1.04: 2000/6/17, add "load" button to svm-toy. README 

file updated.

2.0: 2000/8, major updates. Include nu-svm, one-class svm, and svr

2.01: 2000/9/22, correct the calculation of obj value and number of bounded support vectors

2.02: 2000/9/29, replace b^2/2 newsvm to regular nu svm.

2.03: 2000/10/24 some improvements on the computational speed

2.1: 2000/12/19 Java version included, regression demonstrated in svm-toy

2.2: 2001/1/16 multi-class classification, nu-SVR, remove epsilon_a

2.3: 2001/3/15 c+-, cross validation, fix some minor bugs

2.31: 2001/4/12 fix one bug on one-class SVM, use float for Cache

2.32: 2001/9/23 

	1. max line number of svm-scale now dynamic

	2. gcc 3.0 problem: now g++ always used

	3. java code in a "libsvm" package

	4. avoid a problem when compiled with Borland C++ builder

2.33: 2001/12/1

	Python interface added

2.34: 2002/6/15

	Add the subroutine svm_check_parameter in svm.cpp

	--> better error handling

	fix bug of python interface for handling different weights

	fix bug of cross validation in svm-train.c

2.35: 2002/6/16

	libsvm.jar was not compiled correctly in 2.34

2.36: 2002/8/4

	grid.py added: contour plot of CV accuracy

	fix several bugs

2.4:  2003/4/1

	svm.cpp

		non-psd kernel using max(...,0) in svm.cpp

	python interface

		python interface bug (nu-svm)

	grid.py

		-log2c and -log2c for grid.py

	        output current best

		coarse grid as default (2)

		ssh for grid.py 

	improvements of scaling 

2.5: 2003/11/13

	subroutines for accessing decision values and number of labels. 

		for svm.cpp, java, and  python interface

	fix bug of svm-scale.c (about -r and -s factors)

	use fscanf but not sscanf in svm-predict.c (faster)

	makefile for windows

	add "using namespace std;" in some .java files

	improve easy.py: output cv rate, error messages printed

		better checking

	better python interface example svm_test.py

	and some minor updates

2.6: 2004/04/01

	Probability estimates for classification/regression 

	Python interface: use swig 1.3 instead of 1.1

	Cross validation becomes a library subroutine

	A few minor corrections: (not completely listed)

		more interface functions such as getting svm_type

		print nu only when Cp=Cn

		floor division in python interface

2.7: 2004/11/10

	Stratified cross validation

	Better faq

	Scaling: support storing the factor of y

	A few minor updates:

		class QMatrix added

		improve the use of easy.py and grid.py on windows 

		grid.py: same CV and same g: use smaller C

		sparse input for python interface

		working set selection: < to <=

2.71: 2004/11/20

	fix a java bug introduced from 2.6 to 2.7

2.8: 2005/04/01

	new working set selection via second order information

	fix minor changes/corrections:

		problem when cache size less than two kernel columns

		-v #data -> stratified CV is not loo -> ensure loo is done

		problem of typing "svm-train -c" only

		problem of "svm-train -n 1 -s 3 ..."

		python interface makefile: -fPIC for 32/64bit

		color change in svmtoy

		makefile in building QT svmtoy

2.81: 2005/11/20

	add a python script subset.py for subsampling

	slightly modify the working set so it's exact the same as the paper

	default cache size to 100 MB

2.82: 2006/04/01

	precomputed kernel

	directly implement a fast powi() function

	poly degree double to int

	minor corrections:

		java code for CV

2.83: 2006/11/17

	Fix the bug of -t 3

	better checking load/save in svm-train.c/svm-predict.c/svm.cpp

	remove redundant var pos in svm_predict_values (thanks to Albert Strasheim)

	Better descriptions in README for the precomputed kernel

2.84: 2007/04/01

	Improve the shrinking code: faster for some cases

	Code more 64-bit friendly: allow large -m

	In Solver, b is replaced by p

	Subroutine max_violating_pair removed. Things are directly

	written in do_shrinking().

	Modify do_shrinking() so variable names are the same as libsvm

	document

2.85: 2007/11/6

	fix minor memory leak in svm-predict.c

	add tools/checkdata.py 

	java to 1.5

	Makefile: CXX?

	Makefile.win: avoid warning from .net 2005

	avoid warning of gcc 4.2

	sigma 0.001 to 1e-12; in Newton direction of prob output

	2 * (long int) l avoid warning of old Visual C++

	grid.py: avoid gnuplot problem on windows

2.86: 2008/04/01

	svm-scale for java

	version number in svm.h and svm.m4

	rename svmtrain.exe to svm-train.exe 

	python: while 1 --> while True, Popen -> call

	show best parameters on the contour of grid.py

2.87: 2008/10/13

	svm-toy/qt updated to qt4 from qt3

	fix a bug in svm-scale.c

		max feature index of -r file is considered

	Makefile: add make lib; add -Wconversion and -fPIC in Makefile

	Add "rb" in load_model of svm.cpp

	Simplify do_shrinking of svm.cpp

	Change the order of loops in reconstrict_gradient of svm.cpp

		save the number of kernel evaluations

	Add python/setup.py

2.88: 2008/10/30

	better gradient reconstructions

	issue a warning	when -h 0 may be faster

2.89: 2009/04/01

	reduce input/loading time of svm-train/svm-predict by half

	pointer function so users can specify their own outputs

	remove info_flush() 

	a extern variable libsvm_version

	svm-train -q option (disable outputs)

	svm-scale: warning if more nonzero produced

	easy.py: popel.communiate() to avoid some deadlock (if lots of

		outputs when #classes is large)

2.9: 2009/11/1

	tools/*.py can be run under python 3.0

	svm_set_quiet() in python interface to disable outputs

	check gamma < 0

	internal functions to be static

2.91: 2010/04/01

	completely new python interface using ctype

	new way to set the print_string function

	Java: able to load model from a BufferedReader directly

	fix grid.py so -log2c can be run under python 2.6 or after

3.0: 2010/09/13

	Move model structure to svm.h

	Two functions for freeing a model (content or the whole model)

	QD from Qfloat to double (better precision because SSE on 64-bit machines less accurate than i387 on 32-bit

	exit status for checkdata.py 

	old python interface (swig) is removed

3.1: 2011/04/01

	MATLAB interface:

		Merge matlab interface to core libsvm

		Using mexPrintf() when calling info() in MATLAB interface.

		Both 32- and 64-bit windows binary files are provided

	Java:

		Math.random is replaced by Random in java interface

	Python interface:

i		subroutines to get SVs 

		relative path to load *.dll and *.so

	svm.cpp: 

		null pointer check before release memory in svm_free_model_content()

		svm_destroy_model() no longer supported.

	svm-train.c and svm-predict.c

		Better format check in reading data labels

	svm-toy:

		fix the svm_toy dialog path

	tools:

		Using new string formatting/encoding in tools/*.py

		clearer png output, fix grid.py legend

3.11: 2011/11/5

	Set max number of iterations in the main loop of solvers

	matlab:

		new make.m for unix/mac/windows and for matlab/octave

	matlab and python:

		fix a problem that decision values returned by svmpredict is empty if number of classes = 1

3.12: 2012/04/01

      svm-toy: support loading/saving of regression data

      python interface: handle the issue of the "0th" feature if using lists

      tools/grid.py: not redrawing contour if c,g, cv doesn't change

      add setlocale when saving and loading model so not affected by users' locale

3.13: 2012/11/06

      svm.c and svm.h:

            add sv_indices in model structure, so users can know which training instances are SVs

	    two library funs svm_get_sv_indices and svm_get_nr_sv are added

	    max_iter warning moved to stderr so -q won't disable it

      svm-train.c:

            usage modified to stress that multiclass is supported

      svm-predict.c:

            add -q for svm-predict

      svm-scale.c:

	    issue a warning if feature indices do not start from 1

	    issue a warning for inconsistency between scaling-factor file and input file

      tools:

	    subset.py is written to be much faster

3.14: 2012/11/16

      fix a bug: we didn't re-compile windows binary after updating svm.def

3.15: 2013/1/27

      fix the bug of not freeing sv_indices

      tools/grid.py:

             -null option: allow the search on C or g only

	     -resume option: resume tasks from an earlier run

	     can be called as a python module            	

      python interface:

      	     local package searched first

	     libsvm options can be str or list

	     param.show() becomes print(param)

3.16: 2013/1/27

      tools/:

             easy.py fails in 3.15. Fix it by modifying grid.py

3.17: 2013/04/01

      svm.cpp:

	     if class labels are 1 and -1, ensure labels[0] = 1 and labels[1] = -1

      	     initialize model->sv_indices as null in svm_load_model

	     if nr_fold > # data, change nr_fold to be # data and ro leave-one-out cv 

      matlab interface: 

      	     handle the problem where output variables are not specified

3.18: 2014/04/01

      svm.cpp and svm-scale.c:

      	     check return values of fscanf

      matlab interface:

      	     Makefile no longer handles octave because make.m should be used

3.19: 2014/10/27

      svm.cpp:

             use size_t rather than long int in some places for 64-bit windows support

      matlab interface:

             use size_t and mwIndex to avoid type conversions

      python interface:

             use __all__

3.20: 2014/11/15

      matlab interface:

             fix a bug in libsvmwrite.c introduced in 3.19 (cannot run on windows)

3.21: 2015/12/14

      pre-built windows exe files changed from 32 to 64 bit

      matlab interface:

             now use #include "../svm.h"

             fix some minor issues in make.m of matlab interface

3.22: 2016/12/22

      probability output:

             if 2 classes, directly output the predited probabilities

	     rather than run the iterative algorithms for multi-class situations

      java:

             libsvm.jar generated by java 1.7 rather than 1.5

             change the use of "_" in svm.java, which won't be allowed in later java

3.23: 2018/7/15

      add more digits of predicted file, model file, scaled data and data from matlab libsvmwrite: to %.17g

      revise svm-scale.c so features in test data that do not appear in training data are scaled to zero.

      remove unnecessary tab or space in all files

      python interface:

            add Scipy support

	    add scaling (csr_find_scale_param and csr_scale functions)

	    put some utility functions identical in LIBLINEAR to commonutil.py.

	    functions for scaling are put in commonutil.py.

	    sort column indices of csr matrix before training as feature indices must be ascending

      convert Qt version of svm-toy from Qt4 to Qt5

      remove gtk svm-toy because we stop maintaining this tool

      minor improvement of descriptions in README