https://github.com/philipperemy/very-deep-convnets-raw-waveforms

Tensorflow - Very Deep Convolutional Neural Networks For Raw Waveforms - https://arxiv.org/pdf/1610.00087.pdf
https://github.com/philipperemy/very-deep-convnets-raw-waveforms

audio-analysis deep-learning deep-neural-networks keras keras-models keras-neural-networks keras-tensorflow machine-learning residual-networks tensorflow tensorflow-1-0 very-deep-cnn waveform

Last synced: 2 months ago
JSON representation

Tensorflow - Very Deep Convolutional Neural Networks For Raw Waveforms - https://arxiv.org/pdf/1610.00087.pdf

• Host: GitHub
• URL: https://github.com/philipperemy/very-deep-convnets-raw-waveforms
• Owner: philipperemy
• License: apache-2.0
• Created: 2017-05-14T08:37:14.000Z (over 7 years ago)
• Default Branch: master
• Last Pushed: 2021-03-19T02:25:37.000Z (almost 4 years ago)
• Last Synced: 2024-10-29T23:42:17.288Z (3 months ago)
• Topics: audio-analysis, deep-learning, deep-neural-networks, keras, keras-models, keras-neural-networks, keras-tensorflow, machine-learning, residual-networks, tensorflow, tensorflow-1-0, very-deep-cnn, waveform
• Language: Python
• Homepage:
• Size: 69.8 MB
• Stars: 73
• Watchers: 11
• Forks: 14
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Very Deep Convolutional Networks For Raw Waveforms

Keras (Tensorflow) implementation of the [[paper](https://arxiv.org/abs/1610.00087)].

## Notes:

- Going really deep does not seem to help much on this dataset. We clearly overfit very easily. Adding more regularization might help. I haven't tried to use the FC layers (though it has been implemented).

- We use the `fold10` folder for the **testing set** and the remaining for the **training set**.

- Models implemented:

```

[x] M3

[x] M5

[x] M11

[x] M18

[x] M34 (ResNet)

```

## How to re-run the experiments?

Dataset can be downloaded here: http://urbansounddataset.weebly.com/urbansound8k.html

```bash

git clone https://github.com/philipperemy/very-deep-convnets-raw-waveforms.git

cd very-deep-convnets-raw-waveforms

sudo pip3 install -r requirements.txt

./run_all.sh # will run M3, M5, M11, M18 and M34

```



  M3 model - best accuracy: 0.673, trainable params = 221,194


  







  M5 model - best accuracy: 0.743, trainable params = 559,114


  







  M11 model - best accuracy: 0.752, trainable params = 1,786,442


  







  M18 model - best accuracy: 0.710, trainable params = 3,683,786


  







  M34 model - best accuracy: 0.725, trainable params = 3,984,154