我正在尝试生成一个[600 x 600] numpy数组,其中包含10个类高斯数组的总和(每个数组都有一个随机生成的中心)。
我尝试使用高斯滤波器来生成单独的类高斯数组,然后将它们相加,但我确信有一种矢量化方法可以实现这一点。即使使用num_centers=10
它也很慢,我可能需要总计多达20个高斯。
这里有一个类似的问题,但它似乎没有一个好的或决定性的答案,我不知道如何将它应用于我的问题。 Sum of Gaussians into fast Numpy?
这是我尝试过的。
import numpy as np
from scipy.ndimage import gaussian_filter
import matplotlib.pyplot as plt
num_centers = 10 # number of Gaussians to sum
sigma = 100 # std. dev. of each Gaussian
result = np.zeros((600, 600))
for _ in range(num_centers):
# Pick a random coordinate within the array as the center
center = np.random.uniform(result.shape).astype(int)
# Make array with 1 at the center and 0 everywhere else
temp = np.zeros_like(result)
temp[center[0], center[1]] = 1
# Apply filter
gaussian = gaussian_filter(temp, sigma)
# Add to result
result += gaussian
# Result should look like a contour map with several hills
plt.imshow(result * 1000) # scale up to see the coloring
plt.show()
您可以消除循环,而是在每个中心创建一个值为1的数组,然后将gaussian_filter
应用于此数组。所有步骤都可以进行矢量化。
这是一个例子。我使sigma
更小,因此更容易区分中心,我将宽度增加到800(没有特别的原因:)。
import numpy as np
from scipy.ndimage import gaussian_filter
import matplotlib.pyplot as plt
num_centers = 10
sigma = 25
size = (600, 800)
impulses = np.zeros(size)
# rows and cols are the row and column indices of the centers
# of the gaussian peaks.
np.random.seed(123456)
rows, cols = np.unravel_index(np.random.choice(impulses.size, replace=False,
size=num_centers),
impulses.shape)
impulses[rows, cols] = 1
# or use this if you want duplicates to sum:
# np.add.at(impulses, (rows, cols), 1)
# Filter impulses to create the result.
result = gaussian_filter(impulses, sigma, mode='nearest')
plt.imshow(result)
plt.show()
这是情节:
您可以尝试使用mode
的gaussian_filter
参数来查看哪种模式最适合您。
我不确定如何以平行的方式处理随机高斯数组的创建,因为这是您的代码中花费最多的时间。 (我用timeit
来确定这一点)。这是可以预料的,因为gaussian_filter
是一个计算密集型函数。
但是,我确实通过在一系列高斯上使用np.sum()
来看到轻微的性能提升。这是因为调用np.sum()
一次比在循环中调用+=
更有效。
import numpy as np
from scipy.ndimage import gaussian_filter
import matplotlib.pyplot as plt
num_centers = 10 # number of Gaussians to sum
sigma = 100 # std. dev. of each Gaussian
holder = np.zeros((num_centers, 600, 600))
for _ in range(num_centers):
# Pick a random coordinate within the array as the center
center = np.random.uniform(result.shape).astype(int)
# Make array with 1 at the center and 0 everywhere else
temp = np.zeros((600, 600))
temp[center[0], center[1]] = 1
# Apply filter
holder[_] = gaussian_filter(temp, sigma)
result = np.sum(holder, axis=0)
# Result should look like a contour map with several hills
plt.imshow(result * 1000) # scale up to see the coloring
plt.show()