FANode training does not work with a large matrix with big values

[Joonas Paalasmaa]

When a large matrix with big values is used, FANode training fails.

~$ python2.5
Python 2.5.4 (r254:67916, Apr 4 2009, 17:55:16)
[GCC 4.3.3] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import mdp
>>> import numpy
>>> fa = mdp.nodes.FANode()
>>> fa.train(numpy.random.randn(200, 50)*1000000)
>>> fa.stop_training()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/var/lib/python-support/python2.5/mdp/signal_node.py", line 493, in stop_training
    self._train_seq[self._train_phase][1](*args, **kwargs)
  File "/var/lib/python-support/python2.5/mdp/nodes/em_nodes.py", line 106, in _stop_training
    log_det_B = numx.log(abs(det(B)))
  File "/usr/local/lib/python2.5/site-packages/scipy/linalg/basic.py", line 480, in det
    a1 = asarray_chkfinite(a)
  File "/usr/lib/python2.5/site-packages/numpy/lib/function_base.py", line 724, in asarray_chkfinite
    raise ValueError, "array must not contain infs or NaNs"
ValueError: array must not contain infs or NaNs
>>> mdp.__version__
'2.5'

ProblemType: Bug
Architecture: i386
DistroRelease: Ubuntu 9.04
NonfreeKernelModules: fglrx
Package: python-mdp 2.4+svn20081126-1
PackageArchitecture: all
ProcEnviron:
 LANGUAGE=
 LANG=en_US.UTF-8
 SHELL=/bin/bash
SourcePackage: mdp
Uname: Linux 2.6.28-11-generic i686