Closed Thunfischpirat closed 1 year ago
Since spectrograms are downsampled before saving the corresponding numpy array, the number of columns of the saved array doesn't correspond anymore to the argument num_mels
of create_features()
. Since fixing this would greatly increase the cognitive complexity of the code, I decided to leave the code as it is. The user can recreate features by setting the argument lazy_mode
to False
.
Function
create_features()
inmodel_baseline/mel_spectrogram.py
doesn't check whether the parametersnum_features
ornum_mels
have been changed when checking whether it should load features and labels from the data folder or generate them.What might happen: The user changes
num_mels
ornum_frames
between two runs, but since features and labels have already been saved in the first run, this change doesn't do anything since create_features() tries to load features and labels by default if they already exist.