🦺 Factory PPE — v20260502

21 binary heads partial-label BCE MobileNetV3-L 訓練日期：2026-04-30 ｜ 來源：cvat2 project 5（66K imgs / 163K person crops）

📊 主要指標（Test split）

📈 訓練曲線

🎯 Per-attr 指標（Test）

📦 資料集

cvat2 project 5（factory_ppe_v20260426_pool）：31 tasks, 66,519 imgs。每張 person bbox 切 crop（15% padding）→ 163,242 person crops。

🚀 模型用法

import torch, torch.nn as nn
import timm
from PIL import Image
import torchvision.transforms as T

ckpt = torch.load("factory_ppe_v20260502_best.pt", weights_only=False)
class GenericClassifier(nn.Module):
    def __init__(self, backbone, n_attr, feat_dim):
        super().__init__()
        self.backbone = timm.create_model(backbone, pretrained=False, num_classes=0, global_pool="avg")
        self.dropout = nn.Dropout(0.3); self.cls = nn.Linear(feat_dim, n_attr)
    def forward(self, x): return self.cls(self.dropout(self.backbone(x)))

model = GenericClassifier(ckpt["backbone_name"], len(ckpt["attrs"]), ckpt["feat_dim"]).cuda().eval()
model.load_state_dict(ckpt["model_state"])
attrs = ckpt["attrs"]; thr = ckpt["thresholds"]

tf = T.Compose([T.Resize((384,192)), T.ToTensor(),
                T.Normalize([0.485,0.456,0.406], [0.229,0.224,0.225])])

def predict(person_crop_pil):
    x = tf(person_crop_pil).unsqueeze(0).cuda()
    with torch.no_grad():
        probs = torch.sigmoid(model(x))[0].cpu().numpy()
    return {a: (float(probs[i]), float(probs[i]) >= thr[a]) for i, a in enumerate(attrs)}

Live RTSP / Cascade 整合

已部署 http://192.168.53.21:7860/。 做為 PPE21 與 forklift_ppe（cascade pipeline）兩個 handler 的分類器。

📋 Hyperparameters

📥 模型下載

R2: https://pub-478929a98a5c440cb22c2241c0bde314.r2.dev/factory_ppe_v20260502/best.pt

Generated 2026-04-30 ｜ 回到目錄