在Flask服务器启动时运行mask-rcnn模型

Question

这是我当前运行良好的Flask代码，它从客户端接收一个带有图像的POST请求，通过模型运行它(基于这个GH：https://github.com/matterport/Mask_RCNN)，然后将一个被屏蔽的图像发送回客户端。

但是，它从Configuration文件加载模型并为每个请求加载权重，这需要很长时间。我想在服务器启动时加载模型和权重，并将其传递给index函数。我试过其他问题的答案，但没有成功。我想知道这是不是因为我加载了一个模型，然后加权，而不是仅仅加载一个h5模型文件？

How do I load a file on initialization in a flask application Run code after flask application has started

Flask应用：

from flask import Flask, jsonify, request
import base64
import cv2
import numpy as np
from Configuration import create_model


app = Flask(__name__)

@app.route('/', methods=['GET', 'POST'])
def index():
    if request.method == "POST":
        # Load the image sent from the client
        imagefile = request.files['image'].read()  # Type: bytes
        jpg_as_np = np.frombuffer(imagefile, dtype=np.uint8) # Convert to numpy array
        img = cv2.imdecode(jpg_as_np, flags=1) # Decode from numpy array to opencv object - This is an array

        ### Enter OpenCV/Tensorflow below ###

        model = create_model()

        image = img[..., ::-1]

        # Detect objects
        r = model.detect([image], verbose=0)[0]

        REDACTED VISUALIATION CODE

        ### ###

        string = base64.b64encode(cv2.imencode('.jpg', masked_image)[1]).decode() # Convert back to b64 string ready for json.

        return jsonify({"count": str(r["masks"].shape[2]), 'image': string})

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

配置：

def create_model():
    device = "/cpu:0"
    weights_path = "weights.h5"
    with tf.device(device):
        model = modellib.MaskRCNN(mode="inference", model_dir=weights_path, config=InferenceConfig())
    model.load_weights(weights_path, by_name=True)
    print("Weights Loaded")
    return model

Answer 1

我使用before_first_request装饰器解决了这个问题。下面是总体结构：

app = Flask(__name__)

@app.before_first_request
def before_first_request_func():
    MOODEL WEIGHT LOADING CODE
    return model

@app.route('/', methods=['POST'])
def index():
    if request.method == "POST":
        REDACTED LOADING CODE
        # Detect objects
        r = model.detect([image], verbose=0)[0]

        REDACTED VISUALISATION CODE

        string = base64.b64encode(cv2.imencode('.jpg', masked_image)[1]).decode() # Convert back to b64 string ready for json.

        return jsonify({"count": str(r["masks"].shape[2]), 'image': string})

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

model存储在内存中，稍后可以在检测函数中引用。它可用于每个POST请求，不需要重新加载。