PyCaretのモデルをONNXに変換すーる（回帰分析、Mac）

はじめに

PyCaretで作成したモデルをONNXに変換しまーす

開発環境

• MacBookPro 2018
• Python 3.8

導入

1.anaconda環境作成

conda create -n py38 python=3.8

2.ライブラリのインストール

pip install pycaret
pip install skl2onnx
pip install onnxruntime==1.10.0
pip install numpy==1.20.3

brew install libomp

3.insurance.csvを回帰分析し、モデルをonnxに変換します

※簡単にage,bmi,children,chargesのみにしています

pycaret2onnx.py

from pycaret.regression import *
import pandas as pd
from skl2onnx import to_onnx

df = pd.read_csv("insurance.csv")
s = setup(df, target = 'charges')

best = compare_models()
# save_model(best, 'insurance') # insurance.pkl

X_sample = get_config('X_train')[:1]
print(X_sample)

onnx_model = to_onnx(best, X_sample.to_numpy(), target_opset={'':15, 'ai.onnx.ml':2})

with open("insurance.onnx", "wb") as f:
    f.write(onnx_model.SerializeToString())

4.変換したONNXを用いて、Pythonで推論します

predict.py

import onnxruntime as rt
import numpy as np

session = rt.InferenceSession("insurance.onnx")

print(session.get_inputs()[0].name) # X
print(session.get_outputs()[0].name) # variable

# age, bmi, children_0, children_1, children_2, children_3, children_4, children_5
inputs = np.array([[19.0,27.9,1.0,0.0,0.0,0.0,0.0,0.0]],dtype=np.float32)
outputs = session.run(["variable"],{"X": inputs})[0]
print(outputs[0][0])

5.C#で推論します

using Microsoft.ML.OnnxRuntime;
using Microsoft.ML.OnnxRuntime.Tensors;

Console.WriteLine(Directory.GetCurrentDirectory());
var session = new InferenceSession("../../../insurance.onnx");

Tensor<float> input = new DenseTensor<float>(new[] { 1, 8 });
input[0, 0] = 19.0f;
input[0, 1] = 27.9f;
input[0, 2] = 1.0f;
input[0, 3] = 0.0f;
input[0, 4] = 0.0f;
input[0, 5] = 0.0f;
input[0, 6] = 0.0f;
input[0, 7] = 0.0f;

var inputs = new List<NamedOnnxValue>() {
    NamedOnnxValue.CreateFromTensor<float>("X",input)
};

var results = session.Run(inputs);
Console.WriteLine(results.First().AsTensor<float>()[0]);

参考文献