Implementation of InterMix: An Interference-based Data Augmentation And Regularization Technique For Automatic Deep Sound Classification by Ramit Sawhney, and Atula Tejaswi Neerkaje.

Python 3.8 & Chainer 7.7.0

Execute the following steps in the same environment:

python3 main.py --data data --dataset [DATASET] --mixup_type sound --netType envnetv2 --batchSize 32 --BC --eligible 1 2 3 4 --strongAugment

To run different variants of InterMix, perform ablation or tune hyperparameters, the following command-line arguments may be used:

  --dataset DATASET     dataset from ['esc10', 'esc50', 'urbansound8k']
  --mixup-type TYPE     sound (p-weighting) vs. normal
  --bc                  perform mixup
  --eligible L          eligible layer set 
  --strongAugment       perform scale and gain augmentation
  --batchSize BATCH_SIZE
                        batch size
  --nEpochs EPOCHS      number of epochs
  --LR RATE             learning rate
  --weightDecay WD      weight decay
  --momentum MOMENTUM   LR momentum
  --split SPLIT         choice of split

Dataset preparation for ESC-50, ESC-10, and UrbanSound8K

• FFmpeg should be installed.

• First of all, please make a directory to save datasets.

    mkdir [path]
  

python esc_gen.py [path]

• Following files will be generated.
  • [path]/esc50/wav16.npz # 16kHz, for EnvNet
  • [path]/esc50/wav44.npz # 44.1kHz, for EnvNet-v2
  • [path]/esc10/wav16.npz
  • [path]/esc10/wav44.npz

1. Download UrbanSound8K dataset from this page.

2. Move UrbanSound8K directory.

    mkdir -p [path]/urbansound8k
    mv UrbanSound8K [path]/urbansound8k/
   

3. Run the following command.

    python urbansound_gen.py [path]
   

• Following files will be generated.
  • [path]/urbansound8k/wav16.npz
  • [path]/urbansound8k/wav44.npz

If our work was helpful in your research, please kindly cite this work:

@inproceedings{Sawhney2022InterMix,
  title={InterMix: An Interference-based Data Augmentation And Regularization Technique For Automatic Deep Sound Classification},
  author={Sawhney, Ramit and 
          Neerkaje, Atula Tejaswi},
  booktitle={ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2022},
  organization={IEEE}
}

[1] Jindal, A., Ranganatha, N. E., Didolkar, A., Chowdhury, A. G., Jin, D., Sawhney, R., & Shah, R. R. (2020, January). SpeechMix-Augmenting Deep Sound Recognition Using Hidden Space Interpolations. In INTERSPEECH (pp. 861-865).

[2] Tokozume, Y., Ushiku, Y., & Harada, T. (2018, February). Learning from Between-class Examples for Deep Sound Recognition. In International Conference on Learning Representations.