Cv::ml::TrainData::loadFromCSV Parameters

edit

thanks a lot!!!!!

SVM.train(TrainData,labelsMat, Mat(), Mat(), params);

this is the normal way of training svm right? so when i load it from csv what is my labelsMat?

wait, no, yours looks like opencv2.4 code.

remember, that traindata already contains the labels, so you would use this overload

Ptr<cv::ml::TrainData> tdata = cv::ml::TrainData::loadFromCSV("my.csv", 0, 0, 1);
Ptr<cv::ml::SVM> svm = cv::ml::SVM::create();
svm->setC(10); // maybe you want to set some params
svm->train(tdata);

if you don't use a csv file, you'd have a Mat features, and another Mat labels, and use it like :

svm->train(features, 0, labels);

Sorry for being so bad at this! i am kinda new to this. Is there a way that i can contact you?

do not feel sorry in the 1st place, it's not your fault, that the api changed significantly, causing a lot of confusion !

then, i somewhat believe, it's better, to keep it here, (or on #opencv), so more folks can look at it, and help !

thanks mate!

1. How can i save the svm.train and access in the classification program?
2. Then how can i classify the image using the trained file?

sure:

svm->save("my.xml");
// careful, this makes a **new** instance !
Ptr<SVM> svm = Algorithm::load<SVM>("my.xml");

i am using opencv2.4 so the code do not work with me. :(

svm->save("my.xml");

is it,

svm.save("my.xmal");

and then after saving is done i want to use that save file to classify the image. are these steps correct?

1. Load the xml file
2. Get features from the image
3. Classify using svm

is there any example for this?

oh, sorry, it seems, i got your opencv version wrong then, all the time ;(

but yes, your steps are correct (once you trained and saved the model, you do not need to reload the csv or retrain it, you can just load the saved model and do your prediction)

ok i managed to do it. i am just using two classes (white & silver) So train the svm and use the model as below,

 CvSVM mySVM1;
    mySVM1.load("/home/damiya/test/Sakura/my.xml");
    .
    .
    .
    .
    float response = mySVM1.predict(sampleMat);

but the modele returns only one class every time. (ex. it returns everything as white even i feed the features that i used to train it)

well, once you got the general pipeline working, the real problems start (tweaking svm params, cleaning data)

what is actually in your csv file ? some sort of (hsv?) pixels ? silver and white are incredibly difficult to seperate anyway.

my csv file contains hsv rgb and yCrCb data for each pixel this is csv file.

first column contains the label. then in order R G B H S V Y Cr Cb and it repeats.

cancelled the load after a couple mb's (how large is it?)

so it seems, you got enough data, maybe you need to play with the svm params (C would be a good start, like 0.1, 1, 10, 100, etc) , try different kernel types(linear/rbm/poly), maybe even train_auto() works on 2.4 (haven't tried)

unfortunately that means retraining your SVM for each param change, so huge time sink.

its 31mb. i tried different kernel types such as linear,rbm, and poly but every time its wrong.

i will try train_auto().. hops that works :) :)

hey,

i tried train_auto() and it takes a lot of time. Does it take a lot of time?

yea, ages. one complete training run for each combination of params

again, huge time-sink ahead ;(

its still running :(