在 scipy 稀疏 csr_matrix 中找到 n 个随机零元素

Find n random zero element in a scipy sparse csr_matrix

我想在稀疏矩阵中找到 n 个零元素。我写下面的代码:

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

不幸的是,它很慢。我想要更有效率的东西。有什么方法可以快速访问 scipy 稀疏矩阵中的零元素?

首先列出所有 0:

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

然后随机选择:

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

用以下方法测试:

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

需要 0.34 秒。

首先,这里有一些代码可以创建一些示例数据:

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

以上代码创建了 W 作为稀疏 numpy 数组,形状为 (10,20),并且只有 7 个非零元素(在 200元素)。所有非零元素都有一个值 20.

下面是从这个稀疏矩阵中选取 zeros=70 个零元素的解决方案:

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 现在是一个形状 (70,2) 数组,给出我们从 W.

中随机选择的 70 元素的位置

请注意,这不涉及任何循环。

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

这是一个基准:

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)

请注意,alt return 是一个没有重复项的列表。 origalt_with_dupes 都可以 return 重复。