Store filepaths for spectograms and frames in a dict, and keep a list of the available timestamps.
If (for now theoretical scenario) either frame or spectogram does not exist, the dataloader will return an all zeroes tensor in those cases.
TODO: keep it all zeroes after applying the model, to signal that this feature vector is meaningless. This is not trivial, and since it is a theoretical scenario for now, I placed this out of scope.
Have done some manual checking, no unit tests (yet).
Also removed spectogram dimensionality from general config: this is obtained from model config.
Store filepaths for spectograms and frames in a dict, and keep a list of the available timestamps. If (for now theoretical scenario) either frame or spectogram does not exist, the dataloader will return an all zeroes tensor in those cases. TODO: keep it all zeroes after applying the model, to signal that this feature vector is meaningless. This is not trivial, and since it is a theoretical scenario for now, I placed this out of scope.
Have done some manual checking, no unit tests (yet).
Also removed spectogram dimensionality from general config: this is obtained from model config.