[SciPy-dev] Segmentation fault triggered by optimize/lbfgsb.py

[Travis Oliphant]

Ed Schofield wrote:

>The code in optimize/lbfgsb.py raises an exception and then dumps core:
>
>Traceback (most recent call last):
>  ...
>  File
>"/home/schofield/Test/lib/python2.4/site-packages/scipy/optimize/lbfgsb.py",
>line 182, in fmin_l_bfgs_b
>    isave, dsave)
>ValueError: string_from_pyobj failed in converting 14th argument `csave'
>of _lbfgsb.setulb to C string
>Segmentation fault (core dumped)
>
>
>What is interesting is that I can get a segfault with the same backtrace
>even *before* the call to the Fortran function, just by executing some
>of the statements up to line 182 and exiting the python interpreter with
>^D, implying that the interpreter is no longer in a stable state.
>
>
># scipy.base.__version__    gives '0.4.3.1343'
>
>  
>
>>>>import scipy
>>>>        
>>>>
>
>  
>
>>>>task = scipy.zeros((60,), scipy.Character)
>>>>task[:] = 'blah'
>>>>^D
>>>>        
>>>>
>Segmentation fault
>
>Here's a backtrace from gdb:
>#0  0x080de87d in Py_GetPath ()
>#1  0x08087860 in PyTuple_Pack ()
>#2  0x080df0c0 in _PyObject_GC_Track ()
>#3  0x080df897 in PyGC_Collect ()
>#4  0x080d88f9 in Py_Finalize ()
>#5  0x080556f0 in Py_Main ()
>#6  0x4938dea2 in __libc_start_main () from /lib/tls/i686/cmov/libc.so.6
>#7  0x08054f21 in _start ()
>
>Anyone want to take the baton? ;)
>
>  
>
One part of the problem has been fixed in new scipy core. 

Now, f2py may need to change to handle character arrays differently.

Character arrays work a little differently than they used to.  Before 
they were just special-cases of UBYTE arrays.  Now, character arrays are 
arrays of strings.  Thus, the lbfgsb.py code should use

task = zeros(1, 'S60')  so that
task[:] = 'blah'  will do the right thing.

task = zeros(1,Character) is equivalent to
task = zeros(1,'S1')

so that

task[:] = 'blah' is the same as
task[:] = array('blah','S1')

which fills the array with 'b'


-Travis


-Travis