在Google Cloud Function中部署训练好的YoloV7模型

一、前期准备 #

• 精简YOLOv7的推理代码，移除训练相关冗余模块，只保留检测核心逻辑。
• 将你的自定义.pt模型上传至Google Cloud Storage（GCS）指定路径（比如gs://your-bucket/models/yolov7-custom.pt），Cloud Function本地存储空间有限，需从GCS加载模型。
• 确认Cloud Function使用Python 3.8及以上版本，尽量与训练模型时的Python版本一致，减少依赖冲突。

二、构建Cloud Function代码结构 #

1. 依赖配置（requirements.txt） #

torch==1.12.1+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
torchvision==0.13.1+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
opencv-python>=4.5.5
pillow>=9.0.0
google-cloud-storage>=2.0.0

注：指定CPU版本的预编译Torch包，避免Cloud Function环境下的编译错误。

2. 主函数代码（main.py） #

import os
import torch
import numpy as np
import cv2
from PIL import Image
from google.cloud import storage

# 初始化GCS客户端
storage_client = storage.Client()

# 全局加载YOLOv7模型（仅函数初始化时加载一次，降低冷启动耗时）
def load_yolov7_model():
    bucket_name = "your-bucket-name"
    model_blob = storage_client.bucket(bucket_name).blob("models/yolov7-custom.pt")
    temp_model_path = "/tmp/yolov7-custom.pt"
    model_blob.download_to_filename(temp_model_path)
    
    # 加载模型并设置为推理模式
    model = torch.load(temp_model_path, map_location=torch.device('cpu'))['model'].float().fuse().eval()
    return model

model = load_yolov7_model()

def detect_objects(request):
    # 解析请求参数
    request_json = request.get_json()
    if not request_json or 'image_gcs_path' not in request_json:
        return {"error": "缺少必要参数：image_gcs_path"}
    
    image_gcs_path = request_json['image_gcs_path']
    image_name = request_json.get('image_name', 'unknown.jpg')
    
    # 从GCS下载图片到临时目录
    bucket_name, blob_path = image_gcs_path.replace("gs://", "").split("/", 1)
    bucket = storage_client.bucket(bucket_name)
    blob = bucket.blob(blob_path)
    temp_image_path = f"/tmp/{image_name}"
    blob.download_to_filename(temp_image_path)
    
    # 图片预处理（适配YOLOv7输入要求）
    img = cv2.imread(temp_image_path)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = Image.fromarray(img).resize((640, 640))
    img = np.array(img) / 255.0
    img = np.transpose(img, (2, 0, 1))
    img = np.expand_dims(img, 0)
    img = torch.tensor(img).float()
    
    # 执行目标检测
    with torch.no_grad():
        outputs = model(img)
    
    # 解析检测结果
    detections = []
    for output in outputs:
        boxes = output[:, :4].cpu().numpy()  # [x1, y1, x2, y2]
        confidences = output[:, 4].cpu().numpy()
        class_ids = output[:, 5].cpu().numpy()
        
        for box, conf, cls_id in zip(boxes, confidences, class_ids):
            if conf > 0.5:  # 过滤低置信度结果
                detections.append({
                    "box": [float(coord) for coord in box],
                    "confidence": float(conf),
                    "class_id": int(cls_id)
                })
    
    return {"image_name": image_name, "detections": detections}

三、部署与配置 #

• 内存分配：建议给Cloud Function分配至少2GB内存，避免加载模型或推理时出现内存不足错误。
• 权限设置：为Cloud Function的服务账号添加GCS对象读取权限，确保能访问模型文件和待检测图片。
• 部署命令：使用gcloud工具部署（替换占位符为你的实际信息）：

gcloud functions deploy detect_objects \
--runtime python39 \
--trigger-http \
--memory 2048MB \
--service-account your-service-account@your-project.iam.gserviceaccount.com

四、测试函数 #

发送POST请求验证功能：

curl -X POST https://[REGION]-[PROJECT_ID].cloudfunctions.net/detect_objects \
-H "Content-Type: application/json" \
-d '{"image_gcs_path": "gs://your-bucket/images/test.jpg", "image_name": "test.jpg"}'

内容的提问来源于stack exchange，提问作者Natty Minds