Open damellis opened 8 years ago
Yes, I've been meaning to improve this as there are many cases when you want to mix both. The main logic for the original setup is that a pre-processing module is something that keeps the dimensionality the same (so the number of inputs/outputs will always match), whereas a feature extraction module could have varying dimensionality (so a 3 dimensional input vector could result in a 10 dimensional feature vector for instance)....but I agree it makes more sense to have a more generic ProcessingNode (or something) which can be chained together in the pipeline and consist of either pre-processing or feature extraction algorithms.
I'll look at reconfiguring this over the weekend.
There are times when it might be useful to mix these, e.g. taking the average over time of the result of an FFTFeatures module. Is there a reason that pre-processing and feature extraction need to be distinct? Aren't they both just functions from a vector to another vector?