Fix import style

[OrkoHunter]

With this change and updated #3, networkx-metis works fine on python 2. But has errors with python 3.

Python 3.4.0 (default, Apr 11 2014, 13:05:11) 
Type "copyright", "credits" or "license" for more information.

In [1]: import metis
---------------------------------------------------------------------------
ImportError: Traceback (most recent call last)
<ipython-input-1-f3c558d985ea> in <module>()
----> 1 import metis

/metis.py in <module>()
     16 
     17 from decorators import convert_exceptions
---> 18 from _types import MetisError, MetisOptions
     19 from _metis import node_nd, part_graph, compute_vertex_separator
     20 __all__ = ['node_nested_dissection', 'partition', 'vertex_separator',

/_types.py in <module>()
      2 
      3 from enums import *
----> 4 from _metis import set_default_options
      5 
      6 __all__ = ['MetisOptions', 'MetisError']

ImportError: _metis.so: undefined symbol: _Py_ZeroStruct

[OrkoHunter]

Regarding ImportError: _metis.so: undefined symbol: _Py_ZeroStruct, after some research I came to know that _Py_ZeroStruct symbol is not defined over Python 3. The evidence of this result can also be verified with a grep in /usr/bin/ which gives positive result for the binary file python2.7 but no result for other versions of python.

[OrkoHunter]

Updated. we should consider using import modules; module.item() instead of from module import item

[OrkoHunter]

Updated _metis.pyx with

from networkx.addons.metis import _types

from networkx.addons.metis cimport _api

EDIT: It won't build at first place because cython would be able not find _api at networkx.addons.metis to cimport.

[OrkoHunter]

This diff works for StringIO instance has no attribute 'fileno' Now the tests are passing.

diff --git a/_metis.pyx b/_metis.pyx
index e13d742..b956af3 100644
--- a/_metis.pyx
+++ b/_metis.pyx
@@ -264,7 +264,7 @@ def part_graph(xadj, adjncy, nparts, vwgt=None, vsize=None, adjwgt=None,
         _recursive = bool(recursive)

         with tempfile.TemporaryFile() as tmp:
-            with redirect(sys.stdout, tmp), nogil:
+            with redirect(sys.__stdout__, tmp), nogil:
                 if _recursive:
                     result = _api.METIS_PartGraphRecursive(
                         &nvtxs, &ncon, _xadj, _adjncy, _vwgt, _vsize, _adjwgt,
@@ -274,7 +274,7 @@ def part_graph(xadj, adjncy, nparts, vwgt=None, vsize=None, adjwgt=None,
                         &nvtxs, &ncon, _xadj, _adjncy, _vwgt, _vsize, _adjwgt,
                         &_nparts, _tpwgts, _ubvec, _options, &objval, _part)
             tmp.seek(0)
-            msg = unicode(tmp.read(), sys.stdout.encoding)
+            msg = unicode(tmp.read(), sys.__stdout__.encoding)

         check_result(result, msg)

@@ -354,11 +354,11 @@ def node_nd(xadj, adjncy, vwgt=None, options=None):
         _iperm = <_api.idx_t*> checked_malloc(sizeof(_api.idx_t) * nvtxs)

         with tempfile.TemporaryFile() as tmp:
-            with redirect(sys.stdout, tmp), nogil:
+            with redirect(sys.__stdout__, tmp), nogil:
                 result = _api.METIS_NodeND(&nvtxs, _xadj, _adjncy, _vwgt, _options,
                                       _perm, _iperm)
             tmp.seek(0)
-            msg = unicode(tmp.read(), sys.stdout.encoding)
+            msg = unicode(tmp.read(), sys.__stdout__.encoding)

         check_result(result, msg)

@@ -425,11 +425,11 @@ def compute_vertex_separator(xadj, adjncy, vwgt=None, options=None):
         _part = <_api.idx_t*> checked_malloc(sizeof(_api.idx_t) * nvtxs)

         with tempfile.TemporaryFile() as tmp:
-            with redirect(sys.stdout, tmp), nogil:
+            with redirect(sys.__stdout__, tmp), nogil:
                 result = _api.METIS_ComputeVertexSeparator(
                     &nvtxs, _xadj, _adjncy, _vwgt, _options, &sepsize, _part)
             tmp.seek(0)
-            msg = unicode(tmp.read(), sys.stdout.encoding)
+            msg = unicode(tmp.read(), sys.__stdout__.encoding)

         check_result(result, msg)

[chebee7i]

Can you explain why you are reassigning sys.__stdout__ rather than sys.stdout? The double-underscored stdout is supposed to be a reference to the one true stdout. When people reassign sys.stdout, they revert to the double-underscored version. https://docs.python.org/2/library/sys.html#sys.__stdout__

[OrkoHunter]

I didn't read much about it but this was a fix for the failing test, I found on stackoverflow. I'll look for a possible explanation.

[debugger22]

You might want to read this https://docs.python.org/3.3/library/sys.html#sys.__stdout__

[OrkoHunter]

It seems that nosetests is the culprit. They probably replace sys.stdout with a StringIO object so that it can capture the output easily. But maybe they do not interfere with sys.__stdout__. I realize very well about the similarities of the two, and both works while importing/compiling but nosetests are making the difference.

Source : http://stackoverflow.com/questions/30961173/sys-stdout-works-but-sys-stdout-does-not#comment49954775_30961173

[OrkoHunter]

There's more!

$ nosetests -s
....
----------------------------------------------------------------------
Ran 4 tests in 0.009s

OK

works with sys.stdout.

This plugin captures stdout during test execution. If the test fails or raises an error, the captured output will be appended to the error or failure output. It is enabled by default but can be disabled with the options -s or --nocapture.

http://nose.readthedocs.org/en/latest/plugins/capture.html

[chebee7i]

The other option mentioned on that SO post seems more robust to me than what is currently implemented (since it doesn't assume that stdout has an fd).

[OrkoHunter]

Or should we go with nosetests -s on Travis without any changes?

[chebee7i]

The nosetests -s still requires that other code has not replaced sys.stdout with a StringIO instance. So I don't think it is a general solution.

[ysitu]

The reason why we need to redirect stdout is that METIS outputs error messages to stdout. We need to remember that METIS is a C library. It knows nothing about Python's sys.stdout (or sys.__stdout__) object. Therefore, what is important is that the file descriptor of stdout in C, which should be 1 on common operating systems. So I think that it would not be too bad that we use

fd = sys.__stdout__.fileno() if sys.__stdout__ else 1

or even

fd = 1

Separately from obtaining the file descriptor of stdout, this thread brings up a very good problem of reliably flushing stdout. We need to handle that.

[chebee7i]

Got it. That makes the implementation a bit more specific than I thought---namely, it is specifically targeting stdout generated from C functions rather than Python programs. Makes sense.

[ysitu]

Separately from obtaining the file descriptor of stdout, this thread brings up a very good problem of reliably flushing stdout. We need to handle that.

BTW, we are dealing with this in Cython. So we do have (somewhat) direct access to C facilities. In this case, cimporting fflush and stdout from libc.stdio and fflush(stdout) should suffice.

[OrkoHunter]

I guess this 66d4f92f375a90b054e0de10467e44023afbb9cd is pretty much okay to do the job. sys.stdout.encoding also needed to be changed.

[ysitu]

Can we separate the stdout changes to another PR?

[OrkoHunter]

Sounds good to create new PR for last commit.

[OrkoHunter]

Will send a PR shortly with the changes related to stdout. Last commit removed from this PR.

[OrkoHunter]

@ysitu Thanks!