Model Serialization: Joblib, Pickle & ONNX Runtime

Once an algorithm has converged after hours or days of training, you must save its learned mathematical weights to persistent storage. Model Serialization converts in-memory model objects into standardized byte streams that can be stored, versioned, and transferred across network boundaries for inference.

1. The Danger of Pickle: Remote Code Execution (RCE)

Standard Python uses the built-in pickle module (and its wrapper joblib) to serialize Python objects. However, pickle suffers from a catastrophic security vulnerability:

2. ONNX (Open Neural Network Exchange): The Universal Standard

Created by Microsoft, Meta, and Amazon, ONNX is an open-source format for representing machine learning models. Instead of storing arbitrary Python bytecode, ONNX compiles the model into a strict, static Computational Graph consisting of standard mathematical operators (MatMul, Add, Relu, Conv).

Exporting and Serving a Scikit-Learn Model with ONNX

from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import load_iris
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType
import onnxruntime as rt
import numpy as np

# 1. Train Model
iris = load_iris()
clf = RandomForestClassifier(n_estimators=50).fit(iris.data, iris.target)

# 2. Convert to ONNX format
initial_type = [('float_input', FloatTensorType([None, 4]))]  # [batch_size, 4 features]
onnx_model = convert_sklearn(clf, initial_types=initial_type)

with open("model.onnx", "wb") as f:
    f.write(onnx_model.SerializeToString())
print("Model successfully compiled to model.onnx!")

# 3. High-Speed Inference with ONNX Runtime (Zero Python overhead!)
session = rt.InferenceSession("model.onnx", providers=['CPUExecutionProvider'])
input_name = session.get_inputs()[0].name

query = np.array([[5.1, 3.5, 1.4, 0.2]], dtype=np.float32)
pred_onx = session.run(None, {input_name: query})

print("Predicted Class:", pred_onx[0][0])
print("Class Probabilities:", pred_onx[1][0])