App Engine部署的图像预测API无法读取云存储模型求助

已成功将应用部署至App Engine，但调用/predict API时无法完成图像预测，Postman测试返回错误。已配置调用存储模型的Cloud Storage Bucket，但问题仍存在。

app.py #

import os
import numpy as np
import tensorflow as tf
from flask import Flask, request, jsonify
#from tensorflow import keras
#from keras.models import load_model
from PIL import Image
from flask_cors import CORS


app = Flask(__name__)
CORS(app)

# global variable
model = None #model
#db = None

# download model file from cloud storage 
def download_model_file():
    from google.cloud import storage

    # model bucket details
    BUCKET_NAME = "spezia-bucket"
    PROJECT_ID = "capstone-spezia"
    GCS_MODEL_FILE = "spezia_model.h5"

    # initialise a client
    client = storage.Client(PROJECT_ID)

    # create a bucket object for our bucket
    bucket = client.bucket(BUCKET_NAME)

    # create a blob object from the filepath
    blob = bucket.blob(GCS_MODEL_FILE)

    #with blob.open("r") as model:
    #    print(model.read())

    folder = '/tmp/'
    if not os.path.exists(folder) :
        os.makedirs(folder)
    # download the file to a destination
    blob.download_to_filename(folder + "model_spezia.h5")


#model = load_model('spezia_model.h5')

@app.route('/')
def main():
    return 'Welcome to Spezia ML Team API for predict many spices'

@app.route('/predict', methods=['POST'])
def recognize_image():
    try:
        # deploy model
        global model
        if not model:
            download_model_file()
            model = tf.keras.models.load_model('/tmp/model_spezia.h5')

        # open image from request
        img_sample = Image.open(request.files['image'])
        image_path = img_sample
        image_path = image_path.convert('RGB')
        image_path.close()

        # prepare image for prediction
        img = np.array(img_sample.resize((150,150)))
        
        x = np.expand_dims(img, axis=0)
        images = np.vstack([x])
        image_path.close()


        # predict
         prediction_array = model.predict(images)

         class_names =  ['asam jawa', 'cengkeh', 'daun jeruk', 'daun salam', 
                        'jahe', 'kayu manis', 'keluak', 'kemiri', 'ketumbar', 
                        'kunyit', 'lada hitam', 'pekak','serai']
        result = {
            'prediction': class_names[np.argmax(prediction_array)],
            'confidence': '{:2.0f}%'.format(100 * np.max(prediction_array))
        }

        return jsonify(isError=False, message='Success', statusCode=200, data=result), 200

     except Exception as e:
        print(str(e))
        return jsonify(message='Something went wrong'), 500

if __name__ == '__main__':
    app.run(debug=True, port=8080)

requirements.txt #

Flask==2.2.2
Flask-Cors==3.0.10
tensorflow==2.11.0
keras==2.11.0
numpy==1.23.5
Pillow==9.3.0
google-cloud-storage==2.7.0

排查与修复方案 #

1. 图像处理逻辑错误 #

代码中存在提前关闭图像文件的问题：打开图像后将img_sample赋值给image_path，转换为RGB后直接关闭了文件对象，后续尝试使用已关闭的img_sample进行resize操作会触发错误。

修正代码：

# 修正后的图像处理逻辑
img_sample = Image.open(request.files['image'])
img_sample = img_sample.convert('RGB')  # 直接操作原对象
img = np.array(img_sample.resize((150,150)))
img_sample.close()  # 所有处理完成后再关闭文件

同时删除重复的image_path.close()语句。

2. 模型加载验证与权限检查 #

• 为App Engine默认服务账号（格式：PROJECT_ID@appspot.gserviceaccount.com）添加roles/storage.objectViewer角色，确保有权限从Cloud Storage下载模型文件。
• 在模型加载后添加验证代码，确认模型是否正确加载：

if not model:
    download_model_file()
    model = tf.keras.models.load_model('/tmp/model_spezia.h5')
    print("模型输入形状:", model.input_shape)  # 打印模型信息验证加载状态

3. 路径拼接优化 #

使用os.path.join拼接文件路径，避免因分隔符问题导致的文件找不到错误：

blob.download_to_filename(os.path.join(folder, "model_spezia.h5"))

4. 日志增强排查 #

替换简单的print为日志模块输出，方便在App Engine控制台查看详细错误：

import logging

# 异常处理块修改为
except Exception as e:
    logging.error(f"预测错误详情: {str(e)}")
    return jsonify(message='Something went wrong'), 500

5. 运行环境指定 #

在app.yaml中添加Python版本声明，确保环境兼容依赖版本：

runtime: python39

内容的提问来源于stack exchange，提问作者masyara