Wrong handling of nans in boosted decision trees

While pandas supports a column with mixed ints/ floats which are nan, tc does not identify float("nan") as None and thus Xgboost failed on data converted from pandas.

import pandas as pd
data = pd.DataFrame({'a':[1,float("nan"),2],'target':[1,0,1]})
import turicreate as tc
sf_data = tc.SFrame(data)
tc.boosted_trees_classifier.create(sf_data, target='target')

results in this error.

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
/Users/dbicksokn/g3k/lib/python3.6/site-packages/turicreate/extensions.py in __run_class_function(self, fnname, args, kwargs)
    275         try:
--> 276             ret = self._tkclass.call_function(fnname, argument_dict)
    277         except RuntimeError as exc:

turicreate/cython/cy_model.pyx in turicreate.cython.cy_model.UnityModel.call_function()

turicreate/cython/cy_model.pyx in turicreate.cython.cy_model.UnityModel.call_function()

RuntimeError: Detected inf/nan values in feature(s) 'a'. Cannot proceed with model training.

During handling of the above exception, another exception occurred:

ToolkitError                              Traceback (most recent call last)
<ipython-input-7-632ac263ea9e> in <module>()
----> 1 tc.boosted_trees_classifier.create(sf_data, target='target')

/Users/dbicksokn/g3k/lib/python3.6/site-packages/turicreate/toolkits/classifier/boosted_trees_classifier.py in create(dataset, target, features, max_iterations, validation_set, class_weights, max_depth, step_size, min_loss_reduction, min_child_weight, row_subsample, column_subsample, verbose, random_seed, metric, **kwargs)
    635                         verbose = verbose,
    636                         metric = metric,
--> 637                         **kwargs)
    638     return BoostedTreesClassifier(model.__proxy__)

/Users/dbicksokn/g3k/lib/python3.6/site-packages/turicreate/toolkits/_supervised_learning.py in create(dataset, target, model_name, features, validation_set, distributed, verbose, seed, **kwargs)
    330     model = _turicreate.extensions.__dict__[model_name]()
    331     with QuietProgress(verbose):
--> 332         model.train(dataset, target, validation_set, options)
    333
    334     return SupervisedLearningModel(model, model_name)

/Users/dbicksokn/g3k/lib/python3.6/site-packages/turicreate/extensions.py in <lambda>(*args, **kwargs)
    290         elif name in self._functions:
    291             # is it a function?
--> 292             ret = lambda *args, **kwargs: self.__run_class_function(name, args, kwargs)
    293             ret.__doc__ = "Name: " + name + "\nParameters: " + str(self._functions[name]) + "\n"
    294             try:

/Users/dbicksokn/g3k/lib/python3.6/site-packages/turicreate/extensions.py in __run_class_function(self, fnname, args, kwargs)
    277         except RuntimeError as exc:
    278             # Expose C++ exceptions using ToolkitError.
--> 279             raise _ToolkitError(exc)
    280         ret = _wrap_function_return(ret)
    281         return ret

ToolkitError: Detected inf/nan values in feature(s) 'a'. Cannot proceed with model training.

=> boosted trees supports Nones but not float("nan")

apple / turicreate

Wrong handling of nans in boosted decision trees #1589