注意
跳到最后下载完整示例代码。
WOEEncoder 的转换器，来自 categorical_encoder¶
WOEEncoder 是在 categorical_encoder 中实现的转换器，因此，任何转换器都不会包含在 *sklearn-onnx* 中，*sklearn-onnx* 只实现 *scikit-learn* 模型的转换器。无论如何，本示例演示了如何为 *WOEEncoder* 实现自定义转换器。此代码尚未针对原始编码器可能处理的所有情况进行全面测试。
一个简单示例¶
我们以 Iris 数据集为例。每个特征都被转换为整数。
import numpy as np
from onnxruntime import InferenceSession
from sklearn.datasets import load_iris
from sklearn.preprocessing import OrdinalEncoder as SklOrdinalEncoder
from category_encoders import WOEEncoder, OrdinalEncoder
from skl2onnx import update_registered_converter, to_onnx, get_model_alias
from skl2onnx.common.data_types import FloatTensorType
from skl2onnx.common.utils import check_input_and_output_numbers
from skl2onnx.algebra.onnx_ops import OnnxCast
from skl2onnx.algebra.onnx_operator import OnnxSubEstimator
from skl2onnx.sklapi import WOETransformer
import skl2onnx.sklapi.register # noqa: F401
data = load_iris()
X, y = data.data, data.target
X = X.astype(np.int64)[:, :2]
y = (y == 2).astype(np.int64)
woe = WOEEncoder(cols=[0]).fit(X, y)
print(woe.transform(X[:5]))
0 1
0 -1.405712 3
1 -1.724166 3
2 -1.724166 3
3 -1.724166 3
4 -1.405712 3
我们来看看模型的训练参数。似乎 WOEEncoder 使用了一个 OrdinalEncoder，但不是 scikit-learn 中的那个。我们需要为这个模型工具添加一个转换器。
print("encoder", type(woe.ordinal_encoder), woe.ordinal_encoder)
print("mapping", woe.mapping)
print("encoder.mapping", woe.ordinal_encoder.mapping)
print("encoder.cols", woe.ordinal_encoder.cols)
encoder <class 'category_encoders.ordinal.OrdinalEncoder'> OrdinalEncoder(cols=[0],
mapping=[{'col': 0, 'data_type': dtype('int64'),
'mapping': 5.0 1
4.0 2
7.0 3
6.0 4
NaN -2
dtype: int64}])
mapping {0: 0
1 -1.405712
2 -1.724166
3 2.545531
4 0.961411
-1 0.000000
-2 0.000000
dtype: float64}
encoder.mapping [{'col': 0, 'mapping': 5.0 1
4.0 2
7.0 3
6.0 4
NaN -2
dtype: int64, 'data_type': dtype('int64')}]
encoder.cols [0]
OrdinalEncoder 的自定义转换器¶
我们从示例 实现一个新的转换器 开始，然后编写转换代码。
def ordenc_to_sklearn(op_mapping):
"Converts OrdinalEncoder mapping to scikit-learn OrdinalEncoder."
cats = []
for column_map in op_mapping:
col = column_map["col"]
while len(cats) <= col:
cats.append(None)
mapping = column_map["mapping"]
res = []
for i in range(mapping.shape[0]):
if np.isnan(mapping.index[i]):
continue
ind = mapping.iloc[i]
while len(res) <= ind:
res.append(0)
res[ind] = mapping.index[i]
cats[col] = np.array(res, dtype=np.int64)
skl_ord = SklOrdinalEncoder(categories=cats, dtype=np.int64)
skl_ord.categories_ = cats
return skl_ord
def ordinal_encoder_shape_calculator(operator):
check_input_and_output_numbers(operator, input_count_range=1, output_count_range=1)
input_type = operator.inputs[0].type.__class__
input_dim = operator.inputs[0].get_first_dimension()
shape = operator.inputs[0].type.shape
second_dim = None if len(shape) != 2 else shape[1]
output_type = input_type([input_dim, second_dim])
operator.outputs[0].type = output_type
def ordinal_encoder_converter(scope, operator, container):
op = operator.raw_operator
opv = container.target_opset
X = operator.inputs[0]
skl_ord = ordenc_to_sklearn(op.mapping)
cat = OnnxSubEstimator(
skl_ord, X, op_version=opv, output_names=operator.outputs[:1]
)
cat.add_to(scope, container)
update_registered_converter(
OrdinalEncoder,
"CategoricalEncoderOrdinalEncoder",
ordinal_encoder_shape_calculator,
ordinal_encoder_converter,
)
我们来计算一个简短示例的输出。
enc = OrdinalEncoder(cols=[0, 1])
enc.fit(X)
print(enc.transform(X[:5]))
0 1
0 1 1
1 2 1
2 2 1
3 2 1
4 1 1
我们来检查 ONNX 转换是否产生相同的结果。
ord_onx = to_onnx(enc, X[:1], target_opset=14)
sess = InferenceSession(ord_onx.SerializeToString(), providers=["CPUExecutionProvider"])
print(sess.run(None, {"X": X[:5]})[0])
[[1 1]
[2 1]
[2 1]
[2 1]
[1 1]]
这有效。
WOEEncoder 的自定义转换器¶
我们从示例 实现一个新的转换器 开始，然后编写转换代码。
def woeenc_to_sklearn(op_mapping):
"Converts WOEEncoder mapping to scikit-learn OrdinalEncoder."
cats = []
ws = []
for column_map in op_mapping.items():
col = column_map[0]
while len(cats) <= col:
cats.append("passthrough")
ws.append(None)
mapping = column_map[1]
intervals = []
weights = []
for i in range(mapping.shape[0]):
ind = mapping.index[i]
if ind < 0:
continue
intervals.append((float(ind - 1), float(ind), False, True))
weights.append(mapping.iloc[i])
cats[col] = intervals
ws[col] = weights
skl = WOETransformer(intervals=cats, weights=ws, onehot=False)
skl.fit(None)
return skl
def woe_encoder_parser(scope, model, inputs, custom_parsers=None):
if len(inputs) != 1:
raise RuntimeError("Unexpected number of inputs: %d != 1." % len(inputs))
if inputs[0].type is None:
raise RuntimeError("Unexpected type: %r." % (inputs[0],))
alias = get_model_alias(type(model))
this_operator = scope.declare_local_operator(alias, model)
this_operator.inputs.append(inputs[0])
this_operator.outputs.append(
scope.declare_local_variable("catwoe", FloatTensorType())
)
return this_operator.outputs
def woe_encoder_shape_calculator(operator):
check_input_and_output_numbers(operator, input_count_range=1, output_count_range=1)
input_dim = operator.inputs[0].get_first_dimension()
shape = operator.inputs[0].type.shape
second_dim = None if len(shape) != 2 else shape[1]
output_type = FloatTensorType([input_dim, second_dim])
operator.outputs[0].type = output_type
def woe_encoder_converter(scope, operator, container):
op = operator.raw_operator
opv = container.target_opset
X = operator.inputs[0]
sub = OnnxSubEstimator(op.ordinal_encoder, X, op_version=opv)
cast = OnnxCast(sub, op_version=opv, to=np.float32)
skl_ord = woeenc_to_sklearn(op.mapping)
cat = OnnxSubEstimator(
skl_ord,
cast,
op_version=opv,
output_names=operator.outputs[:1],
input_types=[FloatTensorType()],
)
cat.add_to(scope, container)
update_registered_converter(
WOEEncoder,
"CategoricalEncoderWOEEncoder",
woe_encoder_shape_calculator,
woe_encoder_converter,
parser=woe_encoder_parser,
)
我们来计算一个简短示例的输出。
woe = WOEEncoder(cols=[0, 1]).fit(X, y)
print(woe.transform(X[:5]))
0 1
0 -1.405712 -0.035947
1 -1.724166 -0.035947
2 -1.724166 -0.035947
3 -1.724166 -0.035947
4 -1.405712 -0.035947
我们来检查 ONNX 转换是否产生相同的结果。
woe_onx = to_onnx(woe, X[:1], target_opset=14)
sess = InferenceSession(woe_onx.SerializeToString(), providers=["CPUExecutionProvider"])
print(sess.run(None, {"X": X[:5]})[0])
[[-1.4057125 -0.03594739]
[-1.7241662 -0.03594739]
[-1.7241662 -0.03594739]
[-1.7241662 -0.03594739]
[-1.4057125 -0.03594739]]
脚本总运行时间：（0 分 0.512 秒）