Very Deep Convolutional Networks For Raw Waveforms

Keras (Tensorflow) implementation of the [paper].

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

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