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?
解决方案
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)