def _get_interval(t, cut_gaps):
    if pd.isnull(t):
        return 'NA'
    for i, cut_gap in enumerate(cut_gaps):
        if t <= cut_gap:
            if i == 0:
                return "%s-(-inf, %s]" % (i, cut_gap)
            return "%s-(%s, %s]" % (i, cut_gaps[i - 1], cut_gap)
    return "%s-(%s, inf)" % (len(cut_gaps), cut_gaps[-1])



  class BinEncoder(TransformerMixin):
    def __init__(self, cut_gaps):
        self.cut_gaps = cut_gaps

    def fit(self, X, y=None):
        return self

    def transform(self, X):
        return np.array(list(map(lambda t: _get_interval(t, self.cut_gaps), X)))



class WoeEncoder(TransformerMixin):
    def __init__(self):
        pass

    def fit(self, X, y=None):
        a =  at.calc_iv(X, y, return_type='woe')
        self.woe_dict = a.set_index('interval').to_dict()['woe']
        return self

    def transform(self, X):
        res = np.array(list(map(lambda t: self.woe_dict.get(t), X)))
#        return res
        return np.where(pd.isnull(res), 0, res)


# val和oot集边界值超过dev的边界时，需要处理
# OneHot编码时使用LabelBinarizer
#mapper = DataFrameMapper(
#    [(feature_name, [BinEncoder(cut_gaps), LabelBinarizer()])
#     for feature_name, cut_gaps in list(feat_cut_dict.items())]
#)

# woe填充时，使用自定义的WoeEncoder
mapper = DataFrameMapper(
    [(feature_name, [BinEncoder(cut_gaps), WoeEncoder()])
     for feature_name, cut_gaps in list(feat_cut_dict.items())]
)
lr_model = LogisticRegression()
lm = PMMLPipeline([
    ("mapper", mapper),
    ("lr", lr_model),
])