# 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>

edit retag close merge delete