SVM AutoTrain

[opencv-pushbot]

Transferred from http://code.opencv.org/issues/4314

|| Daniel Horn on 2015-05-05 13:05
|| Priority: Normal
|| Affected: branch 'master' (3.0-dev)
|| Category: ml
|| Tracker: Bug
|| Difficulty: 
|| PR: 
|| Platform: x64 / Windows

SVM AutoTrain

the SVM auto_train function doesnt work as expected. Setting the range of a paramter as a ParamGrid should be enough for auto train, but this process doesnt work correctly. To get the auto_train work, i need to manually set an inital value for the needed params. The inital value should be derived from the ParamGrid.

History

Daniel Horn on 2015-05-20 13:09

-   Assignee deleted (Maria Dimashova)

Jens Garstka on 2015-07-01 12:37

I have (had) the same problem. In case of an RBF kernel it is not sufficient to set the values of C and Gamma somewhere within the minimum and the maximum given by the ParamGrid. Even if you set these values, the autoTrain-method leads to C or Gamma values that are often equal to one of the ParamGrid-limits. As a result the classification rates of the trained SVMs are close to zero.

Accordingly, I build my own autoTrain method for now ...