TensorFlow、Keras、Flask部署问题：无法将糖尿病预测CNN模型转为Web应用

Hey there! I’ve helped folks work through similar Flask-Keras deployment snags before, so let’s walk through the most common pitfalls and fixes to get your diabetes risk prediction model up and running on the web smoothly.

Common Issues & Fixes #

1. Model Loading Headaches #

Loading your .hdf5 model in Flask often trips people up—especially if you’re using production servers like Gunicorn. The key is to load the model once when the app starts, not on every request. Here’s a safe implementation:

from flask import Flask, request, jsonify
from tensorflow.keras.models import load_model
import numpy as np

app = Flask(__name__)
model = None

# Load model only once before the first request
@app.before_first_request
def load_prediction_model():
    global model
    # If you used custom Keras layers, add custom_objects={"YourLayer": YourLayer} here
    model = load_model('diabetes_risk_model.hdf5')

@app.route('/predict', methods=['POST'])
def predict():
    if not model:
        return jsonify({"error": "Model not loaded"}), 500
    
    # Grab input data from request
    try:
        data = request.get_json()
        health_metrics = data['health_metrics']
    except KeyError:
        return jsonify({"error": "Missing 'health_metrics' in request"}), 400
    
    # Preprocess data (match exactly what you did during training!)
    processed_data = preprocess_input(health_metrics)
    # Run prediction
    risk_score = model.predict(processed_data)[0][0]
    
    return jsonify({"diabetes_risk": float(risk_score)})

def preprocess_input(metrics):
    # Example: Reshape to match your CNN's input shape (adjust to your model's requirements)
    # If your model expects (batch_size, 10, 1, 1) for 10 health metrics:
    return np.array(metrics).reshape(1, 10, 1, 1)

Pro tip: If you see AttributeError about missing custom layers, pass your custom layer definitions to load_model() using the custom_objects parameter.

2. Mismatched Data Preprocessing #

This is the #1 reason predictions differ between local and Flask. Double-check these:

• Are you scaling/normalizing input data exactly the same way as training? Save your scaler (e.g., StandardScaler) with joblib during training, then load it in Flask to apply the same transformations.
• Does the input shape match what your CNN expects? For example, if your model was trained on (batch_size, 8, 1) for 8 health metrics, make sure Flask reshapes single inputs to (1, 8, 1).

3. Request/Response Format Misalignment #

Make sure your Flask endpoint is parsing input correctly:

• Ensure clients send POST requests with Content-Type: application/json
• Verify the JSON keys match what your code expects (e.g., if your frontend sends {"health_metrics": [7.2, 130, 75, ...]}, your code should pull data['health_metrics']).

4. Dependency Version Conflicts #

Local and Flask environments might have different TensorFlow/Keras versions, which can break model loading. Run these commands in both environments to check:

pip list | grep tensorflow
pip list | grep keras

Pin versions in a requirements.txt file for deployment to avoid mismatches:

flask==2.3.3
tensorflow==2.15.0
numpy==1.24.3
joblib==1.3.2
# Add any other dependencies you used

Quick Debugging Tips #

• Add print statements in your Flask route to log input data, processed data shape, and errors.
• Test the endpoint locally with curl to isolate issues:

curl -X POST -H "Content-Type: application/json" -d '{"health_metrics": [6.1, 120, 70, 25, 100, 22.5, 0.5, 30]}' http://localhost:5000/predict

If you share specific error messages or snippets of your current app.py code, I can help narrow things down even further!

内容的提问来源于stack exchange，提问作者LucyDrops