Ask Your Question

Revision history [back]

click to hide/show revision 1
initial version

SVM train auto problems

Is there a problem with SVM::train_auto? I am asking this because I have tried to train different SVM classifiers to get the best one. I have also used the default grids of OpenCV. But I have met the following problems:

  1. Crash when training of the EPS_SVR with any of the kernel types. I have got the error of OpenCV Error: Assertion failed (sv_count != 0) in do_train at some line of svm.cpp. I have also met the problem in another question, but I wanted to say more.
  2. Trained, but having Inf, -Inf and nan values in .xml file, when using the NU_SVC with SIGMOID kernel type. It also return a nan when calling predict (IMHO: which is normal based on the .xml).
  3. Crash when training the NU_SVR with POLY kernel because of the same OpenCV Error: Assertion failed (sv_count != 0) in do_train at some line of svm.cpp.

Is this a bug? Is it a bug of train_auto or because of the default grids? What are the default grids? How to find the parameters of default grids, if I want to create custom ones? I think I have also met the problem of sv_count != 0 for some values of the parameters, when training manually (with some fixed values of the parameters). I think it is linked about two "big" values or one big and the other not big enough.

SVM train auto problems

Is there a problem with SVM::train_auto? I am asking this because I have tried to train different SVM classifiers to get the best one. I have also used the default grids of OpenCV. But I have met the following problems:

  1. Crash when training of the EPS_SVR with any of the kernel types. I have got the error of OpenCV Error: Assertion failed (sv_count != 0) in do_train at some line of svm.cpp. I have also met the problem in another question, but I wanted to say more.
  2. Trained, but having Inf, -Inf and nan values in .xml file, when using the NU_SVC with SIGMOID kernel type. It also return a nan when calling predict (IMHO: which is normal based on the .xml).
  3. Crash when training the NU_SVR with POLY kernel because of the same OpenCV Error: Assertion failed (sv_count != 0) in do_train at some line of svm.cpp.

Is this a bug? Is it a bug of train_auto or because of the default grids? What are the default grids? How to find the parameters of default grids, if I want to create custom ones? I think I have also met the problem of sv_count != 0 for some values of the parameters, when training manually (with some fixed values of the parameters). I think it is linked about two "big" values or one big and the other not big enough.


The .xml of NU_SVC with SIGMOID kernel type:

<decision_functions>
    <_>
      <sv_count>18</sv_count>
      <rho>.Nan</rho>
      <alpha>
        .Inf .Inf .Inf .Inf .Inf .Inf .Inf .Inf .Inf -.Inf -.Inf -.Inf
        -.Inf -.Inf -.Inf -.Inf -.Inf -.Inf</alpha>
      <index>
        0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17</index></_></decision_functions>

SVM train auto problems

Is there a problem with SVM::train_auto? I am asking this because I have tried to train different SVM classifiers to get the best one. I have also used the default grids of OpenCV. But I have met the following problems:

  1. Crash when training of the EPS_SVR with any of the kernel types. I have got the error of OpenCV Error: Assertion failed (sv_count != 0) in do_train at some line of svm.cpp. (I think is 1346, but not sure it is for all). I have also met the problem in another question, but I wanted to say more.
  2. Trained, but having Inf, -Inf and nan values in .xml file, when using the NU_SVC with SIGMOID kernel type. It also return a nan when calling predict (IMHO: which is normal based on the .xml).
  3. Crash when training the NU_SVR with POLY kernel because of the same OpenCV Error: Assertion failed (sv_count != 0) in do_train at some line of svm.cpp.

Is this a bug? Is it a bug of train_auto or because of the default grids? What are the default grids? How to find the parameters of default grids, if I want to create custom ones? I think I have also met the problem of sv_count != 0 for some values of the parameters, when training manually (with some fixed values of the parameters). I think it is linked about two "big" values or one big and the other not big enough.


The .xml of NU_SVC with SIGMOID kernel type:

<decision_functions>
    <_>
      <sv_count>18</sv_count>
      <rho>.Nan</rho>
      <alpha>
        .Inf .Inf .Inf .Inf .Inf .Inf .Inf .Inf .Inf -.Inf -.Inf -.Inf
        -.Inf -.Inf -.Inf -.Inf -.Inf -.Inf</alpha>
      <index>
        0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17</index></_></decision_functions>