higher work). We recommand first to create a new environment in conda/virtualenv then to activate it.

Pip install

pip install -r requirements.txt

Conda install

conda env create -f environment.yml

Manual install

pip install numpy scikit-learn pandas tqdm albumentations librosa tensorboard torch torchvision oyaml pytorch-lightning numba==0.49
pip install torchaudio -f https://download.pytorch.org/whl/torch_stable.html

You should edit PATH in config.py to match the directory in which everything will be stored.

Use the following to download the dataset and precompute inputs.

WARNING : It requires about 30Go of free space.

python data_prep.py --download --mel

If you want to manualy download and decompress files, you have to put everything in the audio directory (cf config.py). Then you have to use the above command without the --download.

To use relabeling for TALNet modified, copy paste the best2.csv into the SONYC-UST folder.

To use transfer learning, download the pretrained models and copy-paste them into the SONYC-UST/model folder as shown below.

• the pretrained TALNet on Audioset can be found here.
• the pretrained CNN10 on Audioset can be found here.

Your data folder should look like:

.
└── data                        # Given by PATH in config.py
    ├── SONYC-UST                   
         ├── audio              # Put all audio in it 
         ├── melTALNet          # Contain all computed mel-spectrograms 
         ├── best2.csv 
         │  
         ...
         └── model              # Put TALNet and CNN10 weights here
    ├── ESC-50 
         ├── audio             
         ├── melTALNet 
         ├── meta
         │  
         ...
    ├── UrbanSound8k
         ├── audio             
         ├── melTALNet  
         ├── metadata
         │  
         ...

The code should work on both CPU and GPU. If you want to train everything on CPU, remove gpus=1 in the corresponding model_<model_name>_training.py file. The scripts used for the comparison are the test_model_<model_name>.sh file. To run one test, just execute the following command :

sh test_model_<model_name>.sh

WARNING : CNN10TL and TALNETv3 require the pretrained models on Audioset before running them.

Training results are stored in the data folder:

.
└── data                        # Given by PATH in config.py
    ├── SONYC-UST                   
    ├── ESC-50 
    ├── UrbanSound8k
    └── summaries
         ├── CNN10AllDatasets
         │  
         ...
         └── TFNetAllDatasets

@article{ArnaultAnalysis2020,
  title={Urban Sound Classification : striving towards a fair comparison},
  author={Arnault, Augustin and Hanssens Baptiste and Riche, Nicolas},
  journal={arXiv preprint arXiv:2010.11805},
  year={2020}
}