Interoperability with numpy

[cmey]

Would it be possible to support interoperability with numpy?

import numpy as np
import arrayfire as af

A = np.zeros((2,3), dtype=np.complex64)
Ad = af.array(A) # <-- load data from a numpy array
af.display(Ad)
print(Ad.elements(), Ad.type(), Ad.dims(), Ad.numdims())

Currently:

Using opencl backend
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-3-c8c7947108da> in <module>()
      3 
      4 A = np.zeros((2,3), dtype=np.complex64)
----> 5 Ad = af.array(A) # <-- load data from a numpy array
      6 af.display(Ad)
      7 print(Ad.elements(), Ad.type(), Ad.dims(), Ad.numdims())

/Users/cmey/Code/GitHub/arrayfire-python/arrayfire/array.pyc in __init__(self, src, dims)
    133                 type_char = tmp.typecode
    134             else:
--> 135                 raise TypeError("src is an object of unsupported class")
    136 
    137             elements = 1

TypeError: src is an object of unsupported class

[pavanky]

We are trying to ensure ArrayFire works out of the box on a python distribution. Numpy may not always be installed on the machine. That said, I pushed out 14d5bc6 that should add the ability to interop with numpy in the following manner:

import numpy as np
import arrayfire as af

def af_array(src, transpose=False):
    res = af.array(src.ctypes.data, src.shape, src.dtype.char)
    return res if not transpose else af.transpose(res)

a = np.random.random((5,5))
b = af_array(a)

print(a)
af.display(b)

[pavanky]

The transpose flag may be necessary because numpy is row major by default and ArrayFire is column major.

[cmey]

Works great, thanks!

[mklf]

This only work for square matrix with contigous data , for example:

import numpy as np
import arrayfire as af

def af_array(src, transpose=False):
    res = af.array(src.ctypes.data, src.shape, src.dtype.char)
    return res if not transpose else af.transpose(res)

a = np.random.random((5000,5000))[100:103,200:203]
b = af_array(a,transpose=True)

print(a)
af.display(b)

will print:

Using cuda backend
[[ 0.00902216  0.72194459  0.93649477]
 [ 0.74371952  0.94558585  0.49367796]
 [ 0.01659066  0.38938187  0.44101624]]
b
[3 3 1 1]
    0.0090     0.7219     0.9365 
    0.0837     0.7593     0.4400 
    0.1749     0.0054     0.7595 

the data of the two array are different.
a solution to this is convert the numpy.ndarray to fortran order and ensure it contigous:

import numpy as np
import arrayfire as af

def af_array(src):
    src_ = src if np.isfortran(src) else np.asfortranarray(src)
    res = af.array(src_.ctypes.data, src_.shape, src_.dtype.char)
    return res

[pavanky]

@FilipeMaia Can you comment here on how you are achieving this in afnumpy ?

[pavanky]

This is actually present in master as a constructor thanks to @FilipeMaia

The lay out changes are going to be difficult to implicitly handle inside arrayfire. I think it should be left to the user to change it to proper format before calling arrayfire functions.

[FilipeMaia]

The only way to ensure a conversion where the dimensions of the arrayfire array match dimensions of the numpy is like @mklf did in the last code snippet (N.B. this often means the data will have to be transposed as numpy uses C order by default).

In afnumpy I'm currently assuming the numpy arrays are always in C order and have the shape of the arrayfire array be the inverse of the numpy array, e.g.:

# Assuming src is in contiguous C order
A = af.array(src.ctypes.data, src.shape[::-1], src.dtype.char)

[pavanky]

I closed the issue hastily. arrayfire --> numpy is supported but the opposite is not.

[pavanky]

This is fixed in master. All you need to do now is call the following function:

af_arr = af.np_to_af_array(np_arr)

This will take care of all cases @mklf mentions.

[cmey]

Thank you!

[pavanky]

@cmey I have uploaded the latest version to be available via pip