"""
YOLO26/27 座舱边缘部署性能对比

目标芯片：
- Qualcomm QCS8255 (Hexagon NPU, 26 TOPS)
- Intel Core Ultra (OpenVINO)
- NVIDIA Jetson Orin Nano (20 TOPS)
- D-Robotics RDK (边缘SoC)

关键指标：
- 推理延迟 < 33ms (30fps)
- 模型大小 < 10MB (量化后)
- 功耗 < 2W
"""

import numpy as np
from dataclasses import dataclass
from typing import List

@dataclass
class EdgeBenchmark:
    """边缘部署基准"""
    model_name: str
    params_m: float       # 参数量(M)
    mAP_coco: float       # COCO mAP
    flops_g: float        # GFLOPs
    
    # Qualcomm QCS8255 (Hexagon NPU 26 TOPS)
    qcs8255_latency_ms: float
    qcs8255_fps: float
    qcs8255_quantized_mb: float  # INT8量化后大小
    qcs8255_power_w: float
    
    # NVIDIA Jetson Orin Nano (20 TOPS)
    jetson_latency_ms: float
    jetson_fps: float
    
    # Intel Core Ultra (OpenVINO)
    intel_latency_ms: float
    intel_fps: float

# YOLO26/27各尺寸基准（基于Ultralytics官方数据+估算）
benchmarks = [
    EdgeBenchmark(
        "YOLO26n", 2.6, 39.5, 8.1,
        qcs8255_latency_ms=8, qcs8255_fps=125, 
        qcs8255_quantized_mb=3.2, qcs8255_power_w=0.8,
        jetson_latency_ms=6, jetson_fps=167,
        intel_latency_ms=12, intel_fps=83
    ),
    EdgeBenchmark(
        "YOLO26s", 9.4, 47.2, 21.5,
        qcs8255_latency_ms=18, qcs8255_fps=56,
        qcs8255_quantized_mb=11.2, qcs8255_power_w=1.5,
        jetson_latency_ms=14, jetson_fps=71,
        intel_latency_ms=22, intel_fps=45
    ),
    EdgeBenchmark(
        "YOLO27n", 3.1, 42.0, 9.5,  # 改进版
        qcs8255_latency_ms=10, qcs8255_fps=100,
        qcs8255_quantized_mb=3.8, qcs8255_power_w=0.9,
        jetson_latency_ms=7, jetson_fps=143,
        intel_latency_ms=14, intel_fps=71
    ),
    EdgeBenchmark(
        "YOLO27s", 10.2, 50.5, 24.0,
        qcs8255_latency_ms=22, qcs8255_fps=45,
        qcs8255_quantized_mb=12.5, qcs8255_power_w=1.8,
        jetson_latency_ms=16, jetson_fps=62,
        intel_latency_ms=25, intel_fps=40
    ),
]

# 打印对比表
print("=" * 100)
print("YOLO26/27 座舱边缘部署基准对比")
print("=" * 100)
print(f"{'模型':<10} {'参数M':<8} {'mAP':<7} {'QCS8255延迟':<14} {'QCS8255 FPS':<14} "
      f"{'量化MB':<10} {'功耗W':<8} {'Jetson FPS':<12} {'Intel FPS':<10}")
print("-" * 100)
for b in benchmarks:
    print(f"{b.model_name:<10} {b.params_m:<8.1f} {b.mAP_coco:<7.1f} "
          f"{b.qcs8255_latency_ms:<14.0f} {b.qcs8255_fps:<14.0f} "
          f"{b.qcs8255_quantized_mb:<10.1f} {b.qcs8255_power_w:<8.1f} "
          f"{b.jetson_fps:<12.0f} {b.intel_fps:<10.0f}")

"""
YOLOE-27 座舱开放词汇检测示例

场景：检测驾驶员手持手机
传统方法：收集手机使用数据集 → 训练YOLO → 部署
YOLOE方法：直接用文本提示"person holding phone" → 推理

优势：无需标注训练数据，快速部署新检测需求
"""

# YOLOE-27伪代码（实际需YOLOE-27发布后替换）
def cabin_open_vocab_detection():
    """
    YOLOE-27 座舱开放词汇检测示例
    
    支持的提示模式：
    1. 文本提示: "person holding phone"
    2. 视觉提示: 提供一张手机照片
    3. 无提示: 自动检测显著物体
    """
    # 初始化
    model = YOLOE27n(
        task='detect',
        open_vocab=True,
        quantize=True,  # INT8量化
        target='qcs8255'
    )
    
    # 文本提示检测
    prompts = [
        "person holding phone",
        "person not wearing seatbelt", 
        "child in car seat",
        "person sleeping",
        "feet on dashboard",
    ]
    
    # 单次推理检测多个类别
    results = model.predict(
        image=camera_frame,
        prompts=prompts,
        confidence=0.5
    )
    
    return results