cross validation in random forest doesn't work

[mixmixmix]

Greetings.

Random forest training with number of folds different than rtrees->setCVFolds(0), segfaults. Mentioned on answers.opencv.org

Example code

auto rtrees = cv::ml::RTrees::create(); 
rtrees->setMaxDepth(30);
rtrees->setMinSampleCount(12);
rtrees->setRegressionAccuracy(0);
rtrees->setUseSurrogates(false);
rtrees->setMaxCategories(16);
rtrees->setPriors(cv::Mat());
rtrees->setCalculateVarImportance(false);
rtrees->setActiveVarCount(0);
rtrees->setTermCriteria({cv::TermCriteria::MAX_ITER, 100, 0});
//rtrees->setCVFolds(10); //uncommenting that line causes rtrees->train() to crash

rtrees->train(trainingData32F.colRange(1, trainingData32F.cols),
              cv::ml::ROW_SAMPLE, labels);

Please state the information for your system

• OpenCV version: 3.1
• Host OS: Linux (Ubuntu 15.10) and Mac OS X 10.11
• compiled with clang++

In which part of the OpenCV library you got the issue?

ml module

Expected behaviour

The program should complete training

Actual behaviour

Program received signal SIGSEGV, Segmentation fault.
0x00007ffff1d5c438 in cv::ml::DTreesImpl::calcValue(int, std::vector<int, std::allocator<int> > const&) () from /opt/opencv-3.0.0/lib/libopencv_ml.so.3.1
(gdb) bt
#0  0x00007ffff1d5c438 in cv::ml::DTreesImpl::calcValue(int, std::vector<int, std::allocator<int> > const&) () from /opt/opencv-3.0.0/lib/libopencv_ml.so.3.1
#1  0x00007ffff1d5b533 in cv::ml::DTreesImpl::addNodeAndTrySplit(int, std::vector<int, std::allocator<int> > const&) () from /opt/opencv-3.0.0/lib/libopencv_ml.so.3.1
#2  0x00007ffff1d5ba2b in cv::ml::DTreesImpl::addTree(std::vector<int, std::allocator<int> > const&) () from /opt/opencv-3.0.0/lib/libopencv_ml.so.3.1
#3  0x00007ffff1d2b7e2 in cv::ml::DTreesImplForRTrees::train(cv::Ptr<cv::ml::TrainData> const&, int) [clone .constprop.131] () from /opt/opencv-3.0.0/lib/libopencv_ml.so.3.1
#4  0x00007ffff1d4a7d5 in cv::ml::StatModel::train(cv::_InputArray const&, int, cv::_InputArray const&) () from /opt/opencv-3.0.0/lib/libopencv_ml.so.3.1
Python Exception <class 'gdb.error'> There is no member named _M_dataplus.: 
Python Exception <class 'gdb.error'> There is no member named _M_dataplus.: 
#5  0x000000000040d217 in TrainRandomForest (trainingData=..., dir=, setName=, depth=30, samplecount=12) at /home/miko/rachael/source/train.cpp:40

[ohnozzy]

In fact the documentation of the RTree (http://docs.opencv.org/3.0-beta/modules/ml/doc/random_trees.html) claims the following:

In random trees there is no need for any accuracy estimation procedures, such as cross-validation or bootstrap, or a separate test set to get an estimate of the training error.

So RTree is not designed to work with cross validation. In this case shall we just add an assertion to enforce CVFolds to be 0?

[mixmixmix]

That's right @ohnozzy . It is even in the original paper, my bad.
It'd be best to have a message explaining that, because assertion wouldn't give all the information.

[simonhaenisch]

If it shouldn't be changeable by design, can't this line just be removed:
https://github.com/opencv/opencv/blob/master/modules/ml/src/rtrees.cpp#L371
No need to add the setter/getter, right?

CVFolds is still set to zero in the DTreesImpl here:
https://github.com/opencv/opencv/blob/master/modules/ml/src/rtrees.cpp#L78

[Yumin-Sun-00]

But when I apply setCVFolds for DTrees, the method is also not implemented, why?