SciPy中的混合模型拟合(双峰?)使用截断法线。 Python 3

时间:2017-08-04 23:46:19

标签: python optimization scipy statistics distribution

我试图将此作为示例,但似乎无法使其适应我的数据集,因为我需要截断法线: https://stackoverflow.com/questions/35990467/fit-two-gaussians-to-a-histogram-from-one-set-of-data-python#=

我的数据集肯定是2个截断法线的混合。域中的最小值为0,最大值为1.我想创建一个可以适合优化参数的对象,并获得从该分布中抽取一系列数字的可能性。一种选择可能是仅使用KDE模型并使用pdf来获得可能性。但是,我想要2个分布的精确平均值和标准偏差。我想我可以,将数据分成两半,然后分别对2个法线进行建模,但我也想了解如何在optimize中使用SciPy。我刚刚开始尝试这种类型的统计分析,所以如果这看起来很天真,我会道歉。

我不确定如何以这种方式获得pdf,可以集成到1并且域名约束在0和1之间。

import requests
from ast import literal_eval
from scipy import optimize, stats
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np


# Actual Data
u = np.asarray(literal_eval(requests.get("https://pastebin.com/raw/hP5VJ9vr").text))
# u.size ==> 6000
u.min(), u.max()
# (1.3628525454666037e-08, 0.99973136607553781)

# Distribution
with plt.style.context("seaborn-white"):
    fig, ax = plt.subplots()
    sns.kdeplot(u, color="black", ax=ax)
    ax.axvline(0, linestyle=":", color="red")
    ax.axvline(1, linestyle=":", color="red")
kde = stats.gaussian_kde(u)

enter image description here

# KDE Model
def truncated_gaussian_lower(x,mu,sigma,A):
    return np.clip(A*np.exp(-(x-mu)**2/2/sigma**2), a_min=0, a_max=None)
def truncated_gaussian_upper(x,mu,sigma,A):
    return np.clip(A*np.exp(-(x-mu)**2/2/sigma**2), a_min=None, a_max=1)
def mixture_model(x,mu1,sigma1,A1,mu2,sigma2,A2):
    return truncated_gaussian_lower(x,mu1,sigma1,A1) + truncated_gaussian_upper(x,mu2,sigma2,A2)
kde = stats.gaussian_kde(u)

# Estimates: mu sigma A
estimates= [0.1, 1, 3, 
            0.9, 1, 1]
params,cov= optimize.curve_fit(mixture_model,u,kde.pdf(u),estimates )

# ---------------------------------------------------------------------------
# RuntimeError                              Traceback (most recent call last)
# <ipython-input-265-b2efb2ca0e0a> in <module>()
#      32 estimates= [0.1, 1, 3, 
#      33             0.9, 1, 1]
# ---> 34 params,cov= optimize.curve_fit(mixture_model,u,kde.pdf(u),estimates )

# /Users/mu/anaconda/lib/python3.6/site-packages/scipy/optimize/minpack.py in curve_fit(f, xdata, ydata, p0, sigma, absolute_sigma, check_finite, bounds, method, jac, **kwargs)
#     738         cost = np.sum(infodict['fvec'] ** 2)
#     739         if ier not in [1, 2, 3, 4]:
# --> 740             raise RuntimeError("Optimal parameters not found: " + errmsg)
#     741     else:
#     742         # Rename maxfev (leastsq) to max_nfev (least_squares), if specified.

# RuntimeError: Optimal parameters not found: Number of calls to function has reached maxfev = 1400.

回应@Uvar下面非常有用的解释。我试图从0 - 1测试积分,看它是否等于1,但我得到0.3。我想我错过了逻辑中的关键一步:

# KDE Model
def truncated_gaussian(x,mu,sigma,A):
    return A*np.exp(-(x-mu)**2/2/sigma**2)

def mixture_model(x,mu1,sigma1,A1,mu2,sigma2,A2):
    if type(x) == np.ndarray:
        norm_probas = truncated_gaussian(x,mu1,sigma1,A1) + truncated_gaussian(x,mu2,sigma2,A2)
        mask_lower = x < 0
        mask_upper = x > 1
        mask_floor = (mask_lower.astype(int) + mask_upper.astype(int)) > 1
        norm_probas[mask_floor] = 0
        return norm_probas
    else:
        if (x < 0) or (x > 1):
            return 0
        return truncated_gaussian_lower(x,mu1,sigma1,A1) + truncated_gaussian_upper(x,mu2,sigma2,A2)

kde = stats.gaussian_kde(u, bw_method=2e-2)

# # Estimates: mu sigma A
estimates= [0.1, 1, 3, 
            0.9, 1, 1]
params,cov= optimize.curve_fit(mixture_model,u,kde.pdf(u)/integrate.quad(kde, 0 , 1)[0],estimates ,maxfev=5000)
# params
# array([  9.89751700e-01,   1.92831695e-02,   7.84324114e+00,
#          3.73623345e-03,   1.07754038e-02,   3.79238972e+01])

# Test the integral from 0 - 1
x = np.linspace(0,1,1000)
with plt.style.context("seaborn-white"):
    fig, ax = plt.subplots()
    ax.plot(x, kde(x), color="black", label="Data")
    ax.plot(x, mixture_model(x, *params), color="red", label="Model")
    ax.legend()
# Integrating from 0 to 1
integrate.quad(lambda x: mixture_model(x, *params), 0,1)[0]
# 0.3026863969781809

enter image description here

1 个答案:

答案 0 :(得分:2)

您似乎错过了适合程序。 你试图在约束半边界时适应kde.pdf(u)

foo = kde.pdf(u)

min(foo)
Out[329]: 0.22903365654960098

max(foo)
Out[330]: 4.0119283429320332

如您所见,u的概率密度函数不限于[0,1]。 因此,只需删除剪切操作,就可以完全适合。

def truncated_gaussian_lower(x,mu,sigma,A):
    return A*np.exp((-(x-mu)**2)/(2*sigma**2))
def truncated_gaussian_upper(x,mu,sigma,A):
    return A * np.exp((-(x-mu)**2)/(2*sigma**2))
def mixture_model(x,mu1,sigma1,A1,mu2,sigma2,A2):
    return truncated_gaussian_lower(x,mu1,sigma1,A1) + truncated_gaussian_upper(x,mu2,sigma2,A2)

estimates= [0.15, 1, 3, 
            0.95, 1, 1]
params,cov= optimize.curve_fit(f=mixture_model, xdata=u, ydata=kde.pdf(u), p0=estimates)

params
Out[327]: 
array([ 0.00672248,  0.07462657,  4.01188383,  0.98006841,  0.07654998,
        1.30569665])

y3 = mixture_model(u, params[0], params[1], params[2], params[3], params[4], params[5])

plt.plot(kde.pdf(u)+0.1) #add offset for visual inspection purpose

plt.plot(y3)

Perfect overlap, made visible by the +0.1 offset

所以,现在让我说改变我的目标:

plt.figure(); plt.plot(u,y3,'.')

Which does look like what you are trying to achieve

因为,确实:

np.allclose(y3, kde(u), atol=1e-2)
>>True

您可以在域[0, 1]之外将混合模型编辑为0:

def mixture_model(x,mu1,sigma1,A1,mu2,sigma2,A2):
    if (x < 0) or (x > 1):
        return 0
    return truncated_gaussian_lower(x,mu1,sigma1,A1) + truncated_gaussian_upper(x,mu2,sigma2,A2)

然而,这样做会失去在x数组上立即评估函数的选项。因此,为了论证,我暂时将其删除。

无论如何,我们希望我们的积分在域[0, 1]中总计为1,并且有一种方法可以做到这一点(请随意使用stats.gaussian_kde中的带宽估算器...)是将概率密度估计除以其在域上的积分。请注意optimize.curve_fit在此实现中只需要1400次迭代,因此初始参数估计很重要。

from scipy import integrate
sum_prob = integrate.quad(kde, 0 , 1)[0]
y = kde(u)/sum_prob
# Estimates: mu sigma A
estimates= [0.15, 1, 5, 
            0.95, 0.5, 3]
params,cov= optimize.curve_fit(f=mixture_model, xdata=u, ydata=y, p0=estimates)
>>array([  6.72247814e-03,   7.46265651e-02,   7.23699661e+00,
     9.80068414e-01,   7.65499825e-02,   2.35533297e+00])

y3 = mixture_model(np.arange(0,1,0.001), params[0], params[1], params[2], 
    params[3], params[4], params[5])

with plt.style.context("seaborn-white"):
    fig, ax = plt.subplots()
    sns.kdeplot(u, color="black", ax=ax)
    ax.axvline(0, linestyle=":", color="red")
    ax.axvline(1, linestyle=":", color="red")
    plt.plot(np.arange(0,1,0.001), y3) #The red line is now your custom pdf with area-under-curve = 0.998 in the domain..

Total plot

为了检查曲线下方的区域,我使用了这个重新定义mixture_model的hacky解决方案..:

def mixture_model(x):
    mu1=params[0]; sigma1=params[1]; A1=params[2]; mu2=params[3]; sigma2=params[4]; A2=params[5]
    return truncated_gaussian_lower(x,mu1,sigma1,A1) + truncated_gaussian_upper(x,mu2,sigma2,A2)

from scipy import integrate
integrated_value, error = integrate.quad(mixture_model, 0, 1) #0 lower bound, 1 upper bound
>>(0.9978588016186962, 5.222293368393178e-14)

或者以第二种方式做积分:

import sympy
x = sympy.symbols('x', real=True, nonnegative=True)
foo = sympy.integrate(params[2]*sympy.exp((-(x-params[0])**2)/(2*params[1]**2))+params[5]*sympy.exp((-(x-params[3])**2)/(2*params[4]**2)),(x,0,1), manual=True)
foo.doit()

>>0.562981541724715*sqrt(pi) #this evaluates to 0.9978588016186956

实际按照您编辑过的问题中的描述进行操作:

def mixture_model(x,mu1,sigma1,A1,mu2,sigma2,A2):
    return truncated_gaussian_lower(x,mu1,sigma1,A1) + truncated_gaussian_upper(x,mu2,sigma2,A2)
integrate.quad(lambda x: mixture_model(x, *params), 0,1)[0]
>>0.9978588016186962

如果我将带宽设置为您的水平(2e-2),实际上评估结果降至0.92,这比我们之前的0.998更糟糕,但这仍然与您报告的0.3有明显不同。即使在复制代码片段时,我也无法重新创建。您是否可能意外地在某处重新定义了函数/变量?