svmtl
usage: svmtl train [options] trainfile
svmtl classify [options] testfile
svmtl crossval [options] trainfile
svmtl gridsearch [options] trainfile
Please use 'svmtl train --help', etc. for
furter help.
svmtl train
usage: svmtl train [options] trainfile
where options are:
file input output
-----------------
-m, --modelfile filename
filename for model file. Default: training file
name with ".model" extension
training
--------
-dt, --details
0 no details are printed / written to model-file
1 print statistics and write them to model-file
(default)
2 print training infos from all two-class
trainings and write them to model-file
-pt, --print_train_error
0 do not calculate training error (default)
1 calculate the training error (which is just
classification of each training vector) and
print statistics
progress reporting
------------------
-vb, --verbose_level
0 report nothing
1 report parameter tuning
2 report cross validation
3 report multi class svm
4 report two class svm
5 report addtitional training infos
-p, --draw_progress_bar
0 no acsii progress bars
1 draw acsii progress bars (default)
select algorithms
-----------------
-mc, --multi_class_type
one_vs_one multi-class SVM by using the 'one versus one'
approach
one_vs_rest multi-class SVM by using the One versus Rest
approach
-tc, --two_class_type
c_svc Two class SVM using C-SVC algorithm for
training
nu_svc Two class SVM using nu-SVC algorithm for
training
-oc, --one_class_type xxx
no algorithms available
-kf, --kernel_type
kmatrix_rbf cached kernel matrix for radial basis function
kernel: exp(-gamma*|u-v|^2)
kmatrix_linear cached kernel matrix for linear kernel: u'*v
kmatrix_poly cached kernel matrix for polynomial kernel:
(gamma*u'*v + coef0)^degree
kmatrix_sigmoid cached kernel matrix for polynomial kernel:
(gamma*u'*v + coef0)^degree
kmatrix_scaled_rbf
cached kernel matrix for scaled feature vectors
passed to radial basis function kernel:
exp(-gamma*|u-v|^2)
kmatrix_scaled_linear
cached kernel matrix for scaled feature vectors
passed to linear kernel: u'*v
kmatrix_scaled_poly
cached kernel matrix for scaled feature vectors
passed to polynomial kernel: (gamma*u'*v +
coef0)^degree
kmatrix_scaled_sigmoid
cached kernel matrix for scaled feature vectors
passed to polynomial kernel: (gamma*u'*v +
coef0)^degree
rbf radial basis function kernel:
exp(-gamma*|u-v|^2)
linear linear kernel: u'*v
poly polynomial kernel: (gamma*u'*v + coef0)^degree
sigmoid polynomial kernel: (gamma*u'*v + coef0)^degree
svm parameters
--------------
-cs, --cache_size size cache memory size in MB (default 40)
-c, --cost value penalty cost for wrong training vectors in
C-SVC. (default 1)
-e, --epsilon value tolerance of termination criterion (default
0.001)
-n, --nu value parameter nu in nu-SVC. (default 0.5)
-sh, --shrinking
0 don't use the shrinking heuristics
1 use the shrinking heuristics (default)
-w, --weight value weight for positive class samples in two-class
C-SVC. (default 1)
kernel parameters
-----------------
-r, --coef0 value coef0 for polynomial kernel. (default 0)
-r, --coef0 value coef0 for sigmoid kernel. (default 0)
-d, --degree value degree for polynomial kernel. (default 3)
-g, --gamma value gamma for polynomial kernel. (default 1)
-g, --gamma value gamma for rbf-kernel. (default 1.0, or 1/k, when
feature_vector_dim is given
-g, --gamma value gamma for sigmoid kernel. (default 1)
-sa, --scale_algorithm
minmax scale each feature that min becomes -1 and max
becomes +1
stddev scale each feature that mean becomes 0 and
standard deviation becomes 1
-sf, --scale_factor array
array containing scale factors for each feature
-- usually you don't want to specify this
manually
-so, --scale_offset array
array containing offsets for each feature --
usually you don't want to specify this manually
svmtl classify
usage: svmtl classify [options] testfile
where options are:
file input output
-----------------
-m, --modelfile filename
filename of model file. Default: 'svmtl.model'
-o, --outfile filename filename for output file with classification
results. Default: test file name with ".output"
extension
classification
--------------
-dt, --details
0 no details
1 print correct/wrong classifications (requires a
labeled test set) (default)
2 save results from each two-class classifications
to output file
progress reporting
------------------
-vb, --verbose_level
0 report nothing
1 report parameter tuning
2 report cross validation
3 report multi class svm
4 report two class svm
5 report addtitional training infos
-p, --draw_progress_bar
0 no acsii progress bars
1 draw acsii progress bars (default)
svmtl crossval
usage: svmtl crossval [options] trainfile
where options are:
file input output
-----------------
-o, --outfile filename filename for output file with classification
results. Default: test file name with ".output"
extension
cross validation
----------------
-v, --nfold nsubsets number of subsets for cross validation (default
10)
-ss, --shuffle_subsets
0 no shuffling
1 shuffle data before splitting into subsets for
cross validation (default)
-td, --train_details
0 no details
1 print training statistics (default)
-cd, --class_details
0 no details
1 print correct/wrong classifications per class
(default)
progress reporting
------------------
-vb, --verbose_level
0 report nothing
1 report parameter tuning
2 report cross validation
3 report multi class svm
4 report two class svm
5 report addtitional training infos
-p, --draw_progress_bar
0 no acsii progress bars
1 draw acsii progress bars (default)
select algorithms
-----------------
-mc, --multi_class_type
one_vs_one multi-class SVM by using the 'one versus one'
approach
one_vs_rest multi-class SVM by using the One versus Rest
approach
-tc, --two_class_type
c_svc Two class SVM using C-SVC algorithm for
training
nu_svc Two class SVM using nu-SVC algorithm for
training
-oc, --one_class_type xxx
no algorithms available
-kf, --kernel_type
kmatrix_rbf cached kernel matrix for radial basis function
kernel: exp(-gamma*|u-v|^2)
kmatrix_linear cached kernel matrix for linear kernel: u'*v
kmatrix_poly cached kernel matrix for polynomial kernel:
(gamma*u'*v + coef0)^degree
kmatrix_sigmoid cached kernel matrix for polynomial kernel:
(gamma*u'*v + coef0)^degree
kmatrix_scaled_rbf
cached kernel matrix for scaled feature vectors
passed to radial basis function kernel:
exp(-gamma*|u-v|^2)
kmatrix_scaled_linear
cached kernel matrix for scaled feature vectors
passed to linear kernel: u'*v
kmatrix_scaled_poly
cached kernel matrix for scaled feature vectors
passed to polynomial kernel: (gamma*u'*v +
coef0)^degree
kmatrix_scaled_sigmoid
cached kernel matrix for scaled feature vectors
passed to polynomial kernel: (gamma*u'*v +
coef0)^degree
rbf radial basis function kernel:
exp(-gamma*|u-v|^2)
linear linear kernel: u'*v
poly polynomial kernel: (gamma*u'*v + coef0)^degree
sigmoid polynomial kernel: (gamma*u'*v + coef0)^degree
svm parameters
--------------
-cs, --cache_size size cache memory size in MB (default 40)
-c, --cost value penalty cost for wrong training vectors in
C-SVC. (default 1)
-e, --epsilon value tolerance of termination criterion (default
0.001)
-n, --nu value parameter nu in nu-SVC. (default 0.5)
-sh, --shrinking
0 don't use the shrinking heuristics
1 use the shrinking heuristics (default)
-w, --weight value weight for positive class samples in two-class
C-SVC. (default 1)
kernel parameters
-----------------
-r, --coef0 value coef0 for polynomial kernel. (default 0)
-r, --coef0 value coef0 for sigmoid kernel. (default 0)
-d, --degree value degree for polynomial kernel. (default 3)
-g, --gamma value gamma for polynomial kernel. (default 1)
-g, --gamma value gamma for rbf-kernel. (default 1.0, or 1/k, when
feature_vector_dim is given
-g, --gamma value gamma for sigmoid kernel. (default 1)
-sa, --scale_algorithm
minmax scale each feature that min becomes -1 and max
becomes +1
stddev scale each feature that mean becomes 0 and
standard deviation becomes 1
-sf, --scale_factor array
array containing scale factors for each feature
-- usually you don't want to specify this
manually
-so, --scale_offset array
array containing offsets for each feature --
usually you don't want to specify this manually
svmtl gridsearch
usage: svmtl gridsearch [options] trainfile
where options are:
file input output
-----------------
grid search
-----------
-p1, --param1
<keyname>:<from>,add<step>,<to>
specify linear range for row axis. e.g.,
'cost:1,add1,5' results in 1,2,3,4,5
<keyname>:<from>,mul<factor>,<to>
specify logarithmic range for row axis. e.g.,
'gamma:1,mul2,16' results in 1,2,4,8,16
<keyname>:<v1>,<v2>,<v3>,...
specify list of values for rowaxis. e.g.
'cost:-2,5,42,3'
-p2, --param2 <keyname>:<valuespec>
range and parameter for column axis -- syntax of
valuespec is the same as for 'param1'
-pg, --print_grid
0 no output during evaluation
1 print the grid during evaluation (default)
cross validation
----------------
-v, --nfold nsubsets number of subsets for cross validation (default
10)
-ss, --shuffle_subsets
0 no shuffling
1 shuffle data before splitting into subsets for
cross validation (default)
-td, --train_details
0 no details
1 print training statistics (default)
-cd, --class_details
0 no details
1 print correct/wrong classifications per class
(default)
progress reporting
------------------
-vb, --verbose_level
0 report nothing
1 report parameter tuning
2 report cross validation
3 report multi class svm
4 report two class svm
5 report addtitional training infos
-p, --draw_progress_bar
0 no acsii progress bars
1 draw acsii progress bars (default)
select algorithms
-----------------
-mc, --multi_class_type
one_vs_one multi-class SVM by using the 'one versus one'
approach
one_vs_rest multi-class SVM by using the One versus Rest
approach
-tc, --two_class_type
c_svc Two class SVM using C-SVC algorithm for
training
nu_svc Two class SVM using nu-SVC algorithm for
training
-oc, --one_class_type xxx
no algorithms available
-kf, --kernel_type
kmatrix_rbf cached kernel matrix for radial basis function
kernel: exp(-gamma*|u-v|^2)
kmatrix_linear cached kernel matrix for linear kernel: u'*v
kmatrix_poly cached kernel matrix for polynomial kernel:
(gamma*u'*v + coef0)^degree
kmatrix_sigmoid cached kernel matrix for polynomial kernel:
(gamma*u'*v + coef0)^degree
kmatrix_scaled_rbf
cached kernel matrix for scaled feature vectors
passed to radial basis function kernel:
exp(-gamma*|u-v|^2)
kmatrix_scaled_linear
cached kernel matrix for scaled feature vectors
passed to linear kernel: u'*v
kmatrix_scaled_poly
cached kernel matrix for scaled feature vectors
passed to polynomial kernel: (gamma*u'*v +
coef0)^degree
kmatrix_scaled_sigmoid
cached kernel matrix for scaled feature vectors
passed to polynomial kernel: (gamma*u'*v +
coef0)^degree
rbf radial basis function kernel:
exp(-gamma*|u-v|^2)
linear linear kernel: u'*v
poly polynomial kernel: (gamma*u'*v + coef0)^degree
sigmoid polynomial kernel: (gamma*u'*v + coef0)^degree
svm parameters
--------------
-cs, --cache_size size cache memory size in MB (default 40)
-c, --cost value penalty cost for wrong training vectors in
C-SVC. (default 1)
-e, --epsilon value tolerance of termination criterion (default
0.001)
-n, --nu value parameter nu in nu-SVC. (default 0.5)
-sh, --shrinking
0 don't use the shrinking heuristics
1 use the shrinking heuristics (default)
-w, --weight value weight for positive class samples in two-class
C-SVC. (default 1)
kernel parameters
-----------------
-r, --coef0 value coef0 for polynomial kernel. (default 0)
-r, --coef0 value coef0 for sigmoid kernel. (default 0)
-d, --degree value degree for polynomial kernel. (default 3)
-g, --gamma value gamma for polynomial kernel. (default 1)
-g, --gamma value gamma for rbf-kernel. (default 1.0, or 1/k, when
feature_vector_dim is given
-g, --gamma value gamma for sigmoid kernel. (default 1)
-sa, --scale_algorithm
minmax scale each feature that min becomes -1 and max
becomes +1
stddev scale each feature that mean becomes 0 and
standard deviation becomes 1
-sf, --scale_factor array
array containing scale factors for each feature
-- usually you don't want to specify this
manually
-so, --scale_offset array
array containing offsets for each feature --
usually you don't want to specify this manually