ONNX 转换基准测试

示例 训练和部署 scikit-learn 管道 转换了一个简单的模型。此示例采用类似的示例，但使用随机数据，并比较每个选项计算预测所需的时间。

训练管道

import numpy
from pandas import DataFrame
from tqdm import tqdm
from onnx.reference import ReferenceEvaluator
from sklearn import config_context
from sklearn.datasets import make_regression
from sklearn.ensemble import (
    GradientBoostingRegressor,
    RandomForestRegressor,
    VotingRegressor,
)
from sklearn.linear_model import LinearRegression
from sklearn.model_selection import train_test_split
from onnxruntime import InferenceSession
from skl2onnx import to_onnx
from skl2onnx.tutorial import measure_time


N = 11000
X, y = make_regression(N, n_features=10)
X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.01)
print("Train shape", X_train.shape)
print("Test shape", X_test.shape)

reg1 = GradientBoostingRegressor(random_state=1)
reg2 = RandomForestRegressor(random_state=1)
reg3 = LinearRegression()
ereg = VotingRegressor([("gb", reg1), ("rf", reg2), ("lr", reg3)])
ereg.fit(X_train, y_train)

Train shape (110, 10)
Test shape (10890, 10)

VotingRegressor(estimators=[('gb', GradientBoostingRegressor(random_state=1)),
                            ('rf', RandomForestRegressor(random_state=1)),
                            ('lr', LinearRegression())])

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。

测量处理时间

我们使用函数 skl2onnx.tutorial.measure_time()。关于 assume_finite 的页面可能对您很有用，如果您需要优化预测。我们测量每个观测值的处理时间，无论该观测值是否属于批次或单个观测值。

sizes = [(1, 50), (10, 50), (100, 10)]

with config_context(assume_finite=True):
    obs = []
    for batch_size, repeat in tqdm(sizes):
        context = {"ereg": ereg, "X": X_test[:batch_size]}
        mt = measure_time(
            "ereg.predict(X)", context, div_by_number=True, number=10, repeat=repeat
        )
        mt["size"] = context["X"].shape[0]
        mt["mean_obs"] = mt["average"] / mt["size"]
        obs.append(mt)

df_skl = DataFrame(obs)
df_skl

  0%|          | 0/3 [00:00<?, ?it/s]
 33%|███▎      | 1/3 [00:07<00:14,  7.06s/it]
 67%|██████▋   | 2/3 [00:12<00:06,  6.25s/it]
100%|██████████| 3/3 [00:14<00:00,  4.01s/it]
100%|██████████| 3/3 [00:14<00:00,  4.70s/it]

	平均值	偏差	最小执行时间	最大执行时间	重复次数	数量	大小	平均观测值
0	0.014108	0.004352	0.008928	0.029686	50	10	1	0.014108
1	0.011358	0.003103	0.008228	0.020336	50	10	10	0.001136
2	0.013486	0.002790	0.009885	0.018525	10	10	100	0.000135

图形。

df_skl.set_index("size")[["mean_obs"]].plot(title="scikit-learn", logx=True, logy=True)

ONNX 运行时

对两个可用的 ONNX 运行时执行相同的操作。

onx = to_onnx(ereg, X_train[:1].astype(numpy.float32), target_opset=14)
sess = InferenceSession(onx.SerializeToString(), providers=["CPUExecutionProvider"])
oinf = ReferenceEvaluator(onx)

obs = []
for batch_size, repeat in tqdm(sizes):
    # scikit-learn
    context = {"ereg": ereg, "X": X_test[:batch_size].astype(numpy.float32)}
    mt = measure_time(
        "ereg.predict(X)", context, div_by_number=True, number=10, repeat=repeat
    )
    mt["size"] = context["X"].shape[0]
    mt["skl"] = mt["average"] / mt["size"]

    # onnxruntime
    context = {"sess": sess, "X": X_test[:batch_size].astype(numpy.float32)}
    mt2 = measure_time(
        "sess.run(None, {'X': X})[0]",
        context,
        div_by_number=True,
        number=10,
        repeat=repeat,
    )
    mt["ort"] = mt2["average"] / mt["size"]

    # ReferenceEvaluator
    context = {"oinf": oinf, "X": X_test[:batch_size].astype(numpy.float32)}
    mt2 = measure_time(
        "oinf.run(None, {'X': X})[0]",
        context,
        div_by_number=True,
        number=10,
        repeat=repeat,
    )
    mt["pyrt"] = mt2["average"] / mt["size"]

    # end
    obs.append(mt)


df = DataFrame(obs)
df

  0%|          | 0/3 [00:00<?, ?it/s]
 33%|███▎      | 1/3 [00:15<00:31, 15.60s/it]
 67%|██████▋   | 2/3 [00:40<00:21, 21.10s/it]
100%|██████████| 3/3 [01:03<00:00, 21.84s/it]
100%|██████████| 3/3 [01:03<00:00, 21.09s/it]

	平均值	偏差	最小执行时间	最大执行时间	重复次数	数量	大小	skl	ort	pyrt
0	0.012201	0.003428	0.008608	0.021660	50	10	1	0.012201	0.000030	0.018960
1	0.011550	0.003386	0.008357	0.023143	50	10	10	0.001155	0.000021	0.003811
2	0.013936	0.004899	0.009295	0.023858	10	10	100	0.000139	0.000004	0.002129

图形。

df.set_index("size")[["skl", "ort", "pyrt"]].plot(
    title="Average prediction time per runtime", logx=True, logy=True
)

ONNX 运行时比 scikit-learn 预测单个观测值快得多。 scikit-learn 针对训练和批量预测进行了优化。这解释了为什么 scikit-learn 和 ONNX 运行时对于大型批次似乎会收敛。它们使用类似的实现、并行化和语言（C++，openmp）。