Matlab与Python二维卷积性能比较

我喜欢在Matlab中创建算法的原型，但我需要将它们放在一个运行大量Python代码的服务器上。因此，我很快将代码转换为Python并比较了两者。Matlab实现的运行速度快了约1000倍（从计时函数调用-无评测）。有人能马上知道为什么Python的性能如此之慢吗

Matlab

---

好的，多亏@Yves Daust的建议，我的问题解决了

filter

scipy.ndimage.filters.gaussian_filter

利用内核的可分性，并将运行时间减少到matlab实现的单个数量级以内

import numpy as np
from scipy.ndimage.filters import gaussian_filter as gaussian


# Test data parameters
w = 800
h = 1200
npts = 250

# generate data
xvals = np.random.randint(w, size=npts)
yvals = np.random.randint(h, size=npts)

# Heatmap parameters
gaussianSize = 250
nbreaks = 25

# Preliminary function definitions
def populateMat(w, h, xvals, yvals):
    container = np.zeros((w,h))

    for idx in range(0,xvals.size):
        x = xvals[idx]
        y = yvals[idx]
        container[x,y] += 1

    return container


# Create the data matrix
dmat = populateMat(w,h,xvals,yvals)

# Convolve
dmat2 = gaussian(dmat, gaussianSize/7)

# Scaling etc
dmat2 = dmat2 / dmat2.max()
dmat2 = np.round(nbreaks*dmat2)/nbreaks

# Show
imshow(dmat2)

---

使用numpy进行卷积。请看一看，您是说

scipy.signal.convolve2d

中的实现是罪魁祸首吗？我读了这篇文章，似乎说使用numpy的唯一原因是为了避免scipy依赖。我没有那个限制。尝试ndimage.CONVALVE，但这会引发记忆错误。我猜matlab使用了一个优化的高斯滤波器算法，利用了滤波器的可分性，可能还使用了递归近似。尝试在Python下使用可分性。这很简单：定义两个一维高斯分布，水平和垂直，并按顺序应用它们。

import numpy as np
from scipy.signal import convolve2d as conv2

# Test data parameters
w = 800
h = 1200
npts = 250

# generate data
xvals = np.random.randint(w, size=npts)
yvals = np.random.randint(h, size=npts)

# Heatmap parameters
gaussianSize = 250
nbreaks = 25

# Preliminary function definitions
def populateMat(w, h, xvals, yvals):
    container = np.zeros((w,h))

    for idx in range(0,xvals.size):
        x = xvals[idx]
        y = yvals[idx]
        container[x,y] += 1

    return container

def makeGaussian(size, fwhm):

    x = np.arange(0, size, 1, float)
    y = x[:,np.newaxis]

    x0 = y0 = size // 2

    return np.exp(-4*np.log(2) * ((x-x0)**2 + (y-y0)**2) / fwhm**2)


# Create the data matrix
dmat = populateMat(w,h,xvals,yvals)
h = makeGaussian(gaussianSize, fwhm=gaussianSize/2)

# Convolve
dmat2 = conv2(dmat, h, mode='same')

# Scaling etc
dmat2 = dmat2 / dmat2.max()
dmat2 = np.round(nbreaks*dmat2)/nbreaks
# Show
imshow(dmat2)

import numpy as np
from scipy.ndimage.filters import gaussian_filter as gaussian


# Test data parameters
w = 800
h = 1200
npts = 250

# generate data
xvals = np.random.randint(w, size=npts)
yvals = np.random.randint(h, size=npts)

# Heatmap parameters
gaussianSize = 250
nbreaks = 25

# Preliminary function definitions
def populateMat(w, h, xvals, yvals):
    container = np.zeros((w,h))

    for idx in range(0,xvals.size):
        x = xvals[idx]
        y = yvals[idx]
        container[x,y] += 1

    return container


# Create the data matrix
dmat = populateMat(w,h,xvals,yvals)

# Convolve
dmat2 = gaussian(dmat, gaussianSize/7)

# Scaling etc
dmat2 = dmat2 / dmat2.max()
dmat2 = np.round(nbreaks*dmat2)/nbreaks

# Show
imshow(dmat2)