我正在尝试比较sklearn.neighbors.KernelDensity与scipy.stats.gaussian_kde对二维数组的效果。
从this article我看到每个函数中带宽(bw)的处理方式不同。本文给出了在scipy
中设置正确bw的方法,因此它将等同于sklearn
中使用的bw。基本上它将bw除以样本标准差。结果如下:
# For sklearn
bw = 0.15
# For scipy
bw = 0.15/x.std(ddof=1)
其中x
是我用来获取KDE的示例数组。这在1D中工作得很好,但我无法在2D中工作。
这是我得到的MWE
:
import numpy as np
from scipy import stats
from sklearn.neighbors import KernelDensity
# Generate random data.
n = 1000
m1, m2 = np.random.normal(0.2, 0.2, size=n), np.random.normal(0.2, 0.2, size=n)
# Define limits.
xmin, xmax = min(m1), max(m1)
ymin, ymax = min(m2), max(m2)
# Format data.
x, y = np.mgrid[xmin:xmax:100j, ymin:ymax:100j]
positions = np.vstack([x.ravel(), y.ravel()])
values = np.vstack([m1, m2])
# Define some point to evaluate the KDEs.
x1, y1 = 0.5, 0.5
# -------------------------------------------------------
# Perform a kernel density estimate on the data using scipy.
kernel = stats.gaussian_kde(values, bw_method=0.15/np.asarray(values).std(ddof=1))
# Get KDE value for the point.
iso1 = kernel((x1,y1))
print 'iso1 = ', iso[0]
# -------------------------------------------------------
# Perform a kernel density estimate on the data using sklearn.
kernel_sk = KernelDensity(kernel='gaussian', bandwidth=0.15).fit(zip(*values))
# Get KDE value for the point.
iso2 = kernel_sk.score_samples([[x1, y1]])
print 'iso2 = ', np.exp(iso2[0])
(iso2
表示为指数,因为sklearn
返回日志值)
我得到的iso1
和iso2
的结果是不同的,我失去了如何影响带宽(在任一函数中)以使它们相等(应该如此)。 / p>
添加
我在sklearn
聊天(ep)时建议我在使用(x,y)
计算内核之前缩放scipy
中的值,以便获得与{{1}相当的结果}}
所以这就是我所做的:
sklearn
ie:我在使用# Scale values.
x_val_sca = np.asarray(values[0])/np.asarray(values).std(axis=1)[0]
y_val_sca = np.asarray(values[1])/np.asarray(values).std(axis=1)[1]
values = [x_val_sca, y_val_sca]
kernel = stats.gaussian_kde(values, bw_method=bw_value)
获取内核之前缩放了两个维度,同时保留了scipy
中获取内核的行。
这给出了更好的结果,但在获得的内核方面仍存在差异:
其中红点是代码中的sklearn
点。可以看出,密度估计的形状仍然存在差异,尽管非常小。也许这是可以实现的最佳目标?
答案 0 :(得分:4)
几年后,我尝试了这个,并认为我没有重新扩展数据需要它。带宽值确实需要一些扩展:
# For sklearn
bw = 0.15
# For scipy
bw = 0.15/x.std(ddof=1)
对同一点的两个KDE的评估并不完全相同。例如,这是对(x1, y1)
点的评估:
iso1 = 0.00984751705005 # Scipy
iso2 = 0.00989788224787 # Sklearn
但我觉得它足够接近。
这是2D情况的MWE和输出,据我所知,看起来几乎完全相同:
import numpy as np
from scipy import stats
from sklearn.neighbors import KernelDensity
import matplotlib.pyplot as plt
import matplotlib.gridspec as gridspec
# Generate random data.
n = 1000
m1, m2 = np.random.normal(-3., 3., size=n), np.random.normal(-3., 3., size=n)
# Define limits.
xmin, xmax = min(m1), max(m1)
ymin, ymax = min(m2), max(m2)
ext_range = [xmin, xmax, ymin, ymax]
# Format data.
x, y = np.mgrid[xmin:xmax:100j, ymin:ymax:100j]
positions = np.vstack([x.ravel(), y.ravel()])
values = np.vstack([m1, m2])
# Define some point to evaluate the KDEs.
x1, y1 = 0.5, 0.5
# Bandwidth value.
bw = 0.15
# -------------------------------------------------------
# Perform a kernel density estimate on the data using scipy.
# **Bandwidth needs to be scaled to match Sklearn results**
kernel = stats.gaussian_kde(
values, bw_method=bw/np.asarray(values).std(ddof=1))
# Get KDE value for the point.
iso1 = kernel((x1, y1))
print 'iso1 = ', iso1[0]
# -------------------------------------------------------
# Perform a kernel density estimate on the data using sklearn.
kernel_sk = KernelDensity(kernel='gaussian', bandwidth=bw).fit(zip(*values))
# Get KDE value for the point. Use exponential since sklearn returns the
# log values
iso2 = np.exp(kernel_sk.score_samples([[x1, y1]]))
print 'iso2 = ', iso2[0]
# Plot
fig = plt.figure(figsize=(10, 10))
gs = gridspec.GridSpec(1, 2)
# Scipy
plt.subplot(gs[0])
plt.title("Scipy", x=0.5, y=0.92, fontsize=10)
# Evaluate kernel in grid positions.
k_pos = kernel(positions)
kde = np.reshape(k_pos.T, x.shape)
plt.imshow(np.rot90(kde), cmap=plt.cm.YlOrBr, extent=ext_range)
plt.contour(x, y, kde, 5, colors='k', linewidths=0.6)
# Sklearn
plt.subplot(gs[1])
plt.title("Sklearn", x=0.5, y=0.92, fontsize=10)
# Evaluate kernel in grid positions.
k_pos2 = np.exp(kernel_sk.score_samples(zip(*positions)))
kde2 = np.reshape(k_pos2.T, x.shape)
plt.imshow(np.rot90(kde2), cmap=plt.cm.YlOrBr, extent=ext_range)
plt.contour(x, y, kde2, 5, colors='k', linewidths=0.6)
fig.tight_layout()
plt.savefig('KDEs', dpi=300, bbox_inches='tight')