Machine Learning Academy · 课时

使用 joblib 和 pickle 保存模型

您将使用 joblib 和 pickle 分别序列化训练好的管道，将其重新加载，并验证预测结果完全一致，以确认往返过程成功。

使用 joblib 和 pickle 保存模型 是 CoddyKit 上的免费 Machine Learning Academy 课时。这是第 1 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 Machine Learning Academy 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Machine Learning Academy 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

Why Model Persistence Matters

Training a machine learning model is expensive: it can take minutes to hours and consumes significant compute. Model persistence saves the fitted model to disk so you can reload it instantly for inference without retraining. This is the bridge between the data science notebook and a production system — the serialised model file is the deployable artefact that data engineers package and serve.

What Gets Saved in a Model File?

When you serialise a fitted sklearn model or pipeline, the file captures: all fitted parameters (e.g., scaler mean and variance, tree structure, logistic regression coefficients), hyperparameter settings, and the Python class definition reference. It does NOT include the training data. Loading the file reconstructs a Python object ready to call predict immediately.

Saving with joblib.dump

joblib is the recommended serialisation tool for sklearn objects. It handles large NumPy arrays efficiently using memory mapping and supports transparent compression. The standard workflow is: train the model, dump it to a .joblib file, then load it in a separate script or service for inference.

import joblib
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import load_breast_cancer

X, y = load_breast_cancer(return_X_y=True)

pipe = Pipeline([
    ('scaler', StandardScaler()),
    ('clf', LogisticRegression(C=1.0, max_iter=300))
])
pipe.fit(X, y)

# Save
joblib.dump(pipe, '/tmp/cancer_model.joblib')
print('Model saved to /tmp/cancer_model.joblib')

Loading with joblib.load

joblib.load deserialises the file and returns the exact fitted pipeline object. The loaded model has all of the same attributes — named_steps, fitted scaler parameters, classifier coefficients — as the original. You can immediately call predict, predict_proba, or score without any additional setup.

import joblib
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split

X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0)

# Load in a fresh context
model = joblib.load('/tmp/cancer_model.joblib')

predictions = model.predict(X_test[:5])
print('Predictions:', predictions)
print('Test accuracy:', model.score(X_test, y_test).round(4))

Using pickle for Serialisation

Python's standard library pickle module also serialises sklearn objects. Open files in binary mode ('wb' for write, 'rb' for read). The pickle.HIGHEST_PROTOCOL constant uses the most efficient available protocol. For small models or scripting contexts, pickle is perfectly adequate.

import pickle
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import load_iris

X, y = load_iris(return_X_y=True)
model = LogisticRegression().fit(X, y)

# Save
with open('/tmp/iris_model.pkl', 'wb') as f:
    pickle.dump(model, f, protocol=pickle.HIGHEST_PROTOCOL)

# Load
with open('/tmp/iris_model.pkl', 'rb') as f:
    loaded = pickle.load(f)

print('Score:', loaded.score(X, y).round(4))
print('Coefficients shape:', loaded.coef_.shape)

Comparing joblib vs pickle File Sizes

For a large model like a RandomForest with 1000 trees, joblib's memory-mapped NumPy array storage is more efficient. The difference becomes especially pronounced when the model contains large parameter matrices. For small models (LogReg, SVM), the size difference is negligible.

import joblib
import pickle
import os
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import make_classification

X, y = make_classification(n_samples=1000, n_features=20, random_state=0)
rf = RandomForestClassifier(n_estimators=100, random_state=0).fit(X, y)

# joblib
joblib.dump(rf, '/tmp/rf_model.joblib')

# pickle
with open('/tmp/rf_model.pkl', 'wb') as f:
    pickle.dump(rf, f)

print(f'joblib size: {os.path.getsize("/tmp/rf_model.joblib"):,} bytes')
print(f'pickle size: {os.path.getsize("/tmp/rf_model.pkl"):,} bytes')

Compression with joblib

Use joblib.dump(model, path, compress=3) to compress the file using zlib. Compression levels range from 1 (fast, larger) to 9 (slow, smallest). Level 3 is a practical default. For LZ4 compression (faster than zlib): compress=('lz4', 1). Load time slightly increases for compressed files but the network transfer and storage savings are often worth it.

import joblib
import os
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import make_classification

X, y = make_classification(n_samples=1000, random_state=0)
rf = RandomForestClassifier(n_estimators=100, random_state=0).fit(X, y)

for level in [0, 3, 6, 9]:
    path = f'/tmp/rf_compress_{level}.joblib'
    joblib.dump(rf, path, compress=level)
    size = os.path.getsize(path)
    print(f'compress={level}: {size:,} bytes')

Verifying Round-Trip Consistency

After loading, always verify that the loaded model produces identical predictions to the original. This guards against silent corruption, version mismatches, or incomplete file writes. Compare predictions with np.array_equal on the same input.

import joblib
import numpy as np
from sklearn.datasets import load_iris
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression

X, y = load_iris(return_X_y=True)
pipe = Pipeline([('sc', StandardScaler()), ('lr', LogisticRegression())]).fit(X, y)
original_preds = pipe.predict(X)

joblib.dump(pipe, '/tmp/verify_pipe.joblib')
loaded = joblib.load('/tmp/verify_pipe.joblib')
loaded_preds = loaded.predict(X)

if np.array_equal(original_preds, loaded_preds):
    print('Round-trip PASSED: predictions are identical.')
else:
    diff = (original_preds != loaded_preds).sum()
    print(f'Round-trip FAILED: {diff} different predictions!')

Version Pinning and Metadata

A model pickled with scikit-learn 1.2 may not load cleanly in scikit-learn 1.5 due to internal changes. Always store a metadata file alongside the model that records: sklearn version, Python version, training date, dataset version, and key metrics. This is your model card — the governance document that explains what the model is and how it was produced.

import json
import sklearn
import sys
from datetime import datetime

metadata = {
    'model_file': 'cancer_model.joblib',
    'sklearn_version': sklearn.__version__,
    'python_version': sys.version.split()[0],
    'training_date': datetime.utcnow().isoformat(),
    'dataset': 'breast_cancer',
    'test_accuracy': 0.9789,
    'features': 30,
    'algorithm': 'LogisticRegression'
}

with open('/tmp/cancer_model_metadata.json', 'w') as f:
    json.dump(metadata, f, indent=2)

print(json.dumps(metadata, indent=2))

Security: Never Load Untrusted Pickle Files

Critical security warning: both pickle and joblib can execute arbitrary Python code when loading. Never load a model file from an untrusted source — it could be a malicious payload disguised as a model. For models shared across organisations, consider safer formats: ONNX (Open Neural Network Exchange) is a standardised, inspectable format supported by most frameworks.

File Naming Conventions

Good naming conventions embed key information into the filename: algorithm, dataset, date, and metric. This makes the model registry self-documenting and prevents accidentally loading the wrong model version in production.

from datetime import date
from sklearn.metrics import accuracy_score
import joblib

# Example naming convention
dataset = 'breast_cancer'
algorithm = 'logreg'
test_acc = 0.9789
today = date.today().strftime('%Y%m%d')

filename = f'{dataset}_{algorithm}_{today}_acc{int(test_acc*100)}.joblib'
print('Model filename:', filename)
# e.g.: breast_cancer_logreg_20260620_acc97.joblib

# Load the model we saved earlier (demo)
model = joblib.load('/tmp/cancer_model.joblib')
print('Loaded OK')

Quick Check

Test your understanding of model serialisation with joblib and pickle from this lesson.

Lesson Recap

In this lesson you learned: joblib.dump and joblib.load serialise and restore fitted sklearn models efficiently, pickle works too but joblib is preferred for models with large NumPy arrays due to memory mapping, and always store a metadata file alongside the model recording library versions, training date, and key metrics for governance. Next up we design a versioning naming convention and metadata sidecar to track multiple model versions in a model registry.

常见问题解答

「使用 joblib 和 pickle 保存模型」课时是免费的吗？

是的 — 「使用 joblib 和 pickle 保存模型」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Machine Learning Academy 课程的其余内容，请升级到 CoddyKit PRO。 Machine Learning Academy 课程共包含 4 节课。

「使用 joblib 和 pickle 保存模型」这节课中我会学到什么？

您将使用 joblib 和 pickle 分别序列化训练好的管道，将其重新加载，并验证预测结果完全一致，以确认往返过程成功。你通过在浏览器中直接运行的动手代码来练习 Machine Learning Academy，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 Machine Learning Academy 需要有经验吗？

无需任何先前经验。CoddyKit 上的 Machine Learning Academy 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 1 节课，共 4 节。

「使用 joblib 和 pickle 保存模型」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 Machine Learning Academy 课中编写并运行代码吗？

能。每节 Machine Learning Academy 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 Machine Learning Academy