在 scipy 稀疏 csr_matrix 中查找 n 个随机零元素

Question

I want to find n zero elements in a sparse matrix. I write the code below:

counter = 0
while counter < n:
    r = randint(0, W.shape[0]-1)
    c = randint(0, W.shape[1]-1)
    if W[r,c] == 0:
        result.append([r,c])
        counter += 1

Unfortunately, it is very slow. I want something more efficient. Is there any way to access zero elements from scipy sparse matrix quickly?

Answer 1

First, here's some code to create some sample data:

import numpy as np
rows, cols = 10,20   # Shape of W
nonzeros = 7         # How many nonzeros exist in W
zeros = 70           # How many zeros we want to randomly select

W = np.zeros((rows,cols), dtype=int)
nonzero_rows = np.random.randint(0, rows, size=(nonzeros,))
nonzero_cols = np.random.randint(0, cols, size=(nonzeros,))
W[nonzero_rows, nonzero_cols] = 20

The above code has created W as a sparse numpy array, having shape (10,20) , and having only 7 non-zero elements (out of the 200 elements). All the non-zero elements have a value 20 .

Here's the solution to pick zeros=70 zero elements from this sparse matrix:

argwhere_res = np.argwhere(np.logical_not(W))
zero_count = len(argwhere_res)
ids = np.random.choice(range(zero_count), size=(zeros,))
res = argwhere_res[ids]

res would now be a shape (70,2) array giving the locations of the 70 elements that we have randomly chosen from W .

Note that this does not involve any loops.

Answer 2

First make a list of all the 0's:

list_0s = [(j, i) for i in range(len(matrix[j])) for j in range len(matrix) if matrix[j,i] == 0]

Then get your random choices:

random_0s = random.choices(list_0s, k=n)

Testing this with:

 matrix = np.random.randint(1000, size=(1000,1000))
 n = 100

Takes 0.34 seconds.

Answer 3

import numpy as np
import scipy.sparse as sparse
import random
randint = random.randint

def orig(W, n):
    result = list()
    while len(result) < n:
        r = randint(0, W.shape[0]-1)
        c = randint(0, W.shape[1]-1)
        if W[r,c] == 0:
            result.append((r,c))
    return result

def alt(W, n):
    nrows, ncols = W.shape
    density = n / (nrows*ncols - W.count_nonzero())
    W = W.copy()
    W.data[:] = 1
    W2 = sparse.csr_matrix((nrows, ncols))
    while W2.count_nonzero() < n:
        W2 += sparse.random(nrows, ncols, density=density, format='csr')
        # remove nonzero values from W2 where W is 1
        W2 -= W2.multiply(W)
    W2 = W2.tocoo()    
    r = W2.row[:n]
    c = W2.col[:n]
    result = list(zip(r, c))
    return result

def alt_with_dupes(W, n):
    nrows, ncols = W.shape
    density = n / (nrows*ncols - W.count_nonzero())
    W = W.copy()
    W.data[:] = 1
    W2 = sparse.csr_matrix((nrows, ncols))
    while W2.data.sum() < n:
        tmp = sparse.random(nrows, ncols, density=density, format='csr')
        tmp.data[:] = 1
        W2 += tmp
        # remove nonzero values from W2 where W is 1
        W2 -= W2.multiply(W)
    W2 = W2.tocoo()
    num_repeats = W2.data.astype('int')
    r = np.repeat(W2.row, num_repeats)
    c = np.repeat(W2.col, num_repeats)
    idx = np.random.choice(len(r), n)
    result = list(zip(r[idx], c[idx]))
    return result

---

Here's a benchmark with:

W = sparse.random(1000, 50000, density=0.02, format='csr')
n = int((np.multiply(*W.shape) - W.nnz)*0.01)

In [194]: %timeit alt(W, n)
809 ms ± 261 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [195]: %timeit orig(W, n)
11.2 s ± 121 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [223]: %timeit alt_with_dupes(W, n)
986 ms ± 290 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

---

Note that alt returns a list with no duplicates. Both orig and alt_with_dupes may return duplicates.