在Google Cloud Functions中运行Roboflow推理时的问题

我打算通过移动端抓拍图片，将图片发送到运行在Google Cloud Functions上的Python脚本，调用自己训练好的Roboflow目标检测模型做推理。以下是部署在Cloud Functions上的推理脚本：

from flask import Flask, request
from roboflow import Roboflow

app = Flask(__name__)

# 配置Roboflow密钥、模型端点和版本
ROBOFLOW_API_KEY = "XXX"
ROBOFLOW_MODEL_ENDPOINT = "XXX"
ROBOFLOW_VERSION = "X"

# 初始化Roboflow客户端
rf = Roboflow(api_key=ROBOFLOW_API_KEY)
project = rf.workspace().project(ROBOFLOW_MODEL_ENDPOINT)
model = project.version(ROBOFLOW_VERSION).model


@app.route("/upload", methods=["POST"])
def upload(request):
    # 接收移动端传来的图片文件
    image_file = request.files["image"]
    image_data = image_file.read()
    image_filename = image_file.filename

    # 将图片数据转为base64编码字符串
    base64_img_data = base64.b64encode(image_file.read())

    # 调用Roboflow API做推理
    prediction = model.predict(base64_img_data, confidence=40, overlap=30)
    results = prediction.json()

    # 返回推理结果
    return results

if __name__ == "__main__":
    app.run()

用以下代码发送本地图片测试时，Roboflow的classification.py抛出兼容性错误：

# 用multipart/form-data格式发送图片到Cloud Function
response = requests.post(cloud_function_url, files={"image": ("image.jpg", image_data, "image/jpeg")})

错误日志：

Traceback (most recent call last): 
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/flask/app.py", line 2190, in wsgi_app 
response = self.full_dispatch_request() 
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/flask/app.py", line 1486, in full_dispatch_request 
rv = self.handle_user_exception(e) 
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/flask/app.py", line 1484, in full_dispatch_request 
rv = self.dispatch_request() 
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/flask/app.py", line 1469, in dispatch_request 
return self.ensure_sync(self.view_functions[rule.endpoint])(**view_args) 
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/functions_framework/__init__.py", line 99, in view_func 
return function(request._get_current_object()) 
File "/workspace/main.py", line 38, in upload 
prediction = model.predict(base64_img_data) 
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/roboflow/models/classification.py", line 54, in predict 
self.__exception_check(image_path_check=image_path) 
File "/layers/google.python.pip/pip/lib/python3.9/site-packages/roboflow/models/classification.py", line 126, in __exception_check 
raise Exception("Image does not exist at " + image_path_check + "!") 
TypeError: can only concatenate str (not "bytes") to str

1. Base64数据类型错误：base64.b64encode()返回的是bytes类型，但Roboflow的model.predict()方法把传入的bytes当成了文件路径字符串，尝试拼接错误提示时触发类型不兼容报错。
2. 重复读取文件导致空数据：image_file.read()被调用了两次，第一次读取后文件指针移到末尾，第二次读取会得到空bytes，即便类型正确也无法正常推理。
3. Cloud Functions冗余代码：Cloud Functions本身由functions_framework托管，不需要手动创建Flask app和路由装饰器，原代码的Flask配置属于冗余内容。

修改Cloud Functions脚本，解决上述三个问题：

import base64
from roboflow import Roboflow

# 配置Roboflow密钥、模型端点和版本
ROBOFLOW_API_KEY = "XXX"
ROBOFLOW_MODEL_ENDPOINT = "XXX"
ROBOFLOW_VERSION = "X"

# 初始化Roboflow客户端（全局初始化，避免每次请求重复创建）
rf = Roboflow(api_key=ROBOFLOW_API_KEY)
project = rf.workspace().project(ROBOFLOW_MODEL_ENDPOINT)
model = project.version(ROBOFLOW_VERSION).model


def upload(request):
    # 接收图片文件
    image_file = request.files["image"]
    # 只读取一次图片数据
    image_data = image_file.read()
    
    # 将bytes类型的base64编码转为字符串
    base64_img_data = base64.b64encode(image_data).decode("utf-8")
    
    # 调用推理接口，传入字符串格式的base64数据
    prediction = model.predict(base64_img_data, confidence=40, overlap=30)
    return prediction.json()

用原来的测试代码重新发送请求，此时脚本会正确处理图片数据，调用Roboflow模型完成推理并返回结果。

内容的提问来源于stack exchange，提问作者Kédar