Question

I would like to perform checks on the contents of a numpy array every time it is set. Can python properties be used for this? My approach:

import numpy as np

class Obj(): 
    def __init__(self):                                                                                 
        self._np_arr = None                                                                             

    @property                                                                                           
    def np_arr(self):                                                                                   
        if self._np_arr is None:                                                                        
            self._np_arr = np.ones(10)                                                                  
        return self._np_arr                                                                             

    @np_arr.setter
    def np_arr(self, value):
        if np.sum(value)>10:
            raise ValueError('Error message')                                                           
        self._np_arr = value

if __name__ == '__main__':
    o = Obj()                                                                                           
    print o.np_arr
    o.np_arr = np.zeros(10) # ok                                                                        
    o.np_arr = np.ones(10)*2 # not ok                                                                   
    print o.np_arr       

The getter is entered when the object is still None. Once np_arr is a numpy array the getters and setters no longer work.

What am I doing wrong?

Answer 1

First of all. This is difficult (up to, I believe impossible) to do right. The error here is quite simple, it must be a new style class. So just replace class Obj() with class Obj(object).

This may not solve your problem however, since the user can still assign it in place, ie:

o.np_arr[:] = 2

You could circumvent that somewhat by using _np_arr.setflags(write=False), but then the user cannot use such operations at all.

Edit:
you could also subclass ndarray, and define your own __array_finalize__ to circumvent that probably. However, even then the user could do np.asarray(o.np_arr) and modify that in place…

I guess you could provide your own methods then “simulate” an array by wroking internally on a writeable one. But I doubt there is a 100% foolproof or at least elegent method.