I have 1D array of N elements. I need to create a PxR view (where PxR<N) on this array according to strides that re not uniform but rather specified in a aux list. Observe that aux list has exactly P elements. R is a constant.

Example

INPUTS:
L=[10,14,4,12,3,85,43,22,11,7,10]
R=3
aux=[2,5,7]

output= [[4,12,3],[85,43,22],[22,11,7]]

Comprehensive lists is an option.

My code:

output = [[L[a p] for p in range(R)] for a in aux]

However, since L,R and aux are big in my case, I was wandering if there is a numpy function that do the job in a more efficient way.

I was exploring numpy.lib.stride_tricks.as_strided as an option. This would be great since it also enables me reduce memory occupation. But it seems that as_stride only works for constant strides....

Any suggestion to solve my problem?

CodePudding user response：

With lists, that kind of list comprehension is the only option.

But if we start with arrays:

In [298]: L=np.array([10,14,4,12,3,85,43,22,11,7,10])
     ...: R=3
     ...: aux=np.array([2,5,7])

We can make an indexing array with:

In [299]: idx = aux[:,None] np.arange(R)
In [300]: idx
Out[300]: 
array([[2, 3, 4],
       [5, 6, 7],
       [7, 8, 9]])
In [301]: L[idx]
Out[301]: 
array([[ 4, 12,  3],
       [85, 43, 22],
       [22, 11,  7]])

edit

Using a safer, and easier version as `as_strided:

In [323]: np.lib.stride_tricks.sliding_window_view(L,(3,))
Out[323]: 
array([[10, 14,  4],
       [14,  4, 12],
       [ 4, 12,  3],
       [12,  3, 85],
       [ 3, 85, 43],
       [85, 43, 22],
       [43, 22, 11],
       [22, 11,  7],
       [11,  7, 10]])
In [324]: np.lib.stride_tricks.sliding_window_view(L,(3,))[[2,5,7]]
Out[324]: 
array([[ 4, 12,  3],
       [85, 43, 22],
       [22, 11,  7]])

While [323] is a view, the subset (and [301]) is a copy.