I am trying to make a Python object (called a Stroke) that can be used as elements in NumPy arrays and operated on by NumPy universal functions.
I wish to do the following comparison operation:
x = Stroke(...)
y = Stroke(...)
z = Stroke(...)
arr = np.array([x, y, z], dtype=object)
output = np.greater(arr, 5)
# OR
output = np.greater(arr, Stroke(...))
For my case, I would also like to custom define how the universal function, np.greater, works on arrays of Stroke's. I tried making this custom definition using the __numpy_ufunc__ method as described here.
Unfortunately, when calling np.greater, the Stroke.__gt__ dunder method is used to evaluate np.greater rather than my custom defined function. How can I force NumPy to use of the custom defined np.greater function instead of the dunder method without removing the dunder method?
A relevant snippet of the Stroke code is seen below:
from polare._numpy_ufunc_overrides import HANDLED_FUNCTIONS
import numpy as np
class Stroke:
def __gt__(self, other):
return self._binary_operation(np.greater, other)
...
def __array_ufunc__(self, ufunc, method, *inputs, **kwargs):
if method == '__call__':
if ufunc in HANDLED_FUNCTIONS:
func = HANDLED_FUNCTIONS[ufunc]
return self._handle_functions(func, *inputs)
else:
# Handle the ufunc in a different way.
else:
return NotImplemented
where the HANDLED_FUNCTIONS dictionary is populated with the following code:
import numpy as np
HANDLED_FUNCTIONS = {}
def implements(np_function):
def decorator(func):
HANDLED_FUNCTIONS[np_function] = func
return func
return decorator
@implements(np.greater)
def greater(a: np.ndarray, b: np.ndarray) -> list:
...
