注意

转到末尾 以下载完整的示例代码。

逐步输出中间结果

我们重用示例 转换带有 ColumnTransformer 的管道 中的示例，并逐步输出中间结果。转换后的模型很可能由于未正确实现的自定义转换器而产生不同的输出或失败。一种方法是查看 ONNX 图中每个节点的输出。

创建和训练一个复杂的管道

我们重用在示例 带有混合类型的 Column Transformer 中实现的管道。有一个改动，因为 ONNX-ML Imputer 不支持字符串类型。这不能作为最终 ONNX 管道的一部分，必须移除。请查看下面以 --- 开头的注释。

import skl2onnx
import onnx
import sklearn
import matplotlib.pyplot as plt
import os
from onnx.tools.net_drawer import GetPydotGraph, GetOpNodeProducer
from skl2onnx.helpers.onnx_helper import select_model_inputs_outputs
from skl2onnx.helpers.onnx_helper import save_onnx_model
from skl2onnx.helpers.onnx_helper import enumerate_model_node_outputs
from skl2onnx.helpers.onnx_helper import load_onnx_model
import numpy
import onnxruntime as rt
from skl2onnx import convert_sklearn
import pprint
from skl2onnx.common.data_types import (
    FloatTensorType,
    StringTensorType,
    Int64TensorType,
)
import numpy as np
import pandas as pd
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split

titanic_url = (
    "https://raw.githubusercontent.com/amueller/"
    "scipy-2017-sklearn/091d371/notebooks/datasets/titanic3.csv"
)
data = pd.read_csv(titanic_url)
X = data.drop("survived", axis=1)
y = data["survived"]

# SimpleImputer on string is not available
# for string in ONNX-ML specifications.
# So we do it beforehand.
for cat in ["embarked", "sex", "pclass"]:
    X[cat].fillna("missing", inplace=True)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

numeric_features = ["age", "fare"]
numeric_transformer = Pipeline(
    steps=[("imputer", SimpleImputer(strategy="median")), ("scaler", StandardScaler())]
)

categorical_features = ["embarked", "sex", "pclass"]
categorical_transformer = Pipeline(
    steps=[
        # --- SimpleImputer is not available for strings in ONNX-ML specifications.
        # ('imputer', SimpleImputer(strategy='constant', fill_value='missing')),
        ("onehot", OneHotEncoder(handle_unknown="ignore"))
    ]
)

preprocessor = ColumnTransformer(
    transformers=[
        ("num", numeric_transformer, numeric_features),
        ("cat", categorical_transformer, categorical_features),
    ]
)

clf = Pipeline(
    steps=[
        ("preprocessor", preprocessor),
        ("classifier", LogisticRegression(solver="lbfgs")),
    ]
)

clf.fit(X_train, y_train)

Pipeline(steps=[('preprocessor',
                 ColumnTransformer(transformers=[('num',
                                                  Pipeline(steps=[('imputer',
                                                                   SimpleImputer(strategy='median')),
                                                                  ('scaler',
                                                                   StandardScaler())]),
                                                  ['age', 'fare']),
                                                 ('cat',
                                                  Pipeline(steps=[('onehot',
                                                                   OneHotEncoder(handle_unknown='ignore'))]),
                                                  ['embarked', 'sex',
                                                   'pclass'])])),
                ('classifier', LogisticRegression())])
在 Jupyter 环境中，请重新运行此单元格以显示 HTML 表示，或信任该笔记本。
在 GitHub 上，HTML 表示无法渲染，请尝试使用 nbviewer.org 加载此页面。


定义 ONNX 图的输入

sklearn-onnx 不知道用于训练模型的特征，但它需要知道哪个特征具有哪个名称。我们只是重用数据帧的列定义。

print(X_train.dtypes)

pclass         int64
name          object
sex           object
age          float64
sibsp          int64
parch          int64
ticket        object
fare         float64
cabin         object
embarked      object
boat          object
body         float64
home.dest     object
dtype: object

转换后。

def convert_dataframe_schema(df, drop=None):
    inputs = []
    for k, v in zip(df.columns, df.dtypes):
        if drop is not None and k in drop:
            continue
        if v == "int64":
            t = Int64TensorType([None, 1])
        elif v == "float64":
            t = FloatTensorType([None, 1])
        else:
            t = StringTensorType([None, 1])
        inputs.append((k, t))
    return inputs


inputs = convert_dataframe_schema(X_train)

pprint.pprint(inputs)

[('pclass', Int64TensorType(shape=[None, 1])),
 ('name', StringTensorType(shape=[None, 1])),
 ('sex', StringTensorType(shape=[None, 1])),
 ('age', FloatTensorType(shape=[None, 1])),
 ('sibsp', Int64TensorType(shape=[None, 1])),
 ('parch', Int64TensorType(shape=[None, 1])),
 ('ticket', StringTensorType(shape=[None, 1])),
 ('fare', FloatTensorType(shape=[None, 1])),
 ('cabin', StringTensorType(shape=[None, 1])),
 ('embarked', StringTensorType(shape=[None, 1])),
 ('boat', StringTensorType(shape=[None, 1])),
 ('body', FloatTensorType(shape=[None, 1])),
 ('home.dest', StringTensorType(shape=[None, 1]))]

将单个列合并为向量不是计算预测的最有效方法。可以在将管道转换为图之前完成。

将管道转换为 ONNX

try:
    model_onnx = convert_sklearn(clf, "pipeline_titanic", inputs, target_opset=12)
except Exception as e:
    print(e)

scikit-learn 在可能的情况下会进行隐式转换。sklearn-onnx 不会。OneHotEncoder 的 ONNX 版本必须应用于相同类型的列。

X_train["pclass"] = X_train["pclass"].astype(str)
X_test["pclass"] = X_test["pclass"].astype(str)
white_list = numeric_features + categorical_features
to_drop = [c for c in X_train.columns if c not in white_list]
inputs = convert_dataframe_schema(X_train, to_drop)

model_onnx = convert_sklearn(clf, "pipeline_titanic", inputs, target_opset=12)


# And save.
with open("pipeline_titanic.onnx", "wb") as f:
    f.write(model_onnx.SerializeToString())

比较预测

最后一步，我们需要确保转换后的模型产生相同的预测、标签和概率。让我们从scikit-learn 开始。

print("predict", clf.predict(X_test[:5]))
print("predict_proba", clf.predict_proba(X_test[:1]))

predict [0 0 0 1 1]
predict_proba [[0.78796696 0.21203304]]

使用 onnxruntime 进行预测。我们需要删除已删除的列，并将双精度向量更改为单精度向量，因为 onnxruntime 不支持双精度浮点数。onnxruntime 不接受 dataframe。输入必须以字典列表的形式提供。最后一个细节是，每一列都被描述为单个列的矩阵，而不是真正的向量，这解释了最后一行中的reshape。

X_test2 = X_test.drop(to_drop, axis=1)
inputs = {c: X_test2[c].values for c in X_test2.columns}
for c in numeric_features:
    inputs[c] = inputs[c].astype(np.float32)
for k in inputs:
    inputs[k] = inputs[k].reshape((inputs[k].shape[0], 1))

我们已准备好运行 onnxruntime。

sess = rt.InferenceSession("pipeline_titanic.onnx", providers=["CPUExecutionProvider"])
pred_onx = sess.run(None, inputs)
print("predict", pred_onx[0][:5])
print("predict_proba", pred_onx[1][:1])

predict [0 0 0 1 1]
predict_proba [{0: 0.5883165597915649, 1: 0.41168344020843506}]

计算中间输出

不幸的是，目前没有办法让 onnxruntime 检索中间节点的输出。我们需要修改 ONNX，然后再将其提供给 onnxruntime。让我们先看看中间输出的列表。

model_onnx = load_onnx_model("pipeline_titanic.onnx")
for out in enumerate_model_node_outputs(model_onnx):
    print(out)

merged_columns
embarkedout
sexout
pclassout
concat_result
variable
variable2
variable1
transformed_column
label
probabilities
output_label
output_probability

很难确定哪个是什么，因为 ONNX 拥有比原始 scikit-learn 管道更多的运算符。 显示 ONNX 图 中的图有助于我们找到数值管道和文本管道的输出：variable1，variable2。让我们先看看数值管道。

num_onnx = select_model_inputs_outputs(model_onnx, "variable1")
save_onnx_model(num_onnx, "pipeline_titanic_numerical.onnx")

b'\x08\x07\x12\x08skl2onnx\x1a\x061.19.1"\x07ai.onnx(\x002\x00:\xcd\x03\n:\n\x03age\n\x04fare\x12\x0emerged_columns\x1a\x06Concat"\x06Concat*\x0b\n\x04axis\x18\x01\xa0\x01\x02:\x00\n}\n\x0emerged_columns\x12\x08variable\x1a\x07Imputer"\x07Imputer*#\n\x14imputed_value_floats=\x00\x00\xe0A=gDgA\xa0\x01\x06*\x1e\n\x14replaced_value_float\x15\x00\x00\xc0\x7f\xa0\x01\x01:\nai.onnx.ml\n^\n\x08variable\x12\tvariable1\x1a\x06Scaler"\x06Scaler*\x15\n\x06offset=\x17\xc8\xebA=<\t\x07B\xa0\x01\x06*\x14\n\x05scale=D)\x9e==\xe5j\x9e<\xa0\x01\x06:\nai.onnx.ml\x12\x10pipeline_titanic*\x1f\x08\x02\x10\x07:\x0b\xff\xff\xff\xff\xff\xff\xff\xff\xff\x01\tB\x0cshape_tensorZ\x16\n\x06pclass\x12\x0c\n\n\x08\x08\x12\x06\n\x00\n\x02\x08\x01Z\x13\n\x03sex\x12\x0c\n\n\x08\x08\x12\x06\n\x00\n\x02\x08\x01Z\x13\n\x03age\x12\x0c\n\n\x08\x01\x12\x06\n\x00\n\x02\x08\x01Z\x14\n\x04fare\x12\x0c\n\n\x08\x01\x12\x06\n\x00\n\x02\x08\x01Z\x18\n\x08embarked\x12\x0c\n\n\x08\x08\x12\x06\n\x00\n\x02\x08\x01b\x0b\n\tvariable1B\x04\n\x00\x10\x0bB\x0e\n\nai.onnx.ml\x10\x01'

我们来计算数值特征。

sess = rt.InferenceSession(
    "pipeline_titanic_numerical.onnx", providers=["CPUExecutionProvider"]
)
numX = sess.run(None, inputs)
print("numerical features", numX[0][:1])

numerical features [[0.4268576 9.254629 ]]

我们对文本特征也做同样的处理。

print(model_onnx)
text_onnx = select_model_inputs_outputs(model_onnx, "variable2")
save_onnx_model(text_onnx, "pipeline_titanic_textual.onnx")
sess = rt.InferenceSession(
    "pipeline_titanic_textual.onnx", providers=["CPUExecutionProvider"]
)
numT = sess.run(None, inputs)
print("textual features", numT[0][:1])

ir_version: 7
producer_name: "skl2onnx"
producer_version: "1.19.1"
domain: "ai.onnx"
model_version: 0
doc_string: ""
graph {
  node {
    input: "age"
    input: "fare"
    output: "merged_columns"
    name: "Concat"
    op_type: "Concat"
    attribute {
      name: "axis"
      i: 1
      type: INT
    }
    domain: ""
  }
  node {
    input: "embarked"
    output: "embarkedout"
    name: "OneHotEncoder"
    op_type: "OneHotEncoder"
    attribute {
      name: "cats_strings"
      strings: "C"
      strings: "Q"
      strings: "S"
      strings: "missing"
      type: STRINGS
    }
    attribute {
      name: "zeros"
      i: 1
      type: INT
    }
    domain: "ai.onnx.ml"
  }
  node {
    input: "sex"
    output: "sexout"
    name: "OneHotEncoder1"
    op_type: "OneHotEncoder"
    attribute {
      name: "cats_strings"
      strings: "female"
      strings: "male"
      type: STRINGS
    }
    attribute {
      name: "zeros"
      i: 1
      type: INT
    }
    domain: "ai.onnx.ml"
  }
  node {
    input: "pclass"
    output: "pclassout"
    name: "OneHotEncoder2"
    op_type: "OneHotEncoder"
    attribute {
      name: "cats_strings"
      strings: "1"
      strings: "2"
      strings: "3"
      type: STRINGS
    }
    attribute {
      name: "zeros"
      i: 1
      type: INT
    }
    domain: "ai.onnx.ml"
  }
  node {
    input: "embarkedout"
    input: "sexout"
    input: "pclassout"
    output: "concat_result"
    name: "Concat1"
    op_type: "Concat"
    attribute {
      name: "axis"
      i: -1
      type: INT
    }
    domain: ""
  }
  node {
    input: "merged_columns"
    output: "variable"
    name: "Imputer"
    op_type: "Imputer"
    attribute {
      name: "imputed_value_floats"
      floats: 28
      floats: 14.4542
      type: FLOATS
    }
    attribute {
      name: "replaced_value_float"
      f: nan
      type: FLOAT
    }
    domain: "ai.onnx.ml"
  }
  node {
    input: "concat_result"
    input: "shape_tensor"
    output: "variable2"
    name: "Reshape"
    op_type: "Reshape"
    domain: ""
  }
  node {
    input: "variable"
    output: "variable1"
    name: "Scaler"
    op_type: "Scaler"
    attribute {
      name: "offset"
      floats: 29.4727
      floats: 33.7590179
      type: FLOATS
    }
    attribute {
      name: "scale"
      floats: 0.0772271454
      floats: 0.0193380807
      type: FLOATS
    }
    domain: "ai.onnx.ml"
  }
  node {
    input: "variable1"
    input: "variable2"
    output: "transformed_column"
    name: "Concat2"
    op_type: "Concat"
    attribute {
      name: "axis"
      i: 1
      type: INT
    }
    domain: ""
  }
  node {
    input: "transformed_column"
    output: "label"
    output: "probabilities"
    name: "LinearClassifier"
    op_type: "LinearClassifier"
    attribute {
      name: "classlabels_ints"
      ints: 0
      ints: 1
      type: INTS
    }
    attribute {
      name: "coefficients"
      floats: 0.359008849
      floats: 0.0514151268
      floats: -0.355432689
      floats: 0.247025743
      floats: 0.314346433
      floats: -0.270904839
      floats: -1.22721398
      floats: 1.16224873
      floats: -0.955703616
      floats: -0.0454647
      floats: 0.936203
      floats: -0.359008849
      floats: -0.0514151268
      floats: 0.355432689
      floats: -0.247025743
      floats: -0.314346433
      floats: 0.270904839
      floats: 1.22721398
      floats: -1.16224873
      floats: 0.955703616
      floats: 0.0454647
      floats: -0.936203
      type: FLOATS
    }
    attribute {
      name: "intercepts"
      floats: -0.123175517
      floats: 0.123175517
      type: FLOATS
    }
    attribute {
      name: "multi_class"
      i: 0
      type: INT
    }
    attribute {
      name: "post_transform"
      s: "LOGISTIC"
      type: STRING
    }
    domain: "ai.onnx.ml"
  }
  node {
    input: "label"
    output: "output_label"
    name: "Cast"
    op_type: "Cast"
    attribute {
      name: "to"
      i: 7
      type: INT
    }
    domain: ""
  }
  node {
    input: "probabilities"
    output: "output_probability"
    name: "ZipMap"
    op_type: "ZipMap"
    attribute {
      name: "classlabels_int64s"
      ints: 0
      ints: 1
      type: INTS
    }
    domain: "ai.onnx.ml"
  }
  name: "pipeline_titanic"
  initializer {
    dims: 2
    data_type: 7
    int64_data: -1
    int64_data: 9
    name: "shape_tensor"
  }
  input {
    name: "pclass"
    type {
      tensor_type {
        elem_type: 8
        shape {
          dim {
          }
          dim {
            dim_value: 1
          }
        }
      }
    }
  }
  input {
    name: "sex"
    type {
      tensor_type {
        elem_type: 8
        shape {
          dim {
          }
          dim {
            dim_value: 1
          }
        }
      }
    }
  }
  input {
    name: "age"
    type {
      tensor_type {
        elem_type: 1
        shape {
          dim {
          }
          dim {
            dim_value: 1
          }
        }
      }
    }
  }
  input {
    name: "fare"
    type {
      tensor_type {
        elem_type: 1
        shape {
          dim {
          }
          dim {
            dim_value: 1
          }
        }
      }
    }
  }
  input {
    name: "embarked"
    type {
      tensor_type {
        elem_type: 8
        shape {
          dim {
          }
          dim {
            dim_value: 1
          }
        }
      }
    }
  }
  output {
    name: "output_label"
    type {
      tensor_type {
        elem_type: 7
        shape {
          dim {
          }
        }
      }
    }
  }
  output {
    name: "output_probability"
    type {
      sequence_type {
        elem_type {
          map_type {
            key_type: 7
            value_type {
              tensor_type {
                elem_type: 1
              }
            }
          }
        }
      }
    }
  }
}
opset_import {
  domain: ""
  version: 11
}
opset_import {
  domain: "ai.onnx.ml"
  version: 1
}

textual features [[1. 0. 0. 0. 0. 1. 1. 0. 0.]]

显示子 ONNX 图

最后，让我们看看两个子图。首先，数值管道。

pydot_graph = GetPydotGraph(
    num_onnx.graph,
    name=num_onnx.graph.name,
    rankdir="TB",
    node_producer=GetOpNodeProducer(
        "docstring", color="yellow", fillcolor="yellow", style="filled"
    ),
)
pydot_graph.write_dot("pipeline_titanic_num.dot")

os.system("dot -O -Gdpi=300 -Tpng pipeline_titanic_num.dot")

image = plt.imread("pipeline_titanic_num.dot.png")
fig, ax = plt.subplots(figsize=(40, 20))
ax.imshow(image)
ax.axis("off")
plot intermediate outputs
(np.float64(-0.5), np.float64(1229.5), np.float64(2558.5), np.float64(-0.5))

然后是文本管道。

pydot_graph = GetPydotGraph(
    text_onnx.graph,
    name=text_onnx.graph.name,
    rankdir="TB",
    node_producer=GetOpNodeProducer(
        "docstring", color="yellow", fillcolor="yellow", style="filled"
    ),
)
pydot_graph.write_dot("pipeline_titanic_text.dot")

os.system("dot -O -Gdpi=300 -Tpng pipeline_titanic_text.dot")

image = plt.imread("pipeline_titanic_text.dot.png")
fig, ax = plt.subplots(figsize=(40, 20))
ax.imshow(image)
ax.axis("off")
plot intermediate outputs
(np.float64(-0.5), np.float64(5630.5), np.float64(2735.5), np.float64(-0.5))

此示例使用的版本

print("numpy:", numpy.__version__)
print("scikit-learn:", sklearn.__version__)
print("onnx: ", onnx.__version__)
print("onnxruntime: ", rt.__version__)
print("skl2onnx: ", skl2onnx.__version__)

numpy: 2.3.1
scikit-learn: 1.6.1
onnx:  1.19.0
onnxruntime:  1.23.0
skl2onnx:  1.19.1

脚本总运行时间： (0 分钟 4.306 秒)

Sphinx-Gallery 生成的图库