[Scipy-tickets] [SciPy] #392: optimize.leastsq segfaults

SciPy scipy-tickets@scipy....
Sat Mar 24 11:29:39 CDT 2007

#392: optimize.leastsq segfaults
 Reporter:  pwuertz         |       Owner:  somebody                         
     Type:  defect          |      Status:  new                              
 Priority:  normal          |   Milestone:                                   
Component:  scipy.optimize  |     Version:  0.5.2                            
 Severity:  critical        |    Keywords:  minpack optimize leastsq segfault
 Hi, I experienced some strange segfaults when using python/scipy in my
 application. I think I tracked it down to optimize.leastsq. This function
 is highly unstable when using jacobians and no full output.
 For some numbers of equations it will segfault. But its definately going
 to segfault if you are using equation vectors greater than 2^16^.

 This is a minimal example:

 from numpy import ones, array
 from scipy import optimize

 N = 2**16 + 1
 par_init  = ones(2)
 func      = lambda p: ones( N )
 jacobian  = lambda p: ones((N, 2))

 print "these functions are working"
 result = optimize.leastsq(func, par_init, Dfun = jacobian, full_output=1)
 result = optimize.leastsq(func, par_init, full_output=1)
 result = optimize.leastsq(func, par_init)

 print "this one is crashing"
 result = optimize.leastsq(func, par_init, Dfun = jacobian)

 My guess is the bug is in LMDER, where LMDIF is clean, because the
 functions without jacobians are working.

 I dont know how full_output is affecting the minpack binding... maybe a
 different function will be called, no idea. Also the 2^16^ condition is
 strange... looks like a 16 bit int has been used for looping through the
 array or something like that.

 As workaround for this problem one could always use the full_output=1
 option... but I dont assume this is the intended behavior *g*.

Ticket URL: <http://projects.scipy.org/scipy/scipy/ticket/392>
SciPy <http://www.scipy.org/>
SciPy is open-source software for mathematics, science, and engineering.

More information about the Scipy-tickets mailing list