当我尝试对这个numpy数组求和时，为什么Python会崩溃？

Question

I'm working on Ubuntu 14.04 with Python 3.4 (Numpy 1.9.2 and PIL.Image 1.1.7). Here's what I do:

>>> from PIL import Image
>>> import numpy as np

>>> img = Image.open("./tifs/18015.pdf_001.tif")
>>> arr = np.asarray(img)
>>> np.shape(arr)
(5847, 4133)

>>> arr.dtype
dtype('bool')

# all of the following four cases where I incrementally increase
# the number of rows to 700 are done instantly
>>> v = arr[1:100,1:100].sum(axis=0)
>>> v = arr[1:500,1:100].sum(axis=0)
>>> v = arr[1:600,1:100].sum(axis=0)
>>> v = arr[1:700,1:100].sum(axis=0)

# but suddenly this line makes Python crash
>>> v = arr[1:800,1:100].sum(axis=0)

fish: Job 1, “python3” terminated by signal SIGSEGV (Address boundary error)

Seems to me like Python runs out of memory all of a sudden. If that is the case - how can I allocate more memory to Python? As I can see from htop my 32GB memory capacity is not even remotely depleated.

You may download the TIFF image here .

---

If I create an empty boolean array, set the pixels explicitely and then apply the summation - then it works:

>>> arr = np.empty((h,w), dtype=bool)
>>> arr.setflags(write=True)

>>> for r in range(h):
>>>     for c in range(w):
>>>         arr.itemset((r,c), img.getpixel((c,r)))

>>> v=arr.sum(axis=0)

>>> v.mean()
5726.8618436970719

>>> arr.shape
(5847, 4133)

But this "workaround" is not very satisfactory as copying every pixel takes way too long - maybe there is a faster method?

Answer 1

I can reproduce your segfault using numpy v1.8.2/PIL v1.1.7 installed from the Ubuntu repositories.

• If I install numpy 1.8.2 in a virtualenv using pip (still using PIL v1.7.1 from the Ubuntu repos) then I no longer see the segfault.

• If I do the opposite (installing PIL v1.1.7 using pip, and using numpy v1.8.2 from the Ubuntu repos), I still get the segfault.

This leads me to believe that it's caused by an old bug in numpy. I haven't been able to find a good candidate in numpy's issue tracker, but I suspect that updating numpy (eg from the current source or via pip) would probably resolve the issue.

One workaround is to convert the image mode to "P" (unsigned 8-bit ints) before creating the array, then converting it back to boolean:

arr2 = np.asarray(img.convert("P")).astype(np.bool)
v = arr2[1:800,1:100].sum(axis=0)