LIBSVM

Help edit Baike card

LIBSVM is a large Pak Lin Zhiren Taiwan (Lin Chih-Jen), Associate Professor and other evolution and chart of a simple, easy to use and hasty and efficient SVM pattern recognition software package with the flyback,vchj, he not only provides a compilation agreeable series on Windows file system implementation, it likewise provides the source code to assist the improvement, revision and other operating systems applications; the software parameters of SVM involves relatively small adjustment, a lot of default parameters, the use of These default parameters can solve numerous problems; and to provide a cross-validation (Cross Validation) feature. The package can be http://www.csie.ntu.edu.tw/ ~ cjlin / free. The software can solve C-SVM, ν-SVM, ε-SVR and ν-SVR and other issues, including an based on pattern recognition algorithm for multi-class problems.

directory

Overview appended handbook steps in data format accustomed apt use Svmpredict Svmtrain use SVMSCALE usage? Illustrates one overview of essential papers joined SVM for pattern recognition or flyback time, SVM usage and its parameters, the choice of kernel function and its parameters, the present worldwide still has not formed a unified model, the optimal SVM algorithm parameters that can merely be at virtue of experience, tentative contrast, large-scale search, or use packages catered cross-validation feature as optimization. Currently, LIBSVM with Java, Matlab, C #, Ruby, Python, R, Perl, Common LISP, Labview and additional dozens of languages. The maximum usually used namely Matlab, Java, and command-line version. The upon version Lin Zhiren (Lin Chih-Jen) are correlate on the home sheet. Manual LibSVM is the source code and executable files given in 2 ways. If the Windows family of operating systems, can use the archive invested, alternatively you can adjust the compiler; if it is Unix-like systems, you must compile it yourself, the package provides a compiled format, we have an SGI workstation (operating system IRIX6. 5), use the free compiler GNU C + +3.3 compiler. Use the steps LIBSVM using common steps are: 1) by Zhao LIBSVM software pack to prepare the format required by the data set; 2) a easy data zoom; 3) thinking the culling of RBF kernel; 4) using cross-validation choice the best parameters C and g; 5) using the best parameters C and g on the plenary exercising set for exercising for assist vector machine prototype; 6) use the model to test for and prophesying. Data format data format used in the LIBSVM software using training data and test data document format is as with: : : ... which The target is the training data set for classification, it is an identity of certain int (support multiple classes); for flyback, is any real number. is 1-based integer, can be discontinuous; is a real digit, that is, we constantly say that the neutral variables. Test data file name is only used to enumerate the precision or mistake, if it is nameless, fair fill this column with a number, you can quit blank. In the package, also includes a exercise data example: heart_scale, for ease of reference data file format, and train using the software. You can write small programs, their general data format into this format. Which formatdatalibsvm.xls excel file data can lightly be converted to meet the requirements of LIBSVM data format. Svmtrain use Svmtrain (education model) usage: svmtrain [options] training_set_file [model_file] Options: The options available manner that the signification of the following-s svm type: SVM set type (default 0) 0 - C-SVC 1 - v-SVC 2 class SVM 3 - e-SVR 4 - v-SVR-t kernel type: set type of kernel function (default 2) 0 Linear: u'v 1 polynomials: ( r * u'v + coef0) ^ degree 2 RBF function: exp (-r | uv | ^ 2) 3 sigmoid: tanh (r * u'v + coef0)-d degree: set degree in nuclear function (for polynomial kernel) (default 3)-gr (gama): gist function of the gamma function set (for the polynomial / rbf / sigmoid kernel) (default 1 / k)-r coef0: kernel set in coef0 set (for the polynomial / sigmoid kernel function) ((default 0)-c cost: set the C-SVC, e-SVR and v-SVR parameters (loss function) (default 1)-n nu: set the v-SVC , one-class SVM and v-SVR parameters (default 0.5)-pp: Setting e-SVR in the wastage function of p value (default 0.1)-m cachesize: set cache memory size in MB (default 40) - e eps: setting allows the termination criterion (default 0.001)-h shrinking: whether to use heuristics, 0 or 1 (default 1)-wi weight: set the parameters of the premier Ji class C is weight? C (C-SVC in C) (default 1)-vn: n-fold cross-validation mode, n is the digit of fold, which must be greater than or equal to 2 in the k-g discretion means that the input data the number of attributes. option-v randomly split the data for n chapters and calculates cross-validation accuracy and the root average square error. These parameters can be Zhao SVM types and kernel function supports anybody combination of the parameters, if you set the parameters of the function or SVM type does not have an impact, the procedure ambition not approve the contention; if the proper parameters not set correctly, parameters with default values. training_set_file is to conduct the training data set; model_file is produced behind the end of training model files, comprising support vector number of samples, example and support vector parameters must lagrange coefficient; this parameter will be used if you do not set the default file name, you can set your favored file name. Svmpredict use method Svmpredict (using existing models to predict) the use of: svmpredict test_file model_file output_file model_file by svmtrain generated model files; test_file is to portend the data file; Output_file is svmpredict the output file. svm-predict no other discretion items. SVMSCALE usage? data sets scaling intentions to: 1) avoid some of the characteristic value scope is also large while others feature value range is too small; 2) to avoid training in order to calculate the kernel function and the calculation of the production while reasoned by mathematical difficulties. Therefore, the data is commonly scaled to [-1,helpmeprogram.com,1] or [0,1]. Usage: svmscale [-l lower] [-u upper] [-y y_lower y_upper] [-s save_filename] [-r restore_filename] filename (default: lower = -1, upper = 1, no y scaling) one,-l: data limit brand; lower: lower scaled data;-u: data hats apparent; upper: upper limit of the scaled data;-y: whether the target at the same time scale; y_lower for the lower limit, y_upper for the upper limit ;-s save_filename: scaling rules that will retention the file save_filename;-r restore_filename: that the zoom behind loading the rules file restore_filename here zoom; filename: data files to be scaled (necessary to meet the previously depicted format). Scaling rules file can be opened with a txt browser, discern the format: lower upper lval1 uval1 lval2 uval2 with upper and lower them while using the lower and upper set the same signification; index that feature serial number; lval the characteristics lower limit corresponds to the characteristics of the converted value; uval for the upper limit corresponds to the characteristics of the converted value. Scale data sets result in this case the output via a DOS window, of course, the files can be DOS redirection character Use case: 1) svmscale s train3.rangetrain3> train3.scale that the default value (that is, the value of the attribute scale to the [-1,1] scope, the target is not to scale) to scale the data sets train3 operation, the outcome scaling rules file for train3.range, set the zoom zoom save the results of train3.scale. 2) svmscale r train3.rangetest3> test3.scale said load after scaling rules train3.range Zhao its upper and lower limits by the corresponding eigenvalue and the upper and lower limits to the straightforward scaling of the data set test3, save the results for test3. scale. Example svmtrain-s 0 -? C 1000-t 1-g 1-r 1-d 3 data_file training a polynomial kernel (u'v +1) ^ 3 and C = 1000 makeup classifier. svmtrain-s 1-n 0.1-t 2-g 0.5-e 0.00001 data_file in the RBF kernel exp (-0.5 | uv | ^ 2) and termination of allow limits, beneath the conditions of 0.00001, training a?-SVM (? = 0.1) classifier. svmtrain-s 3-p 0.1-t 0-c 10 data_file with straightforward kernel u'v and C = 10 and loss of function? = 0.1 Solving SVM retrace. Output is It’s about time: optimization finished, # iter iterations epsilon = quadratic termination conditions obj =, obj, is that the minimum value of quadratic functions it rho = constant decision to nSV =, the number of important papers on support vector This partition lists Ji treatises on LIBSVM important papers. libsvm: a library for Support Vector Machines A practical guide to SVM category Training and Testing Low-degree Polynomial Data Mappings via Linear SVM Working Set Selection Using Second Order Information for Training Support Vector Machines Further reading: 1

http://www .blog.edu.cn/user2/huangbo929/archives/2007/1686321.shtml

2

http://www.csie.ntu.edu.tw/ ~ cjlin / libsvm /

open Category: Pok machine studying