Saving, Deploying, and Using Classification Models

# CHAPTER 19

Saving, Deploying, and Using Classification Models #

1. Introduction #

2. Learning Objectives #

• Explain what model serialization means.

• Save a trained Scikit-Learn Pipeline using joblib.

• Load a saved model from the hard drive.

• Understand the architecture of Model Deployment.

• Build a basic Flask REST API to serve predictions.

3. Serialization (joblib vs pickle) #

4. Saving the Model (And the Pipeline!) #

import joblib
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
import numpy as np

# 1. Mock Training Data (Income, Debt) -> 1=Default, 0=Safe
X_train = np.array([[50000, 10000], [100000, 5000], [30000, 25000]])
y_train = np.array([0, 0, 1])

# 2. Build and Train the Pipeline
final_pipeline = make_pipeline(StandardScaler(), LogisticRegression())
final_pipeline.fit(X_train, y_train)

# 3. SAVE THE PIPELINE!
# This creates a file named 'fraud_model.pkl' in your folder.
joblib.dump(final_pipeline, 'fraud_model.pkl')
print("Model successfully saved to disk!")

5. Loading the Model (In a New File) #

import joblib
import numpy as np

# 1. Load the binary file back into RAM
loaded_pipeline = joblib.load('fraud_model.pkl')

# 2. Get user input from the website (Income: $40k, Debt: $30k)
user_transaction = np.array([[40000, 30000]])

# 3. Make the prediction!
# The pipeline AUTOMATICALLY applies the StandardScaler to the user input!
prediction = loaded_pipeline.predict(user_transaction)

result = "FRAUD" if prediction[0] == 1 else "SAFE"
print(f"Transaction Status: {result}")

6. Deployment Architecture (REST API) #

1. 1. The Web Server (using Python's Flask or FastAPI framework) loads the joblib file into memory when it boots up.

1. 2. A user makes a transaction on the website. The site sends a JSON package ({"income": 40000, "debt": 30000}) via HTTP POST to your Python server.

1. 3. The server extracts the numbers, converts them to a NumPy array, passes them through the loaded pipeline, and gets the prediction.

1. 4. The server sends the prediction back as JSON ({"status": "FRAUD"}).

7. Mini Project: A Simple Flask API #

from flask import Flask, request, jsonify
import joblib
import numpy as np

app = Flask(__name__)

# Load the model ONCE when the server boots
model = joblib.load('fraud_model.pkl')

@app.route('/predict', methods=['POST'])
def predict():
    # 1. Receive JSON data from the website
    data = request.get_json()
    
    # 2. Extract features
    income = data['income']
    debt = data['debt']
    
    # 3. Format into a 2D array for scikit-learn
    features = np.array([[income, debt]])
    
    # 4. Predict
    prediction = model.predict(features)
    
    # 5. Return JSON response to the user
    return jsonify({
        'status': 'success',
        'is_fraud': int(prediction[0])
    })

if __name__ == '__main__':
    # Run the web server!
    app.run(port=5000, debug=True)

8. Common Mistakes #

• Version Mismatch: If you train and save the model using scikit-learn version 1.2 on your laptop, but the production web server has scikit-learn version 0.24, joblib.load() will likely crash. The library versions must match exactly! Use a requirements.txt file or Docker to enforce this.

• Missing Columns: The user must provide the exact same number of columns, in the exact same order, that the model was trained on.

9. Best Practices #

• Never retrain in production: The web server should *only* execute model.predict(). It should never execute model.fit(). Model training happens offline. The resulting .pkl file is then uploaded to the server.

10. Exercises #

1. 1. What is the fundamental difference in purpose between model.fit() and joblib.dump()?

1. 2. Write the line of code required to load a classification model named spamfilter.pkl into a variable named mymodel.

11. MCQ Quiz with Answers #

12. Interview Questions #

• Q: Describe the end-to-end architecture of how a user on a website receives a classification prediction from a Scikit-Learn model running on a remote server.

• Q: Explain why a library version mismatch between the training environment and the production environment is catastrophic for a pickled/joblib model.

13. FAQs #

14. Summary #

15. Next Chapter Recommendation #