numpy.getbuffer and numpy.frombuffer

Question

I am trying to wrap my head around the numpy implementation of PEP3118. How exactly does buffer access work in numpy.

>>> p = numpy.getbuffer(numpy.arange(10))
>>> p
<read-write buffer for 0x1003e5b10, size -1, offset 0 at 0x1016ab4b0>
>>> numpy.frombuffer(p)
array([  0.00000000e+000,   4.94065646e-324,   9.88131292e-324,
     1.48219694e-323,   1.97626258e-323,   2.47032823e-323,
     2.96439388e-323,   3.45845952e-323,   3.95252517e-323,
     4.44659081e-323])

So I am getting unexpected returns. I would expect to see an array with 10 elements from 0-9. I can get into the array and read/write though.

>>> j = numpy.frombuffer(p)
>>> j
array([  0.00000000e+000,   4.94065646e-324,   9.88131292e-324,
     1.48219694e-323,   1.97626258e-323,   2.47032823e-323,
     2.96439388e-323,   3.45845952e-323,   3.95252517e-323,
     4.44659081e-323])
>>> j += 1
>>> j
array([ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.])

So it looks like the buffer is initializing to all zeros, which I can then write to. The functionality that I am expecting is to be able to build the array (with arange or asarray) directly to the buffer with getbuffer. Is that not possible?

Answer 1

You have a simple dtype problem. The buffer you create with

np.getbuffer(np.arange(10))

has a dtype=int, because np.arange uses dtype=int by default.

Then, when you try to read your buffer with

np.frombuffer(p)

you're in fact using the dtype=float default of np.frombuffer. Instead, use

np.frombuffer(p, dtype=int)

et voilà, you get

array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])