[Scikit-learn-general] On building scikit-learn-0.14.1 with STATIC ATLAS on non-standard location

[Sergio Rojas]

Hello everyone,

I was able to compile scikit-learn-0.14.1 with STATIC version of ATLAS using
the building command:

python setup.py build_ext --include-dirs='/home/myacct/myProg/NumLibs64b/include
/ -L/home/srojas/myProg/NumLibs64b/lib -llapack -lcblas -lf77blas -latlas' --inplace

I should mention that I came up with this way to build scikit-learn after installing mlpy [http://mlpy.sourceforge.net/docs/3.5/install.html#installing-on-nix-from-source ].

I am curious about the following:

The pre-installation test
python nosetests sklearn/

ended with (FULL OUTPUT at [ https://gist.github.com/anonymous/7089302 ] ):
...
----------------------------------------------------------------------
Ran 1860 tests in 137.411s
FAILED (SKIP=14, errors=1, failures=2)

However, the post-installation test
nosetests --exe sklearn

ended with (FULL OUTPUT at [ https://gist.github.com/anonymous/7124307 ] ):
----------------------------------------------------------------------
Ran 1720 tests in 148.112s
FAILED (SKIP=14, errors=1, failures=2)

Should not be the same number of test run by both commands?

I executed successfully a few examples from the page
http://scikit-learn.org/stable/auto_examples/index.html

Is there any other way to test the installation of this
module?


Sergio

[Olivier Grisel]

Post by Sergio Rojas
Hello everyone,
I was able to compile scikit-learn-0.14.1 with STATIC version of ATLAS using
python setup.py build_ext
--include-dirs='/home/myacct/myProg/NumLibs64b/include
/ -L/home/srojas/myProg/NumLibs64b/lib -llapack -lcblas -lf77blas -latlas'
--inplace
I should mention that I came up with this way to build scikit-learn after
installing mlpy
[http://mlpy.sourceforge.net/docs/3.5/install.html#installing-on-nix-from-source
].

Interesting, thanks for the reference.

Post by Sergio Rojas
The pre-installation test
python nosetests sklearn/
...
----------------------------------------------------------------------
Ran 1860 tests in 137.411s
FAILED (SKIP=14, errors=1, failures=2)
However, the post-installation test
nosetests --exe sklearn
----------------------------------------------------------------------
Ran 1720 tests in 148.112s
FAILED (SKIP=14, errors=1, failures=2)
Should not be the same number of test run by both commands?

Weird. I think we should include all the tests in the distribution.

FYI: the test_image.py failures are known: they are caused by a change
in recent versions of scipy (after 0.12). Nobody took the time to
investigate though.

AFAIK the bi-cluster test failure is unknown. I cannot reproduce it my
self. which version of scipy are you using?

Post by Sergio Rojas
I executed successfully a few examples from the page
http://scikit-learn.org/stable/auto_examples/index.html
Is there any other way to test the installation of this
module?

1700+ tests are not enough for you? :)

[Olivier Grisel]

Post by Sergio Rojas
Hello everyone,
I was able to compile scikit-learn-0.14.1 with STATIC version of ATLAS using
python setup.py build_ext
--include-dirs='/home/myacct/myProg/NumLibs64b/include
/ -L/home/srojas/myProg/NumLibs64b/lib -llapack -lcblas -lf77blas -latlas'
--inplace

AFAIK you should not use --include-dirs to pass library link
instructions but rather --libraries and --library-dirs.

[Lars Buitinck]

Post by Olivier Grisel

Post by Sergio Rojas
python setup.py build_ext
--include-dirs='/home/myacct/myProg/NumLibs64b/include
/ -L/home/srojas/myProg/NumLibs64b/lib -llapack -lcblas -lf77blas -latlas'
--inplace

AFAIK you should not use --include-dirs to pass library link
instructions but rather --libraries and --library-dirs.

Indeed, this will pass -L to the compiler but not the linker, and
you'll get a mismatch between the BLAS/LAPACK headers and libraries.
There are, unfortunately, subtle differences between BLAS implems esp.
in terms of return types.

[Sergio Rojas]

AFAIK the bi-cluster test failure is unknown. I cannot reproduce it my > self. which version of scipy are you using? >