make core algorithm a self.variable.

Currently in each specific deep learning framework’s frontend we import from core the ScatteringnD algorithm and call it.

It would be a nice change for power users if we instead imported the algorithm in the base frontend and defined a variable to contain the call i.e. self.scatteringalgo = scatteringnd. then in each specific frontend, the call to the algorithm would be given by self.scatteringalgo. This would allow power users to define a breadth first search algorithm #798 , or define their own scattering implementation where they can retrieve coefficients #445 (tho, these futures should be one day in our code base, please see #735).

[janden]

Are you saying we store a function as an attribute (a sort of callback)? I’m not convinced this is a good idea. The way we currently extend the classes is using inheritance. Adding another way of modifying the core behavior of the classes would just cause confusion, I think.

[lostanlen]

Hm, i don’t know. Then yes, a field for the algorithm is needed: (just like for out_type)

Maybe refactor scattering this way?

def scattering(self, input, sc_algorithm):
    self.check_input(input)
    phi, psi = self.load_filters()
    x = self.pad(input)
    S = sc_algorithm(x,  self.pad, self.unpad, self.backend ... )
    return self.format(S) 

in this way, we’d be exposing new utility functions (check_input, pad, format …) to potential future core modules (wavelets, phase harmonics …)