CNN-for-Raw-Waveforms

This is my attempt to implement the "Very Deep Convolutional Neural Networks For Raw Waveforms" research paper published in 2016. My implementation has been done fully using PyTorch and PyTorchAudio.

The aim of this paper is to process raw audio wav files using only Convolution networks. The paper proposes five different models that range from 3 layers to 34 layers. The dataset used in this paper named "Urban Sound 8K" which contains short audio files (1-4 seconds) of 44,100 HZ stereo wav files.

Notebooks

In this repo, you can find notebooks for both GoogleColab and Kaggle that can be used directly without the need to clone the repo or anything. These files can be found in the notebooks directory.

Pre-requisites

To install the dependencies for this project, you need to run the following command:

pip install -r requirements.txt

Data

The dataset used in this paper is Urban Sound 8K. You can download the dataset simply by following these steps:

1. Create an account on Kaggle if you haven't already.

2. Go to your account and create a new token.

3. This will directly download a JSON file called kaggle.json. Create a .kaggle directory in your home directory and copy this file to ~/.kaggle. This can be done using the following commands:

   mkdir -p ~/.kaggle
   cp kaggle.json ~/.kaggle
   chmod 600 ~/.kaggle/kaggle.json
   

4. Now, get to the urbansound8k directory and run the bash script:

   cd urbansound8k
   bash download_dataset.sh 
   

Models

As we said earlier, the paper proposes five different models that range from 3 layers to 34 layers. These five models are defined in the models.py script like so:

m3 = CNN(channels = [[256], [256]],
         conv_kernels = [80, 3],
         conv_strides = [4, 1],
         conv_padding = [38, 1],
         pool_padding = [0, 0])

m5 = CNN(channels = [[128], [128], [256], [512]],
         conv_kernels = [80, 3, 3, 3],
         conv_strides = [4, 1, 1, 1],
         conv_padding = [38, 1, 1, 1],
         pool_padding = [0, 0, 0, 2])

m11 = CNN(channels = [[64], [64]*2, [128]*2, [256]*3, [512]*2],
          conv_kernels = [80, 3, 3, 3, 3],
          conv_strides = [4, 1, 1, 1, 1],
          conv_padding = [38, 1, 1, 1, 1],
          pool_padding = [0, 0, 0, 2])

m18 = CNN(channels = [[64], [64]*4, [128]*4, [256]*4, [512]*4],
          conv_kernels = [80, 3, 3, 3, 3],
          conv_strides = [4, 1, 1, 1, 1],
          conv_padding = [38, 1, 1, 1, 1],
          pool_padding = [0, 0, 0, 2])

m34_res = CNNRes(channels = [[48], [48]*3, [96]*4, [192]*6, [384]*3],
          conv_kernels = [80, 3, 3, 3, 3],
          conv_strides = [4, 1, 1, 1, 1],
          conv_padding = [38, 1, 1, 1, 1],
          pool_padding = [0, 0, 0, 2])

When you dig a little bit deeper in this script, you will see that there are two main base classes which are CNN and CNNRes. The first class is used with the first four models in the paper. The second class is the same as the first one but with residual blocks instead.

Experiments

In this experiments.py script you can replicate all the experiments found in the paper. You can try any experiment like the following one:

python experiments.py --model M34_res --n 5 --shuffle true --seed 42 --batchSize 128 --dataPath "urbansound8k"

Inference

TODO