Specify the metadata (csv) location in data_df, with the location of the raw wave files in data_dir. The desired output for the features is set by dir_out. Model objects will be saved to outputs/models/pytorch/ or outputs/models/keras/. By default, metadata is stored in /data/metadata/, audio files are found in /data/audio/, and features in /outputs/features.
A more detailed discussion on feature transformations is given in Appendix B of the supplement in /docs/.
Features with default configuration from the VGGish GitHub intended for use with VGGish: 64 log-mel spectrogram coefficients using 96 feature frames of 10 ms duration. To compare feature sets fairly, predictions are aggregated over neighbouring windows to create outputs over 1.92 second windows as used in Feat. B. The behaviour of these features can be tweaked to user preference within lib/vggish/vggish_input.py, lib/feat_vggish.py and their dependencies.
The feature extraction uses log-mel features with librosa, configurable in config.py with the sample rate rate, to which data is re-sampled on loading, a window size win_size which determines the size of a training window (in number of feature windows), step_size, which determines the step size taken by the window, NFFT, and n_hop, which are parameters for the core STFT transform upon which log-mel feature extraction is based. Finally, n_feat determines the number of mel-frequency bands.
In librosa, we can calculate the value of win_size to achieve a user's desired duration for a label as follows:
win_size = duration / frame_duration, where frame_duration = n_hop/rate. Librosa uses a default hop_length of NFFT/4.
The default values in config.py are optimised for rate = 8000 with win_size = 30, NFFT = 2048, n_hop = default, to achieve a label duration of 30 * 2048/4 * (1/8000) = 1.92 (s).
tau,lengthscale, are parameters used for L2 weight regularization supplied in runKeras.py. dropout controls the dropout rate, batch_size controls the batch size, and epochs, set the number of epochs to train.
Settings to change the learning rate, lr, the number of maximum overrun steps for a particular training criteria max_overrun, the number of epochs, and the batch_size. Whether to use pre-trained weights for the models is configured by the boolean pretrained. n_classes sets the number of classes to use with the CrossEntropyLoss() function in PyTorch.