遍历中间输出

我们重新使用示例 将带有 ColumnTransformer 的管道进行转换 并遍历中间输出。 转换后的模型很有可能产生不同的输出或由于未正确实现的自定义转换器而失败。 一个选项是查看 ONNX 图中每个节点的输出。

创建和训练复杂管道

我们重新使用在示例 带有混合类型的 Column Transformer 中实现的管道。 由于 ONNX-ML Imputer 不处理字符串类型，因此存在一个变化。 这不能是最终 ONNX 管道的一部分，必须删除。 查找以下以 --- 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 表示形式或信任笔记本。
在 GitHub 上，HTML 表示形式无法渲染，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。 我们只需重新使用数据框的列定义。

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]))]

将单个列合并成向量不是计算预测的最高效方法。 这可以在将管道转换为图形之前完成。

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 在可以的情况下执行隐式转换。 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 1 0 0]
predict_proba [[0.60224126 0.39775874]]

使用 onnxruntime 的预测。 我们需要删除已删除的列并将双向量更改为浮点向量，因为 onnxruntime 不支持双精度浮点数。 onnxruntime 不接受 数据框。 输入必须作为字典列表给出。 最后，每个列都被描述为不是真正的向量，而是具有单列的矩阵，这解释了最后一行的 reshape 操作。

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 1 0 0]
predict_proba [{0: 0.7899309396743774, 1: 0.21006903052330017}]

计算中间输出

不幸的是，实际上没有办法让 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。 让我们首先查看中间输出的列表。

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

很难判断哪个是哪个，因为 ONNX 拥有比原始 scikit-learn 管道更多的操作符。 显示 ONNX 图 中的图有助于我们找到数值和文本管道的输出：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\x061.17.0"\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\xe2A=\xcdLgA\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=\xe05\xedA=\'\xcb\nB\xa0\x01\x06*\x14\n\x05scale=\'l\x9f==\xdd,\x96<\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\x0e\n\nai.onnx.ml\x10\x01B\x04\n\x00\x10\x0b'

让我们计算数值特征。

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

numerical features [[-0.7512866  -0.50364053]]

对于文本特征，我们执行相同的操作。

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
opset_import {
  domain: "ai.onnx.ml"
  version: 1
}
opset_import {
  domain: ""
  version: 11
}
producer_name: "skl2onnx"
producer_version: "1.17.0"
domain: "ai.onnx"
model_version: 0
doc_string: ""
graph {
  node {
    input: "age"
    input: "fare"
    output: "merged_columns"
    name: "Concat"
    op_type: "Concat"
    domain: ""
    attribute {
      name: "axis"
      type: INT
      i: 1
    }
  }
  node {
    input: "embarked"
    output: "embarkedout"
    name: "OneHotEncoder"
    op_type: "OneHotEncoder"
    domain: "ai.onnx.ml"
    attribute {
      name: "cats_strings"
      type: STRINGS
      strings: "C"
      strings: "Q"
      strings: "S"
      strings: "missing"
    }
    attribute {
      name: "zeros"
      type: INT
      i: 1
    }
  }
  node {
    input: "sex"
    output: "sexout"
    name: "OneHotEncoder1"
    op_type: "OneHotEncoder"
    domain: "ai.onnx.ml"
    attribute {
      name: "cats_strings"
      type: STRINGS
      strings: "female"
      strings: "male"
    }
    attribute {
      name: "zeros"
      type: INT
      i: 1
    }
  }
  node {
    input: "pclass"
    output: "pclassout"
    name: "OneHotEncoder2"
    op_type: "OneHotEncoder"
    domain: "ai.onnx.ml"
    attribute {
      name: "cats_strings"
      type: STRINGS
      strings: "1"
      strings: "2"
      strings: "3"
    }
    attribute {
      name: "zeros"
      type: INT
      i: 1
    }
  }
  node {
    input: "embarkedout"
    input: "sexout"
    input: "pclassout"
    output: "concat_result"
    name: "Concat1"
    op_type: "Concat"
    domain: ""
    attribute {
      name: "axis"
      type: INT
      i: -1
    }
  }
  node {
    input: "merged_columns"
    output: "variable"
    name: "Imputer"
    op_type: "Imputer"
    domain: "ai.onnx.ml"
    attribute {
      name: "imputed_value_floats"
      type: FLOATS
      floats: 28.25
      floats: 14.4562502
    }
    attribute {
      name: "replaced_value_float"
      type: FLOAT
      f: nan
    }
  }
  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"
    domain: "ai.onnx.ml"
    attribute {
      name: "offset"
      type: FLOATS
      floats: 29.6513062
      floats: 34.698391
    }
    attribute {
      name: "scale"
      type: FLOATS
      floats: 0.077843
      floats: 0.0183319394
    }
  }
  node {
    input: "variable1"
    input: "variable2"
    output: "transformed_column"
    name: "Concat2"
    op_type: "Concat"
    domain: ""
    attribute {
      name: "axis"
      type: INT
      i: 1
    }
  }
  node {
    input: "transformed_column"
    output: "label"
    output: "probabilities"
    name: "LinearClassifier"
    op_type: "LinearClassifier"
    domain: "ai.onnx.ml"
    attribute {
      name: "classlabels_ints"
      type: INTS
      ints: 0
      ints: 1
    }
    attribute {
      name: "coefficients"
      type: FLOATS
      floats: 0.411349356
      floats: -0.0257858913
      floats: -0.341414243
      floats: 0.0805286616
      floats: 0.334271878
      floats: -0.121588431
      floats: -1.24841082
      floats: 1.20020878
      floats: -0.920275748
      floats: -0.037623141
      floats: 0.909696758
      floats: -0.411349356
      floats: 0.0257858913
      floats: 0.341414243
      floats: -0.0805286616
      floats: -0.334271878
      floats: 0.121588431
      floats: 1.24841082
      floats: -1.20020878
      floats: 0.920275748
      floats: 0.037623141
      floats: -0.909696758
    }
    attribute {
      name: "intercepts"
      type: FLOATS
      floats: -0.147927582
      floats: 0.147927582
    }
    attribute {
      name: "multi_class"
      type: INT
      i: 0
    }
    attribute {
      name: "post_transform"
      type: STRING
      s: "LOGISTIC"
    }
  }
  node {
    input: "label"
    output: "output_label"
    name: "Cast"
    op_type: "Cast"
    domain: ""
    attribute {
      name: "to"
      type: INT
      i: 7
    }
  }
  node {
    input: "probabilities"
    output: "output_probability"
    name: "ZipMap"
    op_type: "ZipMap"
    domain: "ai.onnx.ml"
    attribute {
      name: "classlabels_int64s"
      type: INTS
      ints: 0
      ints: 1
    }
  }
  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
              }
            }
          }
        }
      }
    }
  }
}

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

显示子 ONNX 图

最后，让我们看看两个子图。 首先是数值管道。

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")

(-0.5, 1229.5, 2558.5, -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")

(-0.5, 5630.5, 2735.5, -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: 1.26.4
scikit-learn: 1.6.dev0
onnx:  1.17.0
onnxruntime:  1.18.0+cu118
skl2onnx:  1.17.0