7.1.220. 3dsvm

Link to classic view

Program: 3dsvm

+++++++++++ 3dsvm: support vector machine analysis of brain data +++++++++++

3dsvm - temporally predictive modeling with the support vector machine

This program provides the ability to perform support vector machine (SVM) learning on AFNI datasets using the SVM-light package (version 5) developed by Thorsten Joachims (http://svmlight.joachims.org/).
3dsvm [options]

Examples:

  1. Training: basic options require a training run, category (class) labels for each timepoint, and an output model. In general, it usually makes sense to include a mask file to exclude at least non-brain voxels
3dsvm -trainvol run1+orig
-trainlabels run1_categories.1D
-mask mask+orig
-model model_run1
  1. Training: obtain model alphas (a_run1.1D) and model weights (fim: run1_fim+orig)
3dsvm -alpha a_run1
-trainvol run1+orig -trainlabels run1_categories.1D -mask mask+orig -model model_run1
-bucket run1_fim
  1. Training: exclude some time points using a censor file

    3dsvm -alpha a_run1

    -trainvol run1+orig -trainlabels run1_categories.1D

-censor censor.1D -mask mask+orig -model model_run1 -bucket run1_fim
  1. Training: control svm model complexity (C value)
3dsvm -c 100.0
-alpha a_run1 -trainvol run1+orig -trainlabels run1_categories.1D -censor censor.1D -mask mask+orig -model model_run1 -bucket run1_fim
  1. Training: using a kernel

    3dsvm -c 100.0

-kernel polynomial -d 2 -alpha a_run1 -trainvol run1+orig -trainlabels run1_categories.1D -censor censor.1D -mask mask+orig -model model_run1
  1. Training: using regression
3dsvm -type regression
-c 100.0
-e 0.001 -alpha a_run1 -trainvol run1+orig -trainlabels run1_categories.1D -censor censor.1D -mask mask+orig -model model_run1
  1. Testing: basic options require a testing run, a model, and an output predictions file
3dsvm -testvol run2+orig
-model model_run1+orig
-predictions pred2_model1
  1. Testing: compare predictions with ‘truth’

    3dsvm -testvol run2+orig

    -model model_run1+orig

-testlabels run2_categories.1D -predictions pred2_model1
  1. Testing: use -classout to output integer thresholded class predictions (rather than continuous valued output)

    3dsvm -classout

    -testvol run2+orig -model model_run1+orig -testlabels run2_categories.1D -predictions pred2_model1

options:

——————- TRAINING OPTIONS ——————————————- -type tname Specify tname:

classification [default] regression

to select between classification or regression.

-trainvol trnname A 3D+t AFNI brik dataset to be used for training.

-mask mname Specify a mask dataset to only perform the analysis
on non-zero mask voxels.
++ If ‘-mask’ is not used ‘-nomodelmask must be
specified.

For example, a mask of the whole brain can be generated by using 3dAutomask, or more specific ROIs could be generated with the Draw Dataset plugin or converted from a thresholded functional dataset. The mask is specified during training but is also considered part of the model output and is automatically applied to test data.

-nomodelmask Flag to enable the omission of a mask file. This is required if ‘-mask’ is not used.
-trainlabels lname lname = filename of class category .1D labels

corresponding to the stimulus paradigm for the training data set. The number of labels in the selected file must be equal to the number of time points in the training dataset. The labels must be arranged in a single column, and they can be any of the following values:

0 - class 0 1 - class 1 n - class n (where n is a positive integer) 9999 - censor this point

See also -censor.

-censor cname Specify a .1D censor file that allows the user
to ignore certain samples in the training data. To ignore a specific sample, put a 0 in the row corresponding to the time sample - i.e., to ignore sample t, place a 0 in row t of the file. All samples that are to be included for training must have a 1 in the corresponding row. If no censor file is specified, all samples will be used for training. Note the lname file specified by trainlabels can also be used to censor time points (see -trainlabels).

-kernel kfunc kfunc = string specifying type of kernel function:

linear : <u,v> [default] polynomial : (s<u,v> + r)^d rbf : radial basis function

exp(-gamma ||u-v||^2)

sigmoid : tanh(s <u,v> + r))

note: kernel parameters use SVM-light syntax:

-d\ int : d parameter in polyniomial kernel
3 [default]
-g\ float : gamma parameter in rbf kernel
1.0 [default]
-s\ float : s parameter in sigmoid/poly kernel
1.0 [default]
-r\ float : r parameter in sigmoid/poly kernel
1.0 [default]
-max_iterations int Specify the maximum number of iterations for the
optimization. 1 million [default].

-alpha aname Write the alphas to aname.1D

-wout Flag to output sum of weighted linear support
vectors to the bucket file. This is one means of generating an “activation map” from linear kernel SVMs see (LaConte et al., 2005). NOTE: this is currently not required since it is the only output option.
-bucket bprefix Currently only outputs the sum of weighted linear
support vectors written out to a functional (fim) brik file. This is one means of generating an “activation map” from linear kernel SVMS (see LaConte et al, 2005).

——————- TRAINING AND TESTING MUST SPECIFY MODNAME —————— -model modname modname = basename for the model brik.

Training: modname is the basename for the output brik containing the SVM model

3dsvm -trainvol run1+orig
-trainlabels run1_categories.1D -mask mask+orig -model model_run1

Testing: modname is the name for the input brik containing the SVM model.

3dsvm -testvol run2+orig
-model model_run1+orig -predictions pred2_model1

——————- TESTING OPTIONS ——————————————– -testvol tstname A 3D or 3D+t AFNI brik dataset to be used for testing.

A major assumption is that the training and testing volumes are aligned, and that voxels are of same number, volume, etc.
-predictions pname pname = basename for .1D prediction file(s).
Prediction files contain a single column, where each line holds the predicted value for the corresponding volume in the test dataset. By default, the predicted values take on a continuous range; to output integer-valued class decision values use the -classout flag. For classification: Values bellow 0.5 correspond to (class A) and values above 0.5 to (class B), where A < B. For more than two classes a separate prediction file for each possible pair of training classes and one additional “overall” file containing the predicted (integer-valued) class membership is generated. For regression: Each value is the predicted parametric rate for the corresponding volume in the test dataset.
-classout Flag to specify that pname files should be integer- valued, corresponding to class category decisions.
-nopredcensored Do not write predicted values for censored time-points
to prediction file
-nodetrend Flag to specify that pname files should not be
linearly de-trended (detrend is the current default).
-testlabels tlname tlname = filename of ‘true’ class category .1D labels
for the test dataset. It is used to calculate the prediction accuracy performance of SVM classification. If this option is not specified, then performance calculations are not made. Format is the same as lname specified for -trainlabels.
-multiclass mctype mctype specifies the multiclass algorithm for

classification. Current implementations use 1-vs-1 two-class SVM models.

mctype must be one of the following:

DAG : Directed Acyclic Graph [default] vote : Max Wins from votes of all 1-vs-1 models

see http:\lacontelab.org3dsvm.html for details and references.

——————- INFORMATION OPTIONS —————————————

-help this help

-version print version history including rough description
of changes

——————– SVM-light learn help —————————–

SVM-light V5.00: Support Vector Machine, learning module 30.06.02stim

Copyright: Thorsten Joachims, thorsten@ls8.cs.uni-dortmund.de

This software is available for non-commercial use only. It must not be modified and distributed without prior permission of the author. The author is not responsible for implications from the use of this software.

usage: svm_learn [options] example_file model_file
Arguments:
example_file-> file with training data model_file -> file to store learned decision rule in
General options:
-? -> this help
-v [0..3] -> level (default 1)

Learning options:

-z {c,r,p} -> select between classification (c), regression (r),
and preference ranking (p) (default classification)
-c float -> C: trade-off between training error and margin (default [avg. x*x]^-1)
-w [0..] -> epsilon width of tube for regression
(default 0.1)
-j float -> Cost: cost-factor, by which training errors on
positive examples outweight errors on negative examples (default 1) (see [4])
-b [0,1] -> use biased hyperplane (i.e. x*w+b>0) instead
of unbiased hyperplane (i.e. x*w>0) (default 1)
-i [0,1] -> remove inconsistent training examples
and retrain (default 0)

Performance estimation options:

-x [0,1] -> compute leave-one-out estimates (default 0)
(see [5])
-o ]0..2] -> value of rho for XiAlpha-estimator and for pruning
leave-one-out computation (default 1.0) (see [2])
-k [0..100] -> search depth for extended XiAlpha-estimator
(default 0)

Transduction options (see [3]):

-p [0..1] -> fraction of unlabeled examples to be classified
into the positive class (default is the ratio of positive and negative examples in the training data)

Kernel options:

-t int -> type of kernel function:
0: linear (default) 1: polynomial (s a*b+c)^d 2: radial basis function exp(-gamma ||a-b||^2) 3: sigmoid tanh(s a*b + c) 4: user defined kernel from kernel.h
-d int -> parameter d in polynomial kernel
-g float -> parameter gamma in rbf kernel
-s float -> parameter s in sigmoid/poly kernel
-r float -> parameter c in sigmoid/poly kernel
-u string -> parameter of user defined kernel

Optimization options (see [1]):

-q [2..] -> maximum size of QP-subproblems (default 10)
-n [2..q] -> number of new variables entering the working set
in each iteration (default n = q). Set n<q to prevent zig-zagging.
-m [5..] -> size of cache for kernel evaluations in MB (default 40)
The larger the faster...
-e float -> eps: Allow that error for termination criterion [y [w*x+b] - 1] >= eps (default 0.001)
-h [5..] -> number of iterations a variable needs to be
optimal before considered for shrinking (default 100)
-f [0,1] -> do final optimality check for variables removed
by shrinking. Although this test is usually positive, there is no guarantee that the optimum was found if the test is omitted. (default 1)

Output options:

-l string -> file to write predicted labels of unlabeled
examples into after transductive learning
-a string -> write all alphas to this file after learning
(in the same order as in the training set)

More details in: [1] T. Joachims, Making Large-Scale SVM Learning Practical. Advances in

Kernel Methods - Support Vector Learning, B. Schoelkopf and C. Burges and A. Smola (ed.), MIT Press, 1999.
[2] T. Joachims, Estimating the Generalization performance of an SVM
Efficiently. International Conference on Machine Learning (ICML), 2000.
[3] T. Joachims, Transductive Inference for Text Classification using Support
Vector Machines. International Conference on Machine Learning (ICML), 1999.
[4] K. Morik, P. Brockhausen, and T. Joachims, Combining statistical learning
with a knowledge-based approach - A case study in intensive care monitoring. International Conference on Machine Learning (ICML), 1999.
[5] T. Joachims, Learning to Classify Text Using Support Vector
Machines: Methods, Theory, and Algorithms. Dissertation, Kluwer, 2002.

——————– SVM-light classify help —————————–

SVM-light V5.00: Support Vector Machine, classification module 30.06.02

Copyright: Thorsten Joachims, thorsten@ls8.cs.uni-dortmund.de

This software is available for non-commercial use only. It must not be modified and distributed without prior permission of the author. The author is not responsible for implications from the use of this software.

usage: svm_classify [options] example_file model_file output_file
options: -h -> this help

-v [0..3] -> verbosity level (default 2) -f [0,1] -> 0: old output format of V1.0

-> 1: output the value of decision function (default)
Jeff W. Prescott, William A. Curtis, Ziad Saad, Rick Reynolds, R. Cameron Craddock, Jonathan M. Lisinski, and Stephen M. LaConte

Original version written by JP and SL, August 2006 Released to general public, July 2007

Questions/Comments/Bugs - email slaconte@vtc.vt.edu

Reference: LaConte, S., Strother, S., Cherkassky, V. and Hu, X. 2005. Support vector

machines for temporal classification of block design fMRI data. NeuroImage, 26, 317-329.

Specific to real-time fMRI: S. M. LaConte. (2011). Decoding fMRI brain states in real-time.

NeuroImage, 56:440-54.

S. M. LaConte, S. J. Peltier, and X. P. Hu. (2007). Real-time fMRI using brain-state classification. Hum Brain Mapp, 208:1033–1044.

Please also consider to reference: T. Joachims, Making Large-Scale SVM Learning Practical.

Advances in Kernel Methods - Support Vector Learning, B. Schoelkopf and C. Burges and A. Smola (ed.), MIT Press, 1999.
RW Cox. AFNI: Software for analysis and visualization of
functional magnetic resonance neuroimages. Computers and Biomedical Research, 29:162-173, 1996.

Table Of Contents

This Page