Machine Learning Academy · 课时

使用 joblib 保存和加载管道

您将使用 joblib.dump 将拟合好的 Pipeline 序列化到磁盘，并在全新的 Python 会话中重新加载它，无需重新训练即可进行预测。

第 4 / 4 课13 个步骤

使用 joblib 保存和加载管道 是 CoddyKit 上的免费 Machine Learning Academy 课时。这是第 4 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 Persist a Trained Pipeline?

Training a machine learning pipeline can take minutes or hours. Once fitted, you want to save it to disk so you can reload it later for predictions without re-training. Persistence is also essential for deployment: you train on a development machine and serve predictions on a production server. The saved file must include both the preprocessing steps and the model weights.

Two Serialisation Options: pickle and joblib

Python's built-in pickle module can serialise any Python object, including sklearn pipelines. joblib is a third-party library (bundled with scikit-learn) that is generally preferred for ML objects because it is more efficient for large NumPy arrays — using memory mapping instead of copying — and can compress the output file automatically.

import pickle
import joblib

# Both approaches work; joblib is recommended for sklearn objects
print('pickle version:', pickle.HIGHEST_PROTOCOL)
import sklearn
print('sklearn version:', sklearn.__version__)

Saving with joblib.dump

joblib.dump(obj, filename) serialises the pipeline to a file. You can optionally set compress=3 to use zlib compression (levels 1-9; 3 balances speed and size). The function returns a list of files created. For most pipelines, a single .pkl or .joblib file is created.

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

X, y = load_iris(return_X_y=True)

pipe = Pipeline([
    ('scaler', StandardScaler()),
    ('clf', LogisticRegression(C=1.0, max_iter=200))
])
pipe.fit(X, y)

# Save
joblib.dump(pipe, '/tmp/iris_pipeline.joblib')
print('Pipeline saved!')

Loading with joblib.load

joblib.load(filename) deserialises the pipeline back into a Python object. The loaded pipeline is identical to the original: it has the same fitted scaler parameters (mean, variance) and the same model weights. You can call predict, predict_proba, or score immediately without re-fitting.

import joblib
from sklearn.datasets import load_iris
import numpy as np

X, y = load_iris(return_X_y=True)

# Load the saved pipeline
loaded_pipe = joblib.load('/tmp/iris_pipeline.joblib')

# Predict and verify
predictions = loaded_pipe.predict(X[:5])
print('Predictions:', predictions)
print('Test accuracy:', loaded_pipe.score(X, y).round(4))

Verifying Round-Trip Fidelity

After loading, confirm that the loaded pipeline produces identical predictions to the original. Any mismatch indicates a serialisation bug or a version incompatibility. A simple check is to compare predictions element-wise using np.array_equal.

import joblib
import numpy as np
from sklearn.datasets import load_iris

X, _ = load_iris(return_X_y=True)

# Reload and compare
loaded = joblib.load('/tmp/iris_pipeline.joblib')

# Reload the original reference predictions
# (in practice, save original predictions before reload)
original_preds = loaded.predict(X)  # use loaded as reference
loaded2 = joblib.load('/tmp/iris_pipeline.joblib')
reloaded_preds = loaded2.predict(X)

print('Predictions match:', np.array_equal(original_preds, reloaded_preds))

Compression Options in joblib

Large pipelines (e.g., with RandomForest of 1000 trees) can be hundreds of MB. Use joblib.dump(pipe, path, compress=3) to compress on the fly. Alternatively, specify the compressor explicitly: compress=('zlib', 3) or compress=('lz4', 1) for maximum speed. LZ4 is the fastest; zlib gives smaller files but is slower.

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

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

# Uncompressed
joblib.dump(pipe, '/tmp/pipe_raw.joblib')
# Compressed
joblib.dump(pipe, '/tmp/pipe_compressed.joblib', compress=3)

print('Raw size:', os.path.getsize('/tmp/pipe_raw.joblib'), 'bytes')
print('Compressed size:', os.path.getsize('/tmp/pipe_compressed.joblib'), 'bytes')

Using pickle as an Alternative

If joblib is not available, pickle works for sklearn pipelines. Use binary mode ('rb'/'wb') when opening the file. For small models or scripted one-off tools, pickle is perfectly fine; for production systems handling large NumPy arrays, joblib is strongly preferred.

import pickle
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import load_iris

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

# Save with pickle
with open('/tmp/model.pkl', 'wb') as f:
    pickle.dump(pipe, f)

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

print('Loaded score:', loaded.score(X, y).round(4))

Version Compatibility Warnings

A critical production concern: a pipeline pickled with scikit-learn 1.2 may not load correctly in scikit-learn 1.5. Always record the library versions used at training time in a metadata file alongside the saved model. Use pip freeze > requirements.txt or record versions programmatically and store them next to the model file.

import sklearn
import numpy as np
import json
import os

metadata = {
    'sklearn_version': sklearn.__version__,
    'numpy_version': np.__version__,
    'model_file': 'iris_pipeline.joblib'
}

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

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

Loading a Pipeline in a Production Script

In a production service, the workflow is: load the pipeline once at startup (not per request), receive input features, preprocess with the pipeline's built-in transforms, and return predictions. Because the pipeline includes all preprocessing, the serving code does not need to know about scaling, encoding, or PCA — all of that is encapsulated in the saved object.

import joblib
import numpy as np

# At startup (once)
model = joblib.load('/tmp/iris_pipeline.joblib')

def predict(sepal_length, sepal_width, petal_length, petal_width):
    features = np.array([[sepal_length, sepal_width, petal_length, petal_width]])
    label = model.predict(features)[0]
    proba = model.predict_proba(features)[0]
    return {'label': int(label), 'confidence': round(float(proba.max()), 4)}

result = predict(5.1, 3.5, 1.4, 0.2)
print('Prediction result:', result)

Security Considerations with Pickled Models

Never load a pickle file from an untrusted source. Pickle files can execute arbitrary code on loading — this is a fundamental Python security constraint. For sharing models externally, consider format-specific safer alternatives: ONNX for cross-framework serialisation, or joblib files shared only within trusted infrastructure. Always verify the file checksum before loading.

import hashlib

def file_sha256(path):
    h = hashlib.sha256()
    with open(path, 'rb') as f:
        for chunk in iter(lambda: f.read(65536), b''):
            h.update(chunk)
    return h.hexdigest()

checksum = file_sha256('/tmp/iris_pipeline.joblib')
print('Model SHA-256:', checksum)

# In production: compare this checksum with the one stored in your model registry

Quick Load-Predict Sanity Test

A final best practice: include a short sanity test script in your model package that loads the pipeline, runs a known input, and asserts the expected output. Run this test in your CI/CD pipeline every time the model is promoted to production, confirming the file was not corrupted and the environment is compatible.

import joblib
import numpy as np

# Sanity test
model = joblib.load('/tmp/iris_pipeline.joblib')

# Known input (setosa): sepal_length=5.1, sepal_width=3.5, petal_length=1.4, petal_width=0.2
X_test = np.array([[5.1, 3.5, 1.4, 0.2]])
pred = model.predict(X_test)[0]

# Iris class 0 = setosa
assert pred == 0, f'Expected setosa (0) but got {pred}'
print('Sanity test PASSED — model predicts setosa correctly.')

Quick Check

Test your understanding of saving and loading pipelines from this lesson.

在交互式课程中回答

Lesson Recap

In this lesson you learned: joblib.dump and joblib.load save and restore a complete fitted pipeline including all preprocessing parameters, always record library versions alongside the saved model to ensure reproducible loading, and never load pickle files from untrusted sources as they can execute arbitrary code. Next up we tackle imbalanced datasets — detecting class imbalance and understanding why accuracy is a misleading metric in that setting.

免费开始

用 AI 导师学习 Python — 免费

在浏览器中编写并运行真实代码，获得全天候 AI 导师的即时帮助，并在网页或应用中继续学习。

课程: 30
课程: 120

免费开始学习查看完整课程

常见问题解答

「使用 joblib 保存和加载管道」课时是免费的吗？

是的 — 「使用 joblib 保存和加载管道」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 保存和加载管道」这节课中我会学到什么？

您将使用 joblib.dump 将拟合好的 Pipeline 序列化到磁盘，并在全新的 Python 会话中重新加载它，无需重新训练即可进行预测。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 4 节课，共 4 节。

「使用 joblib 保存和加载管道」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