Computer Vision YOLO Scaling Strategy — วิธี

YOLO Object Detection

YOLO You Only Look Once Real-time Object Detection ตรวจจับวัตถุครั้งเดียว YOLOv8 Ultralytics Detection Segmentation Pose Estimation CCTV Autonomous Driving

Scale ได้หลายระดับ Model Nano Small Medium Large GPU TensorRT Kubernetes Batch Processing Edge Device Jetson Raspberry Pi

Model	Parameters	mAP	Speed (ms)	เหมาะกับ
YOLOv8n	3.2M	37.3	1.2	Edge / Mobile
YOLOv8s	11.2M	44.9	2.1	Edge / Light Server
YOLOv8m	25.9M	50.2	4.7	Server
YOLOv8l	43.7M	52.9	7.1	Server GPU
YOLOv8x	68.2M	53.9	10.8	High Accuracy

Performance Optimization

# optimization.py — YOLO Performance Optimization

optimizations = {

    "TensorRT FP16": {

        "speedup": "2-3x faster",

        "accuracy_loss": "< 0.5% mAP",

        "command": "yolo export model=yolov8m.pt format=engine half=True",

        "requirement": "NVIDIA GPU (Compute Capability >= 7.0)",

    },

    "TensorRT INT8": {

        "speedup": "3-5x faster",

        "accuracy_loss": "1-2% mAP",

        "command": "yolo export model=yolov8m.pt format=engine int8=True data=coco.yaml",

        "requirement": "Calibration Dataset needed",

    },

    "ONNX Runtime": {

        "speedup": "1.5-2x faster (CPU)",

        "accuracy_loss": "0%",

        "command": "yolo export model=yolov8m.pt format=onnx simplify=True",

        "requirement": "onnxruntime / onnxruntime-gpu",

    },

    "OpenVINO": {

        "speedup": "2-3x faster (Intel CPU)",

        "accuracy_loss": "< 0.5%",

        "command": "yolo export model=yolov8m.pt format=openvino half=True",

        "requirement": "Intel CPU/iGPU",

    },

    "Image Resize": {

        "speedup": "640->320 = 2-4x faster",

        "accuracy_loss": "5-10% mAP",

        "command": "model.predict(source, imgsz=320)",

        "requirement": "ลด Input Resolution",

    },

    "Batch Inference": {

        "speedup": "2-4x throughput",

        "accuracy_loss": "0%",

        "command": "model.predict(sources, batch=16)",

        "requirement": "GPU Memory เพียงพอ",

    },

}



print("YOLO Performance Optimizations:")

for name, info in optimizations.items():

    print(f"\n  [{name}]")

    print(f"    Speedup: {info['speedup']}")

    print(f"    Accuracy Loss: {info['accuracy_loss']}")

    print(f"    Command: {info['command']}")

    print(f"    Requirement: {info['requirement']}")



# GPU Comparison

gpus = {

    "Jetson Nano": {"VRAM": "4GB", "YOLOv8n": "15 FPS", "cost": "$200"},

    "Jetson Orin": {"VRAM": "8-32GB", "YOLOv8n": "60 FPS", "cost": "$500-2000"},

    "T4": {"VRAM": "16GB", "YOLOv8m": "80 FPS", "cost": "$0.5/hr (Cloud)"},

    "A10G": {"VRAM": "24GB", "YOLOv8m": "150 FPS", "cost": "$1/hr (Cloud)"},

    "RTX 4090": {"VRAM": "24GB", "YOLOv8m": "200 FPS", "cost": "$1,599"},

    "A100": {"VRAM": "80GB", "YOLOv8l": "250 FPS", "cost": "$3/hr (Cloud)"},

}



print(f"\n\nGPU Comparison for YOLO:")

for gpu, info in gpus.items():

    print(f"  {gpu:<14} VRAM: {info['VRAM']:<8} Speed: {info.get('YOLOv8m', info.get('YOLOv8n', info.get('YOLOv8l', 'N/A'))):<10} Cost: {info['cost']}")

เคล็ดลับ

TensorRT: Export YOLO เป็น TensorRT FP16 เร็วขึ้น 2-3 เท่า
Model Size: เลือกขนาดตามงาน Nano สำหรับ Edge Large สำหรับ Server
Batch: ใช้ Batch Inference เพิ่ม Throughput 2-4 เท่า
Queue: ใช้ Redis/Kafka Buffer Requests ป้องกัน Overload
HPA: ใช้ Kubernetes HPA Scale Pods ตาม GPU Utilization
Monitor: ติดตาม FPS Latency GPU Memory ด้วย Prometheus

YOLO คืออะไร

You Only Look Once Real-time Object Detection ตรวจจับวัตถุครั้งเดียว YOLOv8 Ultralytics Detection Segmentation Pose CCTV Autonomous Driving Quality Inspection