Scipy optimize.curve_fit sometimes won't converge

Question

I'm trying to use numpy.optimize.curve_fit to estimate the frequency and phase of an on/off sequence. This is the code I'm using:

from numpy import *
from scipy import optimize

row = array([0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 1.0, 1.0, 1.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 1.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 1.0, 1.0, 0.0, 0.0,])

def fit_func(x, a, b, c, d):
    return c * sin (a * x + b) + d

p0 = [(pi/10.0), 5.0, row.std(), row.mean()]
result = optimize.curve_fit(fit_func, arange(len(row)), row, p0)
print result

This works. But on some rows, even though they seem perfectly ok, it fails. Example of failing row:

row = array([1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0, 0.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0, 0.0,  0.0,  0.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  0.0,  0.0,  0.0,  0.0, 0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  0.0,  0.0, 0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0,  1.0, 1.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0,  0.0,])

The error is:

RuntimeError: Optimal parameters not found: Both actual and predicted relative reductions in the sum of squares are at most 0.000000 and the relative error between two consecutive iterates is at most 0.000000

Which tells me very little about what's happened. A quick test shows that varying the parameters in p0 will cause that row to succeed... and others to fail. Why is that?

Answer 1

I tried both rows of data that you provided and both worked for me just fine. I'm using Scipy 0.8.0rc3. What version are you using? Another thing that might help is to set c and d to fixed values since they really should be the same every time. I set c to 0.6311786 and d to .5. You could also use an fft with zero padding and quadratic fitting around the peak to find the frequency if you want another method. Really, any pitch estimation method is applicable since you are looking for the fundamental frequency.