23

I have a 3-D NumPy array, e.g.

a = np.random.random((2,3,5))

I would like to transpose the last two axes, i.e.

b = a.transpose(0,2,1)

However, I do not want a view with twiddled strides! I want to actually copy the array and reorder it in memory. What is the best way to achieve this?

bdforbes
  • 1,486
  • 1
  • 13
  • 31

2 Answers2

26

The copy() method will reorder to C-contiguous order by default:

b = a.transpose(0,2,1).copy()

Be careful: the copy() function has a different default behavior. With the function, you must explicitly specify the order to ensure a C-contiguous copy:

b = np.copy(a.transpose(0,2,1), order='C')

(Note that the docstring for the function says that the ndarray method is the preferred method for creating an array copy.)

Warren Weckesser
  • 110,654
  • 19
  • 194
  • 214
  • 1
    Thanks. The solution I had _just_ found before you posted yours was `b = np.zeros((2,5,3)); b[:] = a.transpose(0,2,1)`, but yours is much cleaner. – bdforbes Nov 07 '13 at 01:32
4

Under the hood, the stride of b is different than a.

prefer to use ascontiguousarray, which will copy the memory when it's needed. Whereas copy will always copy memory.

Izana
  • 2,537
  • 27
  • 33