you can make an instance of CvGBTreesParams, set your values, and pass that to either the CvGBTrees constructor, or the train method
unfortunately, it's the very last param, so you'll end up with a beast similar to this:
CvGBTreesParams params;
params.loss_function_type = HUBER_LOSS;
CvGBTrees tree( trainData,tflag,responses, Mat(), Mat(), Mat(), Mat(), params ); //lots of default args
hmm, bear with me, i got no deeper knowledge about gbtrees, but ...
looking at modules/ml/src/gbt.cpp, l 475, it looks like alpha is hardcoded to 0.2, and delta is the outcome of a calculation there.
so, only choice you have is to hack into the src, and recompile the module ;(
But I really don't think they are using the right Quantil Function. Because the original code is "delta = residuals[int(ceil(n*alpha))];", which should be quantil(residuals,alpha). If sigma is the standard deviation, mu the mean of residuals and q the inverse value of normal distribution of alpha, then delta = q * sigma + mean. If alpha = 0.99 then q = 2.326348.
shame on me, but that's waaay over my head ;( ( i understand your reasoning above, but total lost at what the code' s doing there..)
so, if you think, there's changes to be made, make an issue here: http://code.opencv.org/projects/opencv/wiki or here (they love pull requests!) https://github.com/itseez/opencv/pulls
Asked: 2013-02-22 03:23:49 -0600
Seen: 1,585 times
Last updated: Feb 22 '13
