Data preprocessing

Data Download and decompress

• First we need to download Coswara and coughbid datasets.

  mkdir datasets
  cd datasets
  git clone https://github.com/iiscleap/Coswara-Data.git # Coswara
  ./decompress.sh # Decompress the dataset. Only Coswara needs it. 
  git clone https://github.com/virufy/virufy-cdf-coughvid.git # coughvid from virufy

Data preparation

• After finishing data download, we will need to do data preparation.

  python data_preprocessing.py --dataset coswara ## preprocess coswara dataset, or
  python data_preprocessing.py --dataset coughvid ## preprocess coughvid dataset

• This will extract features of these wav files and generate a json file (stored at datasets/) according to the dataset being specified.

  # Coswara dataset
  {
      "iV3Db6t1T8b7c5HQY2TwxIhjbzD3": { # patient ID
          "feature_paths": [ # list of the patient's sound file
              "datasets/Coswara-Data/20200424/iV3Db6t1T8b7c5HQY2TwxIhjbzD3counting-normal.npy",
              "datasets/Coswara-Data/20200424/iV3Db6t1T8b7c5HQY2TwxIhjbzD3vowel-o.npy",
              "datasets/Coswara-Data/20200424/iV3Db6t1T8b7c5HQY2TwxIhjbzD3vowel-a.npy",
              "datasets/Coswara-Data/20200424/iV3Db6t1T8b7c5HQY2TwxIhjbzD3vowel-e.npy",
              "datasets/Coswara-Data/20200424/iV3Db6t1T8b7c5HQY2TwxIhjbzD3breathing-shallow.npy",
              "datasets/Coswara-Data/20200424/iV3Db6t1T8b7c5HQY2TwxIhjbzD3counting-fast.npy",
              "datasets/Coswara-Data/20200424/iV3Db6t1T8b7c5HQY2TwxIhjbzD3breathing-deep.npy",
              "datasets/Coswara-Data/20200424/iV3Db6t1T8b7c5HQY2TwxIhjbzD3cough-heavy.npy",
              "datasets/Coswara-Data/20200424/iV3Db6t1T8b7c5HQY2TwxIhjbzD3cough-shallow.npy"
          ],
          "pct_test_result": "untested" # patient's pcr test result
      },
      "AxuYWBN0jFVLINCBqIW5aZmGCdu1": { # another patient ID
          ...
      }
      ...
  }
  

• There are 417/1331/398 positive/negative/untested speakers in Coswara dataset.
• There are 3734/11926/3574 positive/negative/untested speakers in Coswara dataset.
• There are 588/7129/5400 positive/negative/untested speakers in Coughvid dataset.

Model training:

• To train the model, use the following command:

  python train.py --dataset <dataset> --semi <bool> --split_type <split_type>

• Note that we can only choose "random" as the split_type when training on Coughvid dataset.