Description:

Overview

The SHOGUN machine learning toolbox's focus is on large scale kernel methods and especially on Support Vector Machines (SVM). It comes with a generic interface for kernel machines and features 15 different SVM implementations that all access features in a unified way via a general kernel framework or in case of linear SVMs so called "DotFeatures", i.e., features providing a minimalistic set of operations (like the dot product).

Features

SHOGUN includes the LinAdd accelerations for string kernels and the COFFIN framework for on-demand computing of features for the contained linear SVMs. In addition it contains more advanced Multiple Kernel Learning, Multi Task Learning and Structured Output learning algorithms and other linear methods. SHOGUN digests input feature-objects of basically any known type, e.g., dense, sparse or variable length features (strings) of any type char/byte/word/int/long int/float/double/long double.

The toolbox provides efficient implementations to 35 different kernels among them the

• Linear,
• Polynomial,
• Gaussian and
• Sigmoid Kernel

and also provides a number of recent string kernels like the

• Locality Improved,
• Fischer,
• TOP,
• Spectrum,
• Weighted Degree Kernel (with shifts) .

For the latter the efficient LINADD optimizations are implemented. Also SHOGUN offers the freedom of working with custom pre-computed kernels. One of its key features is the combined kernel which can be constructed by a weighted linear combination of a number of sub-kernels, each of which not necessarily working on the same domain. An optimal sub-kernel weighting can be learned using Multiple Kernel Learning. Currently SVM one-class, 2-class, multi-class classification and regression problems are supported. However SHOGUN also implements a number of linear methods like

• Linear Discriminant Analysis (LDA)
• Linear Programming Machine (LPM),
• Perceptrons and features algorithms to train Hidden Markov Models.

The input feature-objects can be read from plain ascii files (tab separated values for dense matrices; for sparse matrices libsvm/svmlight format), a efficient native binary format and general support to the hdf5 based format, supporting

• dense
• sparse or
• strings of various types

that can often be converted between each other. Chains of preprocessors (e.g. subtracting the mean) can be attached to each feature object allowing for on-the-fly pre-processing.

Structure and Interfaces

SHOGUN's core is implemented in C++ and is provided as a library libshogun to be readily usable for C++ application developers. Its common interface functions are encapsulated in libshogunui, such that only minimal code (like setting or getting a double matrix to/from the target language) is necessary. This allowed us to easily create interfaces to Matlab(tm), R, Octave and Python. (note that a modular object oriented and static interfaces are provided to r, octave, matlab, python, python_modular, r_modular, octave_modular, cmdline, libshogun).

Application

We have successfully applied SHOGUN to several problems from computational biology, such as Super Family classification, Splice Site Prediction, Interpreting the SVM Classifier, Splice Form Prediction, Alternative Splicing and Promoter Prediction. Some of them come with no less than 10 million training examples, others with 7 billion test examples.

Documentation

We use Doxygen for both user and developer documentation which may be read online here. More than 600 documented examples for the interfaces python_modular, octave_modular, r_modular, static python, static matlab and octave, static r, static command line and C++ libshogun developer interface can be found in the documentation.

Changes to previous version:

This release also contains several enhancements, cleanups and bugfixes:

Features

• Linear Time MMD two-sample test now works on streaming-features, which allows to perform tests on infinite amounts of data. A block size may be specified for fast processing. The below features were also added. By Heiko Strathmann.
• It is now possible to ask streaming features to produce an instance of streamed features that are stored in memory and returned as a CFeatures* object of corresponding type. See CStreamingFeatures::get_streamed_features().
• New concept of artificial data generator classes: Based on streaming features. First implemented instances are CMeanShiftDataGenerator and CGaussianBlobsDataGenerator. Use above new concepts to get non-streaming data if desired.
• Accelerated projected gradient multiclass logistic regression classifier by Sergey Lisitsyn.
• New CCSOSVM based structured output solver by Viktor Gal
• A collection of kernel selection methods for MMD-based kernel two- sample tests, including optimal kernel choice for single and combined kernels for the linear time MMD. This finishes the kernel MMD framework and also comes with new, more illustrative examples and tests. By Heiko Strathmann.
• Alpha version of Perl modular interface developed by Christian Montanari.
• New framework for unit-tests based on googletest and googlemock by Viktor Gal. A (growing) number of unit-tests from now on ensures basic funcionality of our framework. Since the examples do not have to take this role anymore, they should become more ilustrative in the future.
• Changed the core of dimension reduction algorithms to the Tapkee library.

Bugfixes

• Fix for shallow copy of gaussian kernel by Matt Aasted.
• Fixed a bug when using StringFeatures along with kernel machines in cross-validation which cause an assertion error. Thanks to Eric (yoo)!
• Fix for 3-class case training of MulticlassLibSVM reported by Arya Iranmehr that was suggested by Oksana Bayda.
• Fix for wrong Spectrum mismatch RBF construction in static interfaces reported by Nona Kermani.
• Fix for wrong include in SGMatrix causing build fail on Mac OS X (thanks to @bianjiang).
• Fixed a bug that caused kernel machines to return non-sense when using custom kernel matrices with subsets attached to them.
• Fix for parameter dictionary creationg causing dereferencing null pointers with gaussian processes parameter selection.
• Fixed a bug in exact GP regression that caused wrong results.
• Fixed a bug in exact GP regression that produced memory errors/crashes.
• Fix for a bug with static interfaces causing all outputs to be -1/+1 instead of real scores (reported by Kamikawa Masahisa).

Cleanup and API Changes

• SGStringList is now based on SGReferencedData.
• "confidences" in context of CLabel and subclasses are now "values".
• CLinearTimeMMD constructor changes, only streaming features allowed.
• CDataGenerator will soon be removed and replaced by new streaming- based classes.
• SGVector, SGMatrix, SGSparseVector, SGSparseVector, SGSparseMatrix refactoring: Now contains load/save routines, relevant functions from CMath, and implementations went to .cpp file.

BibTeX Entry: Download

Corresponding Paper BibTeX Entry: Download

Supported Operating Systems: Cygwin, Linux, Macosx, Bsd

Data Formats: Plain Ascii, Svmlight, Binary, Fasta, Fastq, Hdf

Tags: Bioinformatics, Large Scale, String Kernel, Kernel, Kernelmachine, Lda, Lpm, Matlab, Mkl, Octave, Python, R, Svm, Sgd, Icml2010, Liblinear, Libsvm, Multiple Kernel Learning, Ocas, Gaussian Processes, Reg

Archive: download here

Comments

Soeren Sonnenburg (on September 12, 2008, 16:14:36)

In case you find bugs, feel free to report them at [http://trac.tuebingen.mpg.de/shogun](http://trac.tuebingen.mpg.de/shogun).

Tom Fawcett (on January 3, 2011, 03:20:48)

You say, "Some of them come with no less than 10 million training examples, others with 7 billion test examples." I'm not sure what this means. I have problems with mixed symbolic/numeric attributes and the training example sets don't fit in memory. Does SHOGUN require that training examples fit in memory?

Soeren Sonnenburg (on January 14, 2011, 18:12:01)

Shogun does not necessarily require examples to be in memory (if you use any of the FileFeatures). However, most algorithms within shogun are batch type - so using the non in-memory FileFeatures would probably be very slow. This does not matter for doing predictions of course, even though the 7 billion test examples above referred to predicting gene starts on the whole human genome (in memory ~3.5GB and a context window of 1200nt was shifted around in that string). In addition one can compute features (or feature space) on-the-fly potentially saving lots of memory. Not sure how big your problem is but I guess this is better discussed on the shogun mailinglist.

Yuri Hoffmann (on September 14, 2013, 17:12:16)

cannot use the java interface in cygwin (already reported on github) nor in debian.

Leave a comment

You must be logged in to post comments.