the difference of predict result between using xgboost-predictor and using python

[fenxouxiaoquan]

Hi,
i train xgboost model with python and save it as 001.model,for test, i use this 001.model to predict one sample in python, get 0.223，but get 0.0604 with using xgboost-predictor，same model,same sample,but different predicted result,and i test another sample, still get different result,only the same thing is using python to predict can get larger score. i have no idea to deal with this problem.note that,my input is a hashmap,like this:{33 : 1.0,34 : 1.0,125 : 0.04261,185 : 0.01504}

[fenxouxiaoquan]

i try once more,seems the problem is gone,i just change double format to float format in the input hashmap

[hlbkin]

@fenxouxiaoquan
Hello, I've actually found the same with standard xgboost4j.

However, when I tryied this library, I got practically the same result as in python (I inserted float[] array)

In xgboost4j I was inserting DMatrix(float[] mydata) as well

How can this be possible?

[fenxouxiaoquan]

@hlbkin
Hello, I didn't get it in fact. Dou you mean that you get the same predict result when you test it in xgboost-predictor and python? I get different predict result again today( 9.189394587749626E-5 for xgboost-predictor and 0.605 for python),just like about three months ago what i experienced, this time my input still a hashmap with nonezero values,do you have any ider to deal with this?

[cpfarrell]

I would recommend to always be using features as floats. XGBoost is explicit that it treats things as 32 bit due to performance optimizations (one example dmlc/xgboost#1410). If a model has been trained using xgboost its split values will be stored as floats and so giving it doubles may cause inaccurate predictions if hit just the right values.