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:
- Crash when training of the
EPS_SVR
with any of the kernel types. I have got the error ofOpenCV Error: Assertion failed (sv_count != 0) in do_train
at some line ofsvm.cpp
. I have also met the problem in another question, but I wanted to say more. - Trained, but having
Inf
,-Inf
andnan
values in.xml
file, when using theNU_SVC
withSIGMOID
kernel type. It also return anan
when calling predict (IMHO: which is normal based on the.xml
). - Crash when training the
NU_SVR
withPOLY
kernel because of the sameOpenCV Error: Assertion failed (sv_count != 0) in do_train
at some line ofsvm.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.