Getting indices from numpy array and applying to 2nd lower dimensional array to create new array

Question

I have an array Gamma, Gamma.shape=(20,7,90,144), that represents the dimensions (t,z,y,x) and whose respective indices I denote by [l,k,j,i]. For each (t,y,x), I want to find the lowest value of k such that Gamma[l,k,j,i] > g_crit, where g_crit is some constant. Denote this set of lowest k values as k_low; I've found I can do this nicely with

k_low = np.argmax(Gamma > g_crit, axis=1)

There is another array levs, levs.shape=(7,), corresponding to the z dimension as well. I am ultimately trying to create the array levs_low, such that levs_low.shape=(20,90,144) and levs_low[l,j,i]=levs[k_low[l,j,i]]. I'm stuck on this step. Any ideas please? Thanks.

Answer 1

For your case it seems that np.take() is a good choice:

levs_low = np.take( levs, k_low )

OBS: take seems to give an optimized performance, check this question

Answer 2

This should do the trick:

levs_low=levs[k_low]

>>> Gamma=np.random.rand(20,7,90,144)
>>> k_low = np.argmax(Gamma > .3, axis=1)
>>> levs=np.random.rand(7)
>>> levs_low=levs[k_low]
>>> levs_low.shape
(20, 90, 144)

For a small example:

>>> g=np.random.randint(0,5,(4,4))
>>> g
array([[2, 0, 2, 2],
       [2, 0, 1, 0],
       [3, 3, 0, 3],
       [3, 0, 4, 4]])
>>> k=np.arange(5)*-1
>>> k
array([ 0, -1, -2, -3, -4])
>>> k[g]   #Uses indices of g to select values from k. Also same as np.take(k,g)
array([[-2,  0, -2, -2],
       [-2,  0, -1,  0],
       [-3, -3,  0, -3],
       [-3,  0, -4, -4]])

@Saullo Castro's answer is interesting. Strange that there would be a ~5x difference in timings between fancy indexing and np.take.

%timeit levs[k_low]
100 loops, best of 3: 2.3 ms per loop

%timeit np.take( levs, k_low )
1000 loops, best of 3: 439 us per loop

In [33]: np.all(levs[k_low]==np.take(levs,k_low))
Out[33]: True