rerrayne/bc_learning_pytorch

BC learning for sounds PyTorch Port

This is the port of Between-class Examples for Deep Sound Recognition to PyTorch. Dataset generation was taken from the original repo.

Implementation of Learning from Between-class Examples for Deep Sound Recognition by Yuji Tokozume, Yoshitaka Ushiku, and Tatsuya Harada (ICLR 2018).

This also contains training of EnvNet: Learning Environmental Sounds with End-to-end Convolutional Neural Network (Yuji Tokozume and Tatsuya Harada, ICASSP 2017).¹

Contents

• Between-class (BC) learning
  • We generate between-class examples by mixing two training examples belonging to different classes with a random ratio.
  • We then input the mixed data to the model and train the model to output the mixing ratio.
• Training of EnvNet on ESC-50, ESC-10 [1], and UrbanSound8K [2] datasets

Setup

• Install PyTorch.
• Prepare datasets following this page.

Training

• Template:

    python main.py --dataset [esc50, esc10, or urbansound8k] --netType [envnet or envnetv2] --data path/to/dataset/directory/ (--BC) (--strongAugment)
  

• Recipes:

  • Standard learning of EnvNet on ESC-50 (around 29% error²):

      python main.py --dataset esc50 --netType envnet --data path/to/dataset/directory/
    

• Notes:

  • Please check opts.py for other command line arguments.

See also

Between-class Learning for Image Clasification (github)

Reference

[1] Karol J Piczak. Esc: Dataset for environmental sound classification. In ACM Multimedia, 2015.

[2] Justin Salamon, Christopher Jacoby, and Juan Pablo Bello. A dataset and taxonomy for urban sound research. In ACM Multimedia, 2014.

bc_learning_pytorch's Issues

Not Able to replicate results on ESC-50

I have been running the experiments on this repo and I have not been able to replicate the results of ESC50(only BC, EnvNetv2, without strongAugment) mention in the paper. Any assistance would be appreciated.

Input of KLDivLoss

I believed that the input of KLDivLoss should be log-probability.
https://pytorch.org/docs/stable/nn.html?highlight=kldivloss#torch.nn.KLDivLoss

In the code, the output of the model is the softmax.
I think it is suppose to be LogSoftmax() at the last layer.