CODEWITHSUNDEEP
pythonnumpycython
orange
orange

Reputation: 8090

Using numpy masking in cython

I'm trying to convert some of my Python code to Cython and experience some issues when I try to define a function as cdef.

Most of the issues boil down to masking not working as in Python. I wonder if this is a limitation of cdef (works fine if I leave it as def) or whether there's something I could do.

For instance this method

cdef func(double[:,:,:,:] arg1):
    mask = arg1 > 0
    ...

already fails with a compile error:

Error compiling Cython file:
------------------------------------------------------------ ...
    func (double[:,:,:,:] arg1):
        mask = arg1 > 0
                   ^
------------------------------------------------------------

cythonfile.pyx:43:20: Invalid types for '>' (double[:, :, :, :], long)

Upvotes: 2

Views: 548

Answers (2)

MSeifert
MSeifert

Reputation: 152775

The double[:,:,:,:] notation specifies that the argument will be "interpreted" as Typed Memoryview. These support a lot of operations but the vectorized comparisons are not supported.

However it's quite easy to interpret a memory-view as NumPy array inside the function:

import numpy as np

cdef func(double[:,:,:,:] arg1):
    arg1arr = np.asarray(arg1)
    mask = arg1arr > 0.

That doesn't even need a copy so it's basically "free" to do the np.asarray on the memory-view. That allows to combine the advantages of the memory-views with the vectorized operations that are possible on NumPy arrays.

However for vectorized operations you don't need Cython, you could do all the vectorized operations in a pure python function and only use Cython for the heavy-lifting that isn't possible with "normal NumPy" functions.

Upvotes: 1

Nils Werner
Nils Werner

Reputation: 36815

The docs consistently use np.ndarray[...] in function definitions so I would change your function signature to

cdef func(np.ndarray[np.float_t, ndim=4] arg1):

Also, you are comparing a float array with an long integer constant. Change it to

mask = arg1 > 0.

to compare float with float.

Upvotes: 1

Related Questions