Link error on Windows

[Roottan]

When I recomiling and linking the code ,find some error ,like this:

Creating library build\temp.win32-2.7\Release\sklearn\linear_model\sgd_fast.l
ib and object build\temp.win32-2.7\Release\sklearn\linear_model\sgd_fast.exp
sgd_fast.obj : error LNK2019: unresolved external symbol _isfinite referenced in
 function ___pyx_f_7sklearn_12linear_model_8sgd_fast_any_nonfinite
build\lib.win32-2.7\sklearn\linear_model\sgd_fast.pyd : fatal error LNK1120: 1 u
nresolved externals
error: Command "C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\BIN\link.e
xe /DLL /nologo /INCREMENTAL:NO /LIBPATH:C:\Python27\libs /LIBPATH:C:\Python27\P
Cbuild /LIBPATH:build\temp.win32-2.7 /EXPORT:initsgd_fast build\temp.win32-2.7\R
elease\sklearn\linear_model\sgd_fast.obj /OUT:build\lib.win32-2.7\sklearn\linear
_model\sgd_fast.pyd /IMPLIB:build\temp.win32-2.7\Release\sklearn\linear_model\sg
d_fast.lib /MANIFESTFILE:build\temp.win32-2.7\Release\sklearn\linear_model\sgd_f
ast.pyd.manifest" failed with exit status 1120

OS: windows7 64bit
python: 2.7 32bit
VC: 2008 express version

[ogrisel]

It is probably caused by the fact that we generated the windows wheel against a more recent version of numpy than the one you have installed.

Could you please tell us which version of numpy (and scipy) are installed on your system? Can you please try to upgrade to the latest stable version of numpy and scipy and see if that fixes the problem?

[ogrisel]

I think this is the same problem as #3548 but for Windows.

[ogrisel]

Actually this is not the same problem as #3548 as the issue reported by @Roottan is a compile time problem and not a runtime problem.

I think this is a problem in numpy. I found this issue numpy/numpy#2099 that seems related but was fixed a long time ago in numpy. I can reproduce @Roottan's error when building against the current official numpy 1.8.1 for windows. However I don't have the problem when I build against @cgohlke's numpy from http://www.lfd.uci.edu/~gohlke/pythonlibs/ so it seems that @cgohlke's version of numpy has been patched in some way to workaround this issue.

[ogrisel]

@cournape @asmeurer didn't you encounter a similar issue when building your own scikit-learn package for 32 bit Python on Windows?

[cournape]

@ogrisel nope. Which binary installer did you use exactly for numpy ?

[cgohlke]

I think the issue is that npy_isfinite, used by sgd_fast.pyx, is a macro. The definition of the macro depends on configuration settings in _numpyconfig.h, which is generated when building numpy and depends on the compiler used to build numpy. In this case, if numpy is build with mingw, NPY_HAVE_DECL_ISFINITE is defined in _numpyconfig.h and npy_isfinite(x) is isfinite((x)), which does not work with msvc...

[cournape]

@cgohlke's analysis is correct. One cannot currently build numpy with one compiler and other extensions with another. It may be fairly hard to fix in numpy.

[ogrisel]

Alright thanks for you analysis @cgohlke and @cournape.

I will just redefine the _isfinite macro in scikit-learn then. @larsmans do you have a suggestion where we should put that? Just in the sgd module or do you want to mutualize those kind of macros in a common project level header file for scikit-learn?

[larsmans]

I would put it in a little header file next to SGD for now. We can then move that elsewhere if the need arises.

#ifdef _MSC_VER
# include <float.h>
# define isfinite _finite
#else
# include <numpy/npy_math.h>
# define isfinite npy_isfinite
#endif

[ogrisel]

@larsmans I just saw your comment. I will do what you suggest in #3611.