Using as command line tools

Installation

When cl-online-learning is installed with roswell, you will have clol-train and clol-predict files in ~/.roswell/bin/.

ros install masatoi/cl-online-learning

These usage are similar to the libsvm's commands.

Usage of clol-train

clol-train command receives a training-set-file and several options, and then generates a model-file. The dataset file is in libsvm data format (More details for data format).

clol-train [options] training-set-file model-file

options:
-dim : number of feature dimensions (optional)
-n-class : number of classes (optional)
-n-epoch : number of epochs (optional)
-type : type of learning model
  for binary or multi-class classification
        0 -- Perceptron
        1 -- AROW (default)
        2 -- SCW-I
-sparse : whether sparse data or not
        0 -- binary
        1 -- sparse  (default)
-mtype : multiclass classifier type
        0 -- one-vs-rest (default)
        1 -- one-vs-one
-gamma : regularization parameter for AROW (default 10.0)
-eta : decay parameter for SCW-I (default 0.9 (0 < eta < 1))
-c : regularization parameter for SCW-I (default 1.0)

Usage of clol-predict

clol-predict command receives a test-set-file and the model-file generated by clol-train, and then generates an output-file.

clol-predict test-set-file model-file output-file

Examples

binary classification using AROW

• Dataset: a1a
• AROW, gamma=10.0

clol-train a1a a1a.model
clol-predict a1a.t a1a.model a1a.out

Accuracy: 84.461815%, Correct: 26146, Total: 30956

multiclass classification using SCW-I

• Dataset: mnist
• SCW-I, one-vs-one, eta=0.9, c=0.1

clol-train mnist.scale mnist.model -type 2 -mtype 1 -eta 0.9 -c 0.1 -n-epoch 10
clol-predict mnist.scale.t mnist.model mnist.out

Accuracy: 94.36%, Correct: 9436, Total: 10000