week8 More ufunc overrides
This week was spent implementing the API for overriding ufuncs. After a lengthy discussion, I decided that it would be best to proceed with implementing things using a duck typing mechanism, similar to the one Python uses, instead of multimethods. Nathaniel Smith and Pauli Virtanen had a lot of great suggestions on the interface which has been totally changed since last week. __ufunc_override__
has become __numpy_ufunc__
and is now a method which takes arguments like:
def __numpy_ufunc__(self, ufunc, method, i, inputs, kwargs):
...
Where:
 ufunc is the ufunc object that was called.

method is a string indicating which Ufunc method was called
(one of
"__call__"
,"reduce"
,"reduceat"
,"accumulate"
,"outer"
,"inner"
).  i is the index of self in inputs.

inputs is a tuple of the input arguments to the
ufunc

kwargs is a dictionary containing the optional input arguments
of the ufunc. The
out
argument is always contained in kwargs, if given.
The implementation is best described by the NEP:

If one of the input arguments implements
__numpy_ufunc__
it is executed instead of the Ufunc. 
If more than one of the input arguments implements
__numpy_ufunc__
, they are tried in the following order: subclasses before superclasses, otherwise left to right. The first__numpy_ufunc__
method returning something else thanNotImplemented
determines the return value of the Ufunc. This is to accommodate Python's own MRO behavior. 
If all
__numpy_ufunc__
methods of the input arguments returnNotImplemented
, aTypeError
is raised. 
If a
__numpy_ufunc__
method raises an error, the error is propagated immediately.
Demo¶
There have been several models working in a limited fashion so far, but here is an example. I added this to scipy/sparse/base.py
So __numpy_ufunc__
is now a method on all sparse matrices.
import scipy.sparse as sp
import numpy as np
a = np.random.randint(5, size=(3,3))
b = np.random.randint(5, size=(3,3))
asp = sp.csr_matrix(a)
bsp = sp.csr_matrix(b)
np.multiply(a,b)
np.multiply(asp, bsp).todense()
np.multiply(asp,b)
np.multiply(a, bsp)
Comments
Comments powered by Disqus