[pymvpa] shogun 6.6/pymvpa bug

[Yaroslav Halchenko]

Hi Scott,

Thanks for the report... I hadn't tried recent SG thus I didn't hit it but
indeed it seems SG team introduced non-backward compatible changes without
properly announcing (ie really screaming out loud ;-)) it nor by
signaling about that in the version change (afaik change of minor in the
version should be backward compatible). heh heh

All those changes originate from a new feature in SG -- various kernel
normalization strategies which is a great feature to have but once again --
such changes make it really inconvenient for us users to stay compatible with
various versions of shogun.

Here is a snippet from the change in example (rev 3359 change) from
shogun python-modular which reveals what they want us to do:

-	kernel=LinearKernel(feats_train, feats_train, scale)
+	kernel=LinearKernel()
+	kernel.set_normalizer(AvgDiagKernelNormalizer(scale))
+	kernel.init(feats_train, feats_train)

I will ask SG team if they would be kind to introduce compatibility patch,
otherwise we will need to create custom handling for different versions of
shogun (which I would prefer to avoid)

I will keep you updated

Cheers
Yarik

On Sat, 01 Nov 2008, Scott Gorlin wrote:

> Hi,

> I just upgraded my system, and after reinstalling some packages, noticed 
> a few bugs.  Pymvpa does not seem to work with the latest release of 
> Shogun 0.6.6.

> Running clfs_examples gives the following output then error:

> Dummy 2-class univariate with 2 useful features out of 100
>  Dataset / float64 60 x 100
> uniq: 2 labels 6 chunks labels_mapped
> stats: mean=0.00195773 std=0.213342 var=0.0455148 min=-0.752662 max=1
> Counts of labels in each chunk:
>   chunks\labels  0   1
>                 --- ---
>        0         5   5
>        1         5   5
>        2         5   5
>        3         5   5
>        4         5   5
>        5         5   5

> Original labels were mapped using following mapping:
>     L0:    0
>     L1:    1

> Summary per label across chunks
>   label mean std min max #chunks
>    0      5   0   5   5     6
>    1      5   0   5   5     6

> Summary per chunk across labels
>   chunk mean std min max #labels
>    0      5   0   5   5     2
>    1      5   0   5   5     2
>    2      5   0   5   5     2
>    3      5   0   5   5     2
>    4      5   0   5   5     2
>    5      5   0   5   5     2

>  Classifier                                  %corr  #features     train 
> predict  full
>   SMLR(lm=0.1)                            :   85.0%   22       0.07s  
> 0.00s   0.43s
>   SMLR(lm=1.0)                            :   91.7%   10       0.00s  
> 0.00s   0.04s
>   Pairs+maxvote multiclass on SMLR(lm=0.1):   85.0%   100      0.08s  
> 0.00s   0.48s
>   libsvm.LinSVM(C=def)                    :   78.3%   100      0.01s  
> 0.00s   0.10s
>   libsvm.LinSVM(C=10*def)                 :   81.7%   100      0.01s  
> 0.00s   0.11s
>   libsvm.LinSVM(C=1)                      :   81.7%   100      0.01s  
> 0.00s   0.11s
>   libsvm.LinNuSVM(nu=def)                 :   78.3%   100      0.01s  
> 0.00s   0.11s
>   libsvm.RbfSVM()                         :   75.0%   100      0.01s  
> 0.00s   0.08s
>   libsvm.RbfNuSVM(nu=def)                 :   78.3%   100      0.01s  
> 0.00s   0.08s
>   libsvm.PolySVM()                        :   75.0%   100      0.01s  
> 0.00s   0.10s
>   sg.LinSVM(C=def)/libsvm                 :
> Traceback (most recent call last):
>   File "clfs_examples.py", line 90, in <module>
>     main()
>   File "clfs_examples.py", line 71, in main
>     clf.train(training_ds)
>   File "/usr/lib/python2.5/site-packages/mvpa/clfs/base.py", line 375, 
> in train
>     result = self._train(dataset)
>   File "/usr/lib/python2.5/site-packages/mvpa/clfs/sg/svm.py", line 287, 
> in _train
>     *kargs)
>   File "/usr/local/lib/python2.5/site-packages/shogun/Kernel.py", line 
> 1303, in __init__
>     this = _Kernel.new_LinearKernel(*args)
> NotImplementedError: Wrong number of arguments for overloaded function 
> 'new_LinearKernel'.
>   Possible C/C++ prototypes are:
>     CLinearKernel()
>     CLinearKernel(CRealFeatures *,CRealFeatures *)


> I placed the crossvalidation step in a try block, and noticed that most 
> of the shogun classes in clfs_examples failed.  The exceptions (which 
> did work) called sg.SomeClass() without any arguments, leading me to 
> believe there is some weird API change in shogun.  All of the examples 
> in Shogun's python-modular tests work just fine.

> After I reinstalled Shogun 0.6.4, everything worked fine.  Then I 
> reinstalled 0.6.6, and it didn't work again.

> I am running Ubuntu 8.10 in python 2.5 with pymvpa 0.3.1.

> Thanks!
> Scott



> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/mailman/listinfo/pkg-exppsy-pymvpa


-- 
Yaroslav Halchenko
Research Assistant, Psychology Department, Rutgers-Newark
Student  Ph.D. @ CS Dept. NJIT
Office: (973) 353-1412 | FWD: 82823 | Fax: (973) 353-1171
        101 Warren Str, Smith Hall, Rm 4-105, Newark NJ 07102
WWW:     http://www.linkedin.com/in/yarik