Custom Preprocessor and preprocessing pipeline construction from yaml file

[paulpeyret-biophonia]

In OPSO version : 0.10.2 it seems that designing a custom preprocessing/augmentation pipeline requires either to:

• Create a SpectrogramPreprocessorobject then change the pipeline actions manually. or
• Create a new class inheriting from BasePreceprocessorwith a custom pipeline initialized at instanciation.

In order to get the maximum flexibility over the actions and there parameters, Would it make sense to use a preprocessing_config.yaml file to create a custom preprocessing pipeline? In my mind, that could require to add a CustomPreprocessorclass to the package with a class method .from_yaml() which would be an alternative constructor that provide multiple ways to create different instances of this class. The yaml file would contain a dictionary with each action names (in the desired order of the pipeline) and there associated set of parameters.

Something similar have been done here : https://github.com/musikalkemist/praudio

I see some benefits to this:

• Simplicity and flexibility to experiment with augmentation and preprocessing techniques.
• Keep track of augmentation and preprocessing parameters inside the experiment tracking by saving a copy of this input yaml file as an artefact.

Would it make sense? Or is there already a way through flexible Preprocessor customization (I might have missed something)?

[sammlapp]

Hi, we’ve been planning to do this and I actually just started a prototype yesterday :) The challenge/limitstion is that it won’t be able to handle any custom functions defined by the users code. But it will easily save and reload the built in preprocessing functions.

[sammlapp]

initial implementation in this commit https://github.com/kitzeslab/opensoundscape/commit/b729842cea3b2dced46dd599c0bb30b66ec572fb and it will actually support custom functions/classes as long as the user provides them in the code and decorates them so that they are "registered"