Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0: svm_light_faq
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0: How do I compile on a Windows PC?
Daniel@0:
Daniel@0: - The following instructions apply to SVMlight, SVMstruct, SVMperf, SVMcfg, SVMmulticlass,
Daniel@0: and SVMhmm.
Daniel@0: - The easiest is to install CYGWIN and
Daniel@0: use the gcc compiler that comes with the Cygwin distribution. Just open a
Daniel@0: cygwin command window, change to the directory that contains file "Makefile" and type
Daniel@0:
Daniel@0: make
Daniel@0: at the command prompt. This creates executables that run under Cygwin.
Daniel@0: - If you want execuables that do not need Cygwin to run, follow the same
Daniel@0: instructions as above, but make sure that you download the MinGW package as
Daniel@0: part of Cygwin. Then use the command
Daniel@0: make 'SFLAGS=-mno-cygwin'
Daniel@0: to create the executables.
Daniel@0: - You can also use the Visual C compiler, but it will be much more of a
Daniel@0: hassle than using Cygwin. To use Visual C, you need to build the appropriate project file according to the provided "Makefile".
Daniel@0: Make sure you do not include the file "svm_loqo.c" in your project, unless you
Daniel@0: want to use the PL_LOQO optimizer instead of the built-in optimizer.
Daniel@0:
Daniel@0:
Daniel@0: How do I compile on a PowerMac using Code Warrior?
Daniel@0:
Daniel@0:
Daniel@0: - This applies to SVMlight, but should also work for SVMstruct, SVMperf, SVMcfg, SVMmulticlass,
Daniel@0: and SVMhmm.
Daniel@0: - You need to modify the source code a little (as suggested by Jewgeni Starikow). Use #include "console.h" to emulate a UNIX shell. Then add argc=ccommand(&argv) as the first instruction of each main().
Daniel@0: - CPU-timing might cause another problem. The timing routines are used to calculated the runtime of the program. If you do not need this feature, remove the body of get_runtime() in svm_common.c. Otherwise replace the body with the appropriate Mac routines from 'time.h'.
Daniel@0: How do I integrate SVMlight into C++ code?
Daniel@0:
Daniel@0:
Daniel@0: - Compile "svm_learn.c", "svm_common.c", and "svm_hideo.c" as C code.
Daniel@0: - The C++ program you want to call svm_learn/8 and classify_example/2 (or classify_example_linear/2) from needs to include the following headers:
Daniel@0: extern "C" {
Daniel@0: # include "svm_common.h"
Daniel@0: # include "svm_learn.h"
Daniel@0: }
Daniel@0: - Link "svm_learn.o", "svm_common.o", and "svm_hideo.o" to your program.
Daniel@0:
Daniel@0: Is there an option for doing multi-class classification?
Daniel@0:
Daniel@0: - Not in SVMlight. If you do not use kernel, you can
Daniel@0: use SVMmulticlass to
Daniel@0: do multi-class classification. Otherwise, I recommend to split the task
Daniel@0: into multiple binary classification tasks for multi-class or multi-label classification. For
Daniel@0: example, you can do one-class-against-the-rest classification, or pairwise classification. Refer to Chapter 2 in [Joachims,
Daniel@0: 2002a].
Daniel@0:
Daniel@0:
Daniel@0: Is there an option for doing cross-validation?
Daniel@0:
Daniel@0: - Yes, there is such an option in SVMlight, but not in SVMstruct.
Daniel@0: By setting the option "-x 1", SVMlight computes the
Daniel@0: leave-one-out estimates of the prediction error, recall, precision, and F1. By
Daniel@0: default, it also outputs the XiAlpha-Estimates of these quantities. They are
Daniel@0: strictly conservative approximations to the leave-one-out (i.e. they never
Daniel@0: underestimate the leave-one-out error). Refer to Chapter 5 in [Joachims,
Daniel@0: 2002a].
Daniel@0:
Daniel@0: How can I output the dual variables at the solution?
Daniel@0:
Daniel@0: - Use the option -a <filename>. The file contains the value of the
Daniel@0: dual variables (multiplied by the label y) for each training example in the
Daniel@0: order in which the examples appeared in the training file.
Daniel@0:
Daniel@0: How can I get the weight vector of the hyperplane for a linear SVM?
Daniel@0:
Daniel@0: - There is a small PERL script to
Daniel@0: compute the weight vector (similar script in
Daniel@0: PYTHON by Ori Cohen) based on the model file output by svm_learn
Daniel@0: and svm_perf_learn.
Daniel@0: Of course, it works only in the linear case and not for kernels. All the
Daniel@0: script does is compute the weighted sum of the support vectors (first
Daniel@0: element in line is alpha*y, what follows is the feature vector). For
Daniel@0: further info, see the comment in the model file for its format.
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0: What is the format of the learned model file?
Daniel@0:
Daniel@0: - The model files produced by svm_learn and svm_perf_learn are in ASCII and
Daniel@0: easy to read. The first few lines contain parameters. The comments in the
Daniel@0: model file are self-explanatory. The last of these is the threshold b of the
Daniel@0: learned hyperplane classifier sign(w*x - b). The following lines each contain
Daniel@0: a support vector, and the first element of the line is its coefficient (alpha_i
Daniel@0: * y_i) in the kernel expansion. The support vectors are listed in random
Daniel@0: order. Note that the term support vector has a different meaning in SVMperf
Daniel@0: compared to SVMlight.
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0: How can I implement my own Kernel?
Daniel@0:
Daniel@0: - You can write your own kernel by extending the function "double
Daniel@0: custom_kernel(KERNEL_PARM *kernel_parm, SVECTOR *a, SVECTOR *b)" in "kernel.h"
Daniel@0: and then select it via the "-t 4" option. The "a" and "b" are pointers to the
Daniel@0: examples you are computing the kernel for.
Daniel@0:
Daniel@0: - If the data you are working with is vectorial, then you can use the
Daniel@0: existing sparse vector operations defined in svm_common.c. For example,
Daniel@0: pow(sprod_ss(a->words,b->words),2.0)
Daniel@0: implements the homogeneous polynomial kernel of degree 2.
Daniel@0: - If the data is non-vectorial (e.g. strings), then you can use the
Daniel@0: following functionality. In the training and test data files, whatever is
Daniel@0: written behind the "#" is copied into the "userdefined" field of the
Daniel@0: internal datastructure "SVECTOR" as a string. So, for the line
Daniel@0: +1 #abcdefg
Daniel@0: you will find "abcdefg" in "svector->userdefined".
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0: Error messages and known problems:
Daniel@0:
Daniel@0:
Daniel@0: - ERROR: terminating optimizer - choldc failed, matrix not positive definite
Daniel@0:
Daniel@0:
Daniel@0: - If the program terminates after this message, get the lastest version of PR_LOQO and SVM-Light V2.01 (or later).
Daniel@0: - It the program continues after this error message, don't worry :-)
Daniel@0:
Daniel@0: - The CPU-time is negative or looks bogus:
Daniel@0:
Daniel@0:
Daniel@0: - To be compatible with Windows, I used the clock/0 function to measure CPU-time. This function returns a long integer and counts microseconds, so that it wraps around after about 30 minutes. You could use timing routines more appropriate for your system. For example on Solaris, you might want to use the routines from 'sys/times.h'.
Daniel@0:
Daniel@0:
Daniel@0: The program hangs when ...
Daniel@0:
Daniel@0:
Daniel@0: - ... reading in the examples.
Daniel@0:
Daniel@0:
Daniel@0: - Get version 3.02 or later.
Daniel@0:
Daniel@0:
Daniel@0: Convergence during learning is very slow!
Daniel@0:
Daniel@0:
Daniel@0: - In verbose mode 2 I observe that max violation bounces around and does not really converge.
Daniel@0:
Daniel@0:
Daniel@0: - Use a smaller value for the option -n (default n=q). This makes sure that only n new variables enter the working set in each iteration. This can prevent zig-zagging behavior.
Daniel@0: - Use a smaller or larger value for the size of the working set (option -q). Reasonable values are in the range [2:50].
Daniel@0: - You might be using an excessively large value of C (option -c) in relation to your data. The value of C should typically be less than 1000 times 1.0/(the maximum squared Euclidian length of your feature vectors).
Daniel@0: - You have weired data and the convergence simply IS very slow J
Daniel@0: . Sorry, not much you can do about it.
Daniel@0:
Daniel@0: - Nearly all my training examples end up as support vectors.
Daniel@0:
Daniel@0:
Daniel@0: - Use a "stiffer" kernel (e g. a lower value of gamma for the RBF-kernel). If you pick a kernel which is very far away from the optimum, you will not get good generalization performance anyway.
Daniel@0: - Your data is really difficult to separate. Think about a better representation of your data. It is a bad idea - for example - if different features have values in very much different orders of magnitude. You might want to normalize all features to the range [-1,+1].
Daniel@0:
Daniel@0:
Daniel@0: It does not converge!
Daniel@0:
Daniel@0:
Daniel@0: - If you are using the built-in HIDEO optimizer, get the version 3.50 or later. There used to be problems for low dimensional data sets for the old HIDEO optimizer.
Daniel@0: - It should always converge, unless there are numerical problems :-)
Daniel@0: - Numerical problems.
Daniel@0:
Daniel@0:
Daniel@0: - Make sure your data is properly scaled. It is a bad idea - for example - if different features have values in different orders of magnitude. You might want to normalize all features to the range [-1,+1], especially for problems with more than 100 features. Or even better, normalize all feature vectors to Euclidian length 1.
Daniel@0:
Daniel@0:
Daniel@0: It crashes!
Daniel@0:
Daniel@0: - It crashes during learning when using a non-linear kernel.
Daniel@0:
Daniel@0: - Get version 5.00 or later. There was an initialization bug in the
Daniel@0: kernel cache.
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0: The results look bogus!
Daniel@0:
Daniel@0: - When doing transductive learning, for some training set sizes it does
Daniel@0: not obey the -p parameter and the results look bogus.
Daniel@0:
Daniel@0: - Get version 5.00 or later. There was a bug introduced in version 4.00.
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0:
Daniel@0: Last modified July 1st, 2007 by
Thorsten Joachims <thorsten@joachims.org>
Daniel@0:
