YOLO+OpenClaw+SAM微调实战：工业缺陷自动标注的低代码落地

from peft import LoraConfig, get_peft_model
from segment_anything import sam_model_registry

# 加载通用SAM
sam = sam_model_registry["vit_h"](checkpoint="sam_vit_h.pth")

# 配置LoRA（只调整注意力层）
lora_config = LoraConfig(
    r=16,                    # 秩维度
    lora_alpha=32,
    target_modules=["qkv"],  # 只改查询/键/值层
    lora_dropout=0.1
)

# 应用LoRA
model = get_peft_model(sam, lora_config)
# 可训练参数占比仅2.5%，显存需求从24GB降到8GB

def industrial_loss(pred, target):
    """
    Dice + Focal组合
    - Dice：让分割边界更贴合
    - Focal：让模型更关注难分的缺陷区域
    """
    pred_sigmoid = torch.sigmoid(pred)
    
    # Dice Loss（边界贴合）
    intersection = (pred_sigmoid * target).sum()
    dice = 1 - (2 * intersection) / (pred_sigmoid.sum() + target.sum() + 1e-6)
    
    # Focal Loss（关注难例）
    ce = torch.nn.functional.binary_cross_entropy_with_logits(pred, target, reduction='none')
    pt = torch.exp(-ce)
    focal = 0.25 * (1-pt)**2 * ce
    
    return 0.6 * dice + 0.4 * focal.mean()

import cv2, json, numpy as np
from ultralytics import YOLO
from segment_anything import SamPredictor
from peft import PeftModel

class IndustrialAnnotator:
    def __init__(self, yolo_path, sam_path, lora_path):
        self.yolo = YOLO(yolo_path)                      # YOLO检测
        sam = sam_model_registry["vit_h"](checkpoint=sam_path)
        peft_model = PeftModel.from_pretrained(sam, lora_path)  # 加载微调权重
        self.predictor = SamPredictor(peft_model)
    
    def annotate(self, img_path, out_dir):
        img = cv2.imread(img_path)
        self.predictor.set_image(img)
        
        # 1. YOLO粗定位
        results = self.yolo(img)[0]
        
        annotations = []
        for box in results.boxes:
            x1,y1,x2,y2 = map(int, box.xyxy[0])
            
            # 2. SAM微调精分割
            masks,_,_ = self.predictor.predict(box=np.array([x1,y1,x2,y2]))
            mask = masks[0]
            
            annotations.append({
                "type": self.yolo.names[int(box.cls[0])],
                "bbox": [x1,y1,x2,y2],
                "mask": mask.tolist()
            })
        
        # 3. 输出JSON标注文件
        with open(f"{out_dir}/{Path(img_path).stem}.json", 'w') as f:
            json.dump(annotations, f)

用户：每天凌晨2点，用 industrial-annotator 技能处理 /data/defect_raw/ 目录下的图片，
      先用YOLO检测缺陷，再用微调后的SAM生成精细掩码，输出标注文件到 /data/defect_annotated/
OpenClaw：已创建定时任务，每天2:00执行批量标注