动态Python数组切片

Question

I am facing a situation where I have a VERY large numpy.ndarray (really, it's an hdf5 dataset) that I need to find a subset of quickly because they entire array cannot be held in memory. However, I also do not want to iterate through such an array (even declaring the built-in numpy iterator throws a MemoryError ) because my script would take literally days to run.

As such, I'm faced with the situation of iterating through some dimensions of the array so that I can perform array-operations on pared down subsets of the full array. To do that, I need to be able to dynamically slice out a subset of the array. Dynamic slicing means constructing a tuple and passing it.

For example, instead of

my_array[0,0,0]

I might use

my_array[(0,0,0,)]

Here's the problem: if I want to slice out all values along a particular dimension/axis of the array manually, I could do something like

my_array[0,:,0]
> array([1, 4, 7])

However, I this does not work if I use a tuple:

my_array[(0,:,0,)]

where I'll get a SyntaxError .

How can I do this when I have to construct the slice dynamically to put something in the brackets of the array?

Answer 1

You could slice automaticaly using python's slice :

>>> a = np.random.rand(3, 4, 5)
>>> a[0, :, 0]
array([ 0.48054702,  0.88728858,  0.83225113,  0.12491976])
>>> a[(0, slice(None), 0)]
array([ 0.48054702,  0.88728858,  0.83225113,  0.12491976])

The slice method reads as slice(*start*, stop[, step]) . If only one argument is passed, then it is interpreted as slice(0, stop) .

In the example above : is translated to slice(0, end) which is equivalent to slice(None) .

Other slice examples:

:5 -> slice(5)
1:5 -> slice(1, 5)
1: -> slice(1, None)
1::2 -> slice(1, None, 2)

Answer 2

Okay, I finally found an answer just as someone else did.

Suppose I have array:

my_array[...]
>array(
  [[[ 1,  2,  3],
    [ 4,  5,  6],
    [ 7,  8,  9]],

   [[10, 11, 12],
    [13, 14, 15],
    [16, 17, 18]]])

I can use the slice object, which apparently is a thing:

sl1 = slice( None )
sl2 = slice( 1,2 )
sl3 = slice( None )
ad_array.matrix[(sl1, sl2, sl3)]
>array(
  [[[ 4,  5,  6]],

   [[13, 14, 15]]])