在以下代码中,density=True
返回每个仓位处的概率密度函数。现在,如果必须计算P(x),我是否可以说hist
在显示概率?例如,如果第一个bin的平均值为0.5,我可以说在x = 0.5时,hist [0]的概率是多少?我必须使用使用P(x)的KL散度。
x = np.array([0,0,0,0,0,3,3,2,2,2,1,1,1,1,])
hist,bin_edges= np.histogram(x,bins=10,density=True)
答案 0 :(得分:1)
设置density=True
时,NumPy
返回概率密度函数(假设p
)。从理论上讲,p(0.5) = 0
是因为概率定义为PDF曲线下的面积。您可以阅读有关它的更多详细信息here。因此,如果要计算概率,则必须定义所需范围,并对该范围内的所有PDF值求和。
对于KL,我可以分享我的相互信息计算解决方案(基本上是KL):
def mutual_information(x, y, sigma=1):
bins = (256, 256)
# histogram
hist_xy = np.histogram2d(x, y, bins=bins)[0]
# smooth it out for better results
ndimage.gaussian_filter(hist_xy, sigma=sigma, mode='constant', output=hist_xy)
# compute marginals
hist_xy = hist_xy + EPS # prevent division with 0
hist_xy = hist_xy / np.sum(hist_xy)
hist_x = np.sum(hist_xy, axis=0)
hist_y = np.sum(hist_xy, axis=1)
# compute mi
mi = (np.sum(hist_xy * np.log(hist_xy)) - np.sum(hist_x * np.log(hist_x)) - np.sum(hist_y * np.log(hist_y)))
return mi
编辑: KL 可以这样计算(请注意,我没有对此进行测试!):
def kl(x, y, sigma=1):
# histogram
hist_xy = np.histogram2d(x, y, bins=bins)[0]
# smooth it out for better results
ndimage.gaussian_filter(hist_xy, sigma=sigma, mode='constant', output=hist_xy)
# compute marginals
hist_xy = hist_xy + EPS # prevent division with 0
hist_xy = hist_xy / np.sum(hist_xy)
hist_x = np.sum(hist_xy, axis=0)
hist_y = np.sum(hist_xy, axis=1)
kl = -np.sum(hist_x * np.log(hist_y / hist_x ))
return kl
此外,为了获得最佳结果,您应该使用一些启发式方法来计算sigma
,例如A rule-of-thumb bandwidth estimator。