python - curve_fit is seemingly unable to fit sum of gaussians

Question

Here I have defined a function to return the sum of an arbitrary number of Gaussian Distributions:

import numpy
from numpy import *
import matplotlib.pyplot as plt
from scipy.optimize import curve_fit

def GaussSum(x,*p):
    n=len(p)/3
    A=p[:n]
    w=p[n:2*n]
    c=p[2*n:3*n]
    return sum([ A[i]*exp(-(x-c[i])**2./(2.*(w[i])**2.))/(2*pi*w[i]**2)**0.5 for i in range(n)])

I then proceed to generate x and y data for a given set of parameters and ask curve_fit to fit this data with the initial parameters matching the generating set. I tried this for many different sets including both single and multiple Gaussians.

params = [1.,1.,-3.]; #parameters for a single gaussian
#params=[1.,1.,1.,2.,-3.,0.]; #parameters for the sum of two gaussians
xdata=arange(-6,6,0.01)
ydata = array([GaussSum(x,*params) for x in xdata])

popt,pcov = curve_fit(GaussSum,xdata,ydata,p0=params)

print popt
print pcov

Every parameter set gives me a non-fits, even though I already supposedly started the fit at the solution. (In the above single Gaussian):

[    52.18242366   5549.66965192  15678.51803797]
inf

I know that the function itself operates normally, as I have plotted with it and verified its validity.

Conor · Accepted Answer · 2015-07-08T22:10:18.037

The problem here is that curve_fit expects the function you are optimising to take a vector of inputs and return a vector of outputs. As written, the function GaussSum takes a single input or a vector of inputs (i.e. a numpy.array) but either way it returns a single scalar output. The curve_fit function then fails to find a good optimum fit.

In order to clarify what's going on, it's always advisable when using numpy (or any external library) to always use the namespace explicitly, as shown in following working version:

import numpy
import matplotlib.pyplot as plt
from scipy.optimize import curve_fit

def GaussSum(x,*p):
    n=len(p)/3
    A=p[:n]
    w=p[n:2*n]
    c=p[2*n:3*n]
    y = sum([ A[i]*numpy.exp(-(x-c[i])**2./(2.*(w[i])**2.))/(2*numpy.pi*w[i]**2)**0.5 for i in range(n)])
    return y

params = [1.,1.,-3.]; #parameters for a single gaussian                                                                                                                        
#params=[1.,1.,1.,2.,-3.,0.]; #parameters for the sum of two gaussians                                                                                                         
xdata=numpy.arange(-6,6,0.01)
ydata = numpy.array([GaussSum(x,*params) for x in xdata])

popt,pcov = curve_fit(GaussSum,xdata,ydata,p0=params)

Specifically you were implicitly calling numpy.sum, which would aggregate the numpy.array and return a single value, while you needed to use the builtin python sum, which aggregates the native list of numpy.array to one numpy.array.

Thank you. That clarifies what the issue was. One item though, the working code needs to have a return instead of a y = in the function definition. — brad.t, Jul 08 '15 at 22:06

python - curve_fit is seemingly unable to fit sum of gaussians

1 Answers1