How to zero out all values of matrix (2D array) except top n values using numpy?

Question

Suppose i have a 2D numpy array. Given n, i wish to nulify all elements in the matrix except the top n.

I've tried idx = (-y_pred).argsort(axis=-1)[:, :n] to determine what are the indices of the largest n values, but idx shape is [H,W,n], and i don't understand why.

I've tried -

sorted_list = sorted(y_pred, key=lambda x: x[0], reverse=True)
top_ten = sorted_list[:10]

But it didn't really return top 10 indices.

Is there an efficient way to find top n indices and zero the rest?

EDIT input is a NxM matrix of values, and output is the same matrix of size NxM, such that all values are 0 except in indices that correspond to top 10 values

@Divakar the approach I tried creates a new array. Is there a way to get these indices of `top_n` elements in a *view* and then zero out the rest in the input array? — kmario23, May 12 '19 at 14:06
@kmario23 A view is not possible. There isn't a specific striding pattern there. — Divakar, May 12 '19 at 14:08

kmario23 · Answer 1 · 2019-05-12T13:29:06.980

Here's one approach using numpy.argpartition() based on the idea of How do I get indices of N maximum values in a NumPy array?

# sample input to work with
In [62]: arr = np.random.randint(0, 30, 36).reshape(6, 6)

In [63]: arr
Out[63]: 
array([[ 8, 25, 12, 26, 21, 29],
       [24, 22,  7, 14, 23, 13],
       [ 1, 22, 18, 20, 10, 19],
       [26, 10, 27, 19,  6, 28],
       [17, 28,  9, 13, 11, 12],
       [18, 25, 15, 29, 25, 25]])


# initialize an array filled with zeros
In [59]: nullified_arr = np.zeros_like(arr)
In [64]: top_n = 10

# get top_n indices of `arr`
In [57]: top_n_idxs = np.argpartition(arr.reshape(-1), -top_n)[-top_n:]

# copy `top_n` values to output array
In [60]: nullified_arr.reshape(-1)[top_n_idxs] = arr.reshape(-1)[top_n_idxs]


In [71]: nullified_arr
Out[71]: 
array([[ 0, 25,  0, 26,  0, 29],
       [ 0,  0,  0,  0,  0,  0],
       [ 0,  0,  0,  0,  0,  0],
       [26,  0, 27,  0,  0, 28],
       [ 0, 28,  0,  0,  0,  0],
       [ 0,  0,  0, 29, 25, 25]])

Yuya Takashina · Answer 2 · 2019-05-12T13:31:21.370

1

The following code will nullify an NxM matrix X.

threshold = np.sort(X.ravel())[-n]  # get the nth largest value
idx = X < threshold
X[idx] = 0

Note: this method can return a matrix which has more than n nonzero elements when there are duplicated values.

edited May 12 '19 at 13:31

answered May 12 '19 at 13:25

Yuya Takashina

592
6
13

How to zero out all values of matrix (2D array) except top n values using numpy?

2 Answers2