首先按照官方教程安装CVAT：https://docs.cvat.ai/docs/administration/basics/installation/，建议使用Ubuntu22.04版本，较为稳定。
CVAT使用无服务器函数进行自动标注模型的部署。此外，当需要使用自动标注模型的时候，不应直接使用docker compose up -d直接启动服务，而应该执行如下命令：
docker compose -f docker-compose.yml -f components/serverless/docker-compose.serverless.yml up -d如果之前使用这个命令运行CVAT，需要进行停止：
docker compose -f docker-compose.yml -f components/serverless/docker-compose.serverless.yml down安装nuctl（使用版本1.13.0）：
wget https://github.com/nuclio/nuclio/releases/download/1.13.0/nuctl-1.13.0-linux-amd64给予权限和软连接：
sudo chmod +x nuctl-1.13.0-linux-amd64
sudo ln -sf $(pwd)/nuctl-1.13.0-linux-amd64 /usr/local/bin/nuctl部署无服务器函数（当且仅当CVAT启动完毕之后，再执行如下命令）：
./serverless/deploy_cpu.sh serverless/openvino/dextr
./serverless/deploy_cpu.sh serverless/openvino/omz/public/yolo-v3-tf使用GPU，需要对函数进行修改，并使用./serverless/depoly_gpu.sh。
示例：部署一个yolov8训练的权重的配置文件：
在serverless目录下创建自定义函数目录
mkdir -p serverless/pytorch/ultralytics/yolov8/nuclio创建镜像构建文件function-gpu.yaml
metadata:
name: emdetector
namespace: cvat
annotations:
name: emdetect
type: detector
framework: pytorch
spec: |
[
{ "id": 0, "name": "EM14", "type": "rectangle" },
{ "id": 1, "name": "EM18", "type": "rectangle" },
{ "id": 2, "name": "EM17", "type": "rectangle" },
]
spec:
description: 工位检测
runtime: "python:3.11"
handler: main:handler
eventTimeout: 30s
build:
image: cvat.pth.yolo11.emdetector:latest-gpu
baseImage: python:3.9
directives:
preCopy:
- kind: ENV
value: DEBIAN_FRONTEND=noninteractive
- kind: RUN
value: apt-get update && apt-get install -y libgl1 libglib2.0-0 && apt-get clean
- kind: RUN
value: pip install ultralytics torch torchvision opencv-python-headless && pip cache purge
triggers:
myHttpTrigger:
numWorkers: 1
kind: 'http'
workerAvailabilityTimeoutMilliseconds: 10000
attributes:
# Set value from the calculation of tracking of 100 objects at the same time on a 4k image
maxRequestBodySize: 268435456 # 256MB
volumes:
- volume:
name: model-volume
hostPath:
path: /path/to/models/emdetection/best.pt # 可选：如果使用HostPath挂载模型
volumeMount:
name: model-volume
mountPath: /opt/nuclio/best.pt
resources:
limits:
nvidia.com/gpu: 1
platform:
attributes:
restartPolicy:
name: always
maximumRetryCount: 3
mountMode: volume
创建函数入口文件main.py
import json
import base64
# from PIL import Image
import io
from model_handler import ModelHandler
def init_context(context):
context.logger.info("Init context... 0%")
context.logger.info("Initializing EMDetection model...")
context.user_data.model_handler = ModelHandler()
context.logger.info("Init context...100%")
def handler(context, event):
context.logger.info("Run EMDetection model")
data = event.body
image_data = base64.b64decode(data["image"])
threshold = float(data.get("threshold", 0.5))
results = context.user_data.model_handler.infer(image_data, threshold)
return context.Response(body=json.dumps(results), headers={},
content_type='application/json', status_code=200)
创建推理函数文件model_handler.py
import io
from PIL import Image
from ultralytics import YOLO
class ModelHandler:
def __init__(self):
"""加载 YOLOv11 模型"""
self.model = YOLO("/opt/nuclio/best.pt") # 确保路径正确
def infer(self, image_data, threshold=0.3):
"""
执行推理
:param image_data: 图片的二进制数据
:param threshold: 置信度阈值（默认0.3）
:return: 符合阈值的检测结果
"""
image = Image.open(io.BytesIO(image_data))
results = self.model(image)
detections = []
for result in results:
for box in result.boxes.data.tolist():
x1, y1, x2, y2, score, class_id = box
if score >= threshold: # 过滤低置信度目标
detections.append({
"confidence": score,
"label": self.model.names[int(class_id)],
"points": [x1, y1, x2, y2],
"type": "rectangle",
})
return detections
将这些文件上传至服务器指定目录，并运行GPU部署命令。即可完成部署。