Week 1｜YOLOv8n × BDD100K（5k 子集）复现与解读
原创2025/8/24大约 10 分钟
一周复现系列（W1）。本篇记录我从数据下载到训练评测的完整过程，并提供一键复现命令清单与结果解读（mAP 的含义与读法）。
TL;DR（结果速览）
- 训练集：BDD100K 5k 子集（按 8 类筛选）；验证集：官方 val 全量 10k
- 模型与设置：
yolov8n.pt，imgsz=960，batch=16，GPU: Tesla T4 16GB - 训练：50 epochs，耗时 ≈2.37 h
- 指标（val, 10k 图）：mAP50=0.489，mAP50-95=0.281；P=0.581，R=0.459
- 结论：总体可作为入门基线；**小目标类（person/bicycle）**召回偏低，提升分辨率与增广可改善。
一、环境与目录
mamba create -n adv_w1 python=3.10 -y
mamba activate adv_w1
pip install ultralytics opencv-python-headless tqdm目录（ETHZ 镜像解压后的常见结构）：
data/bdd100k/
images/100k/{train,val,test}/*.jpg
labels/
det_20/{det_train.json, det_val.json} # Detection 2020（Scalabel/COCO混合来源）
100k/{train,val} # YOLO 文本标注（生成后/软链）软链对齐（让
images/100k/...能映射到labels/100k/...）：
mkdir -p data/bdd100k/labels/100k
ln -sfn ../train data/bdd100k/labels/100k/train
ln -sfn ../val data/bdd100k/labels/100k/val二、数据准备与标注转换
1) 下载（ETHZ 镜像）
# 建议逐个下载，避免 416
aria2c -c -x8 -s8 -k1M https://dl.cv.ethz.ch/bdd100k/data/100k_images_train.zip
aria2c -c -x8 -s8 -k1M https://dl.cv.ethz.ch/bdd100k/data/100k_images_val.zip
aria2c -c -x8 -s8 -k1M https://dl.cv.ethz.ch/bdd100k/data/bdd100k_det_20_labels_trainval.zip
unzip -q 100k_images_train.zip
unzip -q 100k_images_val.zip
unzip -q bdd100k_det_20_labels_trainval.zip也可使用自定义脚本：
bash bdd100k_downloader.sh --size 100k --dir ./data
2) 生成 5k 训练子集（兼容 Scalabel/COCO）
python w1/scripts/make_subset_any.py \
--labels-json data/bdd100k/labels/det_20/det_train.json \
--images-root data/bdd100k/images/100k/train \
--save-json data/bdd100k/labels/det_20/det_train_5k.json \
--max-images 5000 \
--classes car bus truck person bicycle motorcycle "traffic light" "traffic sign"3) 转 YOLO 文本标注
# train（5k 子集）
python w1/scripts/coco_to_yolo.py \
--coco-json data/bdd100k/labels/det_20/det_train_5k.json \
--images-root data/bdd100k/images/100k/train \
--labels-root data/bdd100k/labels/train \
--classes car bus truck person bicycle motorcycle "traffic light" "traffic sign"
# val（先转 COCO 再转 YOLO；避免 Scalabel 直接读失败）
python w1/scripts/make_subset_any.py \
--labels-json data/bdd100k/labels/det_20/det_val.json \
--images-root data/bdd100k/images/100k/val \
--save-json data/bdd100k/labels/det_20/det_val_coco.json \
--max-images 999999 \
--classes car bus truck person bicycle motorcycle "traffic light" "traffic sign"
python w1/scripts/coco_to_yolo.py \
--coco-json data/bdd100k/labels/det_20/det_val_coco.json \
--images-root data/bdd100k/images/100k/val \
--labels-root data/bdd100k/labels/val \
--classes car bus truck person bicycle motorcycle "traffic light" "traffic sign"4) 数据集配置（data/bdd100k/bdd5k.yaml）
path: data/bdd100k
train: train_5k.txt # 相对 path，真实文件为 data/bdd100k/train_5k.txt
val: images/100k/val
names:
0: car
1: bus
2: truck
3: person
4: bicycle
5: motorcycle
6: traffic light
7: traffic sign将 5k 训练图片清单写入
data/bdd100k/train_5k.txt（避免 7 万无标注图被当背景）：
python - <<'PY'
import json, os
subset='data/bdd100k/labels/det_20/det_train_5k.json'
img_root=os.path.abspath('data/bdd100k/images/100k/train')
d=json.load(open(subset,'r',encoding='utf-8'))
names={os.path.basename(x['file_name']) for x in d['images']}
with open('data/bdd100k/train_5k.txt','w') as f:
for n in sorted(names): f.write(os.path.join(img_root,n)+'\n')
print('wrote', len(names), 'lines')
PY三、训练
yolo detect train data=data/bdd100k/bdd5k.yaml \
model=yolov8n.pt imgsz=960 epochs=50 batch=16 workers=8 device=0 \
name=bdd5k_yolov8n_960_subset5k2- Ultralytics: 8.3.179；PyTorch 2.8.0+cu128；GPU: Tesla T4 16GB
- 优化器自动选择：AdamW(lr≈0.000833)
- 训练耗时：≈2.371 h / 50 epochs
四、评测结果（mAP 详解）
- 整体：
P=0.581,R=0.459,mAP50=0.489,mAP50-95=0.281 - 按类（mAP50）：car 0.767，traffic light 0.587，motorcycle 0.544，truck 0.499，bus 0.485，person 0.312，bicycle 0.226
- 速度：pre 0.2ms / infer 2.4ms / post 1.2ms per image
mAP 解释
- AP：在某个 IoU 阈值下，PR 曲线下面积；mAP：对所有类别的 AP 取平均。
- mAP50：IoU=0.50（宽松）；mAP50-95：IoU 0.50→0.95 每 0.05 一档的平均（严格）。
- 你的结果显示召回偏低（R≈0.46），小目标类更难 → 可通过更高分辨率与增广改进。
生成曲线与混淆矩阵：
yolo detect val data=data/bdd100k/bdd5k.yaml \
model=runs/detect/bdd5k_yolov8n_960_subset5k2/weights/best.pt \
plots=True save_json=True device=0五、误差分析与可视化
- 远距小目标（person/bicycle）：细节不足导致漏检、框偏小
- 背光/夜间：低照/高对比 → 误检
- 遮挡/密集：NMS 后漏检
导出预测样例（做“误检/漏检案例墙”）：
yolo detect predict source=data/bdd100k/images/100k/val \
model=runs/detect/bdd5k_yolov8n_960_subset5k2/weights/best.pt \
conf=0.25 save=True max_det=300 device=0六、立刻可做的改进实验
每条独立尝试，更换
name=，避免覆盖。
# 1) 更高分辨率（冲小目标）
yolo detect train data=data/bdd100k/bdd5k.yaml model=yolov8n.pt imgsz=1280 batch=8 epochs=50 device=0 name=bdd5k_yolov8n_1280
# 2) 轻度增广（mosaic/mixup）
yolo detect train data=data/bdd100k/bdd5k.yaml model=yolov8n.pt imgsz=960 epochs=50 mosaic=0.8 mixup=0.1 hsv_h=0.015 hsv_s=0.7 hsv_v=0.4 device=0 name=bdd5k_yolov8n_aug
# 3) 更长训练 + 余弦退火
yolo detect train data=data/bdd100k/bdd5k.yaml model=yolov8n.pt imgsz=960 epochs=100 cos_lr=True device=0 name=bdd5k_yolov8n_100ep
# 4) 更大模型（提升上限）
yolo detect train data=data/bdd100k/bdd5k.yaml model=yolov8s.pt imgsz=960 epochs=50 batch=16 device=0 name=bdd5k_yolov8s_960七、导出与部署（可选）
# 导出 ONNX
yolo export model=runs/detect/bdd5k_yolov8n_960_subset5k2/weights/best.pt format=onnx opset=12
# 用 ONNX 推理一次校验
yolo detect predict source=data/bdd100k/images/100k/val \
model=runs/detect/bdd5k_yolov8n_960_subset5k2/weights/best.onnx \
conf=0.25 save=True device=0八、踩坑与排错清单
no labels found：对齐images/100k/...↔labels/100k/...；- Scalabel 读取报错：先转 COCO 再转 YOLO；
- OpenCV 导入失败：用
opencv-python-headless或安装libgl1 libglib2.0-0； - 下载 416：分开下载、降低并发或用
wget -c； path与train重复前缀：train应相对path，或将path设为.并用绝对路径。
九、结语
本篇完成了 BDD100K 上 YOLOv8n 的从零到一复现与指标解读。作为入门基线，它已经可用于下游（如 MOT 的 ByteTrack 输入）。下一篇将在车道线检测或单目深度之间选择作为 W2 主题。
附录
训练脚本
#!/usr/bin/env bash
# Week1 end-to-end pipeline: YOLOv8n × BDD100K (5k subset)
# 作用：环境准备 -> 下载/解压 -> 标注转换 -> 训练/验证/预测
# 需求：Ubuntu + Mambaforge/Conda（建议）+ NVIDIA 驱动（GPU 可选）
# 用法：bash week1_all.sh
set -euo pipefail
### ----------------------- 基础变量 -----------------------
PROJ_ROOT="${PROJ_ROOT:-$PWD}"
DATA_DIR="$PROJ_ROOT/data"
BDD_DIR="$DATA_DIR/bdd100k"
ENV_NAME="${ENV_NAME:-adv_w1}"
RUN="mamba run -n $ENV_NAME" # 不改变父 shell 的情况下使用环境
RUN_NAME="${RUN_NAME:-bdd5k_yolov8n_960_subset5k}"
TRAIN_IMGSZ="${TRAIN_IMGSZ:-960}"
EPOCHS="${EPOCHS:-50}"
BATCH="${BATCH:-16}"
DEVICE="${DEVICE:-0}" # 指定 GPU；CPU 调试可设为 cpu
echo "==> PROJ_ROOT=$PROJ_ROOT"
echo "==> DATA_DIR=$DATA_DIR"
echo "==> BDD_DIR=$BDD_DIR"
mkdir -p "$BDD_DIR" "$PROJ_ROOT/w1/scripts"
### ----------------------- 系统依赖 -----------------------
echo "==> 安装系统依赖（aria2/unzip；GPU 不需要额外包）"
if ! command -v aria2c >/dev/null 2>&1; then
(apt-get update -y || true) && apt-get install -y aria2 || true
fi
apt-get install -y unzip || true
### ----------------------- Python 环境 -----------------------
if ! command -v mamba >/dev/null 2>&1 && command -v conda >/dev/null 2>&1; then
# 没有 mamba 就用 conda
alias mamba=conda
fi
if ! mamba env list | grep -q "^$ENV_NAME"; then
echo "==> 创建环境 $ENV_NAME"
mamba create -y -n "$ENV_NAME" python=3.10
fi
echo "==> 安装 Python 依赖（ultralytics / pillow / opencv-python-headless / tqdm）"
$RUN python - <<'PY'
import sys, subprocess
pip = [sys.executable, "-m", "pip"]
subprocess.check_call(pip + ["install", "-U", "pip"])
subprocess.check_call(pip + ["install", "ultralytics", "pillow", "opencv-python-headless", "tqdm"])
PY
### ----------------------- 下载 BDD100K -----------------------
cd "$DATA_DIR"
echo "==> 下载 BDD100K（ETHZ 镜像；images 100k + det_20 labels）"
download() {
local url="$1"
if command -v aria2c >/dev/null 2>&1; then
aria2c -c -x8 -s8 -k1M "$url"
else
wget -c "$url"
fi
}
download https://dl.cv.ethz.ch/bdd100k/data/100k_images_train.zip
download https://dl.cv.ethz.ch/bdd100k/data/100k_images_val.zip
download https://dl.cv.ethz.ch/bdd100k/data/bdd100k_det_20_labels_trainval.zip
# 只有 images 有 md5；labels 没有 md5 是正常的
download https://dl.cv.ethz.ch/bdd100k/data/100k_images_train.zip.md5 || true
download https://dl.cv.ethz.ch/bdd100k/data/100k_images_val.zip.md5 || true
if command -v md5sum >/dev/null 2>&1; then
[ -f 100k_images_train.zip.md5 ] && md5sum -c 100k_images_train.zip.md5
[ -f 100k_images_val.zip.md5 ] && md5sum -c 100k_images_val.zip.md5
fi
echo "==> 解压..."
unzip -q -n 100k_images_train.zip
unzip -q -n 100k_images_val.zip
unzip -q -n bdd100k_det_20_labels_trainval.zip
# 目录对齐与软链
cd "$PROJ_ROOT"
echo "==> 创建对齐软链 images/train|val -> images/100k/*"
ln -sfn "$BDD_DIR/images/100k/train" "$BDD_DIR/images/train"
ln -sfn "$BDD_DIR/images/100k/val" "$BDD_DIR/images/val"
mkdir -p "$BDD_DIR/labels/100k" "$BDD_DIR/labels/train" "$BDD_DIR/labels/val"
ln -sfn ../train "$BDD_DIR/labels/100k/train"
ln -sfn ../val "$BDD_DIR/labels/100k/val"
### ----------------------- 准备脚本（兼容 Scalabel/COCO） -----------------------
echo "==> 写入 make_subset_any.py（兼容 Scalabel/COCO；输出 COCO 子集）"
cat > "$PROJ_ROOT/w1/scripts/make_subset_any.py" <<'PY'
#!/usr/bin/env python
# 兼容 BDD100K 的 COCO 与 Scalabel，抽样为 COCO 子集
import os, json, argparse, random
from pathlib import Path
try:
import cv2
except Exception:
cv2 = None
def parse_args():
ap = argparse.ArgumentParser()
ap.add_argument("--labels-json", required=True)
ap.add_argument("--images-root", required=True)
ap.add_argument("--save-json", required=True)
ap.add_argument("--max-images", type=int, default=5000)
ap.add_argument("--classes", nargs="+", required=True)
ap.add_argument("--seed", type=int, default=42)
return ap.parse_args()
def load_json(p): return json.load(open(p, "r", encoding="utf-8"))
def is_coco(x): return isinstance(x, dict) and all(k in x for k in ["images","annotations","categories"])
def to_coco_from_scalabel(frames, images_root, keep_names):
if cv2 is None:
raise SystemExit("需要 opencv-python-headless：pip install opencv-python-headless")
seen, cat2id = [], {}
for fr in frames:
for lb in fr.get("labels", []):
c = lb.get("category")
if c and c not in seen:
seen.append(c)
for i,n in enumerate([n for n in seen if n in keep_names], start=1):
cat2id[n] = i
images, annotations = [], []
iid, aid = 1, 1
for fr in frames:
name = fr.get("name");
if not name: continue
fn = Path(name).name
p = Path(images_root)/fn
if not p.exists(): p = Path(images_root)/name
if not p.exists(): continue
im = cv2.imread(str(p))
if im is None: continue
h,w = im.shape[:2]
labs=[]
for lb in fr.get("labels", []):
cat=lb.get("category"); box=lb.get("box2d")
if not box or cat not in cat2id: continue
x1,y1 = float(box.get("x1",0)), float(box.get("y1",0))
x2,y2 = float(box.get("x2",0)), float(box.get("y2",0))
bw,bh = max(0.0,x2-x1), max(0.0,y2-y1)
if bw<=1e-6 or bh<=1e-6: continue
labs.append((cat2id[cat], x1,y1,bw,bh))
if not labs: continue
images.append({"id": iid, "file_name": fn, "width": w, "height": h})
for (cid,x,y,bw,bh) in labs:
annotations.append({"id": aid,"image_id": iid,"category_id": cid,"bbox":[x,y,bw,bh],"area":bw*bh,"iscrowd":0})
aid+=1
iid+=1
categories=[{"id":v,"name":k} for k,v in cat2id.items()]
return {"images":images,"annotations":annotations,"categories":categories}
def filter_and_sample_coco(coco, keep_names, max_images, seed=42):
name2id={c["name"]:c["id"] for c in coco["categories"]}
keep_ids={name2id[n] for n in keep_names if n in name2id}
from collections import defaultdict
by_img=defaultdict(list)
for a in coco["annotations"]:
if a.get("category_id") in keep_ids: by_img[a["image_id"]].append(a)
valid=[img for img in coco["images"] if by_img.get(img["id"])]
random.seed(seed); random.shuffle(valid)
chosen={img["id"] for img in valid[:max_images]}
images=[i for i in coco["images"] if i["id"] in chosen]
anns=[a for a in coco["annotations"] if a["image_id"] in chosen and a["category_id"] in keep_ids]
cats=[c for c in coco["categories"] if c["id"] in keep_ids]
return {"images":images,"annotations":anns,"categories":cats}
def main():
args=parse_args(); labs=load_json(args.labels_json)
keep=set(args.classes)
if is_coco(labs):
sub=filter_and_sample_coco(labs, keep, args.max_images, args.seed)
else:
frames=labs["frames"] if isinstance(labs,dict) and "frames" in labs else labs
coco_all=to_coco_from_scalabel(frames, args.images_root, keep)
sub=filter_and_sample_coco(coco_all, keep, args.max_images, args.seed)
os.makedirs(Path(args.save_json).parent, exist_ok=True)
json.dump(sub, open(args.save_json,"w",encoding="utf-8"))
print(f"子集完成：images={len(sub['images'])}, anns={len(sub['annotations'])}, 类别={len(sub['categories'])}")
if __name__=="__main__": main()
PY
chmod +x "$PROJ_ROOT/w1/scripts/make_subset_any.py"
echo "==> 写入 coco_or_scalabel_to_yolo.py（直接转 YOLO，兼容两种标注）"
cat > "$PROJ_ROOT/w1/scripts/coco_or_scalabel_to_yolo.py" <<'PY'
#!/usr/bin/env python
# 将 COCO 或 Scalabel 直接转为 YOLO 文本标注
import os, json, argparse
from pathlib import Path
from collections import defaultdict
def parse_args():
ap=argparse.ArgumentParser()
ap.add_argument("--labels-json", required=True)
ap.add_argument("--images-root", required=True)
ap.add_argument("--labels-root", required=True)
ap.add_argument("--classes", nargs="+", required=True)
return ap.parse_args()
def load_json(p): return json.load(open(p, "r", encoding="utf-8"))
def is_coco(x): return isinstance(x, dict) and all(k in x for k in ["images","annotations","categories"])
def get_wh(p):
try:
from PIL import Image
with Image.open(p) as im: return im.size[0], im.size[1]
except Exception:
try:
import cv2
im=cv2.imread(str(p));
if im is None: raise RuntimeError
h,w=im.shape[:2]; return w,h
except Exception:
raise SystemExit("无法读取图像尺寸，请安装 pillow 或 opencv-python-headless")
def coco_to_yolo(coco, images_root, labels_root, keep):
os.makedirs(labels_root, exist_ok=True)
name2cid={c["name"]:c["id"] for c in coco["categories"]}
keep=[n for n in keep if n in name2cid]
name2yid={n:i for i,n in enumerate(keep)}
keep_cids={name2cid[n] for n in keep}
by_img=defaultdict(list)
for a in coco["annotations"]:
if a.get("category_id") in keep_cids and "bbox" in a:
by_img[a["image_id"]].append(a)
cid2name={v:k for k,v in name2cid.items()}
wh_cache={}
for img in coco["images"]:
fn=Path(img.get("file_name")).name
w,h=img.get("width"), img.get("height")
if not w or not h:
p=Path(images_root)/fn
if fn not in wh_cache: wh_cache[fn]=get_wh(p)
w,h=wh_cache[fn]
lines=[]
for a in by_img.get(img["id"], []):
name=cid2name.get(a["category_id"])
if name not in name2yid: continue
x,y,bw,bh=a["bbox"]
x_c=(x+bw/2)/w; y_c=(y+bh/2)/h; bw/=w; bh/=h
lines.append(f"{name2yid[name]} {x_c:.6f} {y_c:.6f} {bw:.6f} {bh:.6f}")
out=Path(labels_root)/(Path(fn).stem+".txt")
open(out,"w",encoding="utf-8").write("\n".join(lines))
def scalabel_to_yolo(frames, images_root, labels_root, keep):
os.makedirs(labels_root, exist_ok=True)
seen=[]
for fr in frames:
for lb in fr.get("labels", []):
c=lb.get("category")
if c and c not in seen: seen.append(c)
keep=[n for n in seen if n in keep]
name2yid={n:i for i,n in enumerate(keep)}
for fr in frames:
name=fr.get("name");
if not name: continue
fn=Path(name).name
p=Path(images_root)/fn
if not p.exists(): p=Path(images_root)/name
if not p.exists(): continue
w,h=get_wh(p)
lines=[]
for lb in fr.get("labels", []):
cat=lb.get("category"); b=lb.get("box2d")
if not b or cat not in name2yid: continue
x1,y1=float(b.get("x1",0)), float(b.get("y1",0))
x2,y2=float(b.get("x2",0)), float(b.get("y2",0))
bw,bh=max(0.0,x2-x1), max(0.0,y2-y1)
if bw<=1e-6 or bh<=1e-6: continue
x=(x1+bw/2)/w; y=(y1+bh/2)/h; bw/=w; bh/=h
lines.append(f"{name2yid[cat]} {x:.6f} {y:.6f} {bw:.6f} {bh:.6f}")
out=Path(labels_root)/(Path(fn).stem+".txt")
open(out,"w",encoding="utf-8").write("\n".join(lines))
def main():
a=parse_args(); d=load_json(a.labels_json); keep=a.classes
if is_coco(d): coco_to_yolo(d, a.images_root, a.labels_root, keep)
else:
frames=d["frames"] if isinstance(d,dict) and "frames" in d else d
scalabel_to_yolo(frames, a.images_root, a.labels_root, keep)
print(f"完成：YOLO 标签写入 {a.labels_root}")
if __name__=="__main__": main()
PY
chmod +x "$PROJ_ROOT/w1/scripts/coco_or_scalabel_to_yolo.py"
### ----------------------- 生成 5k 子集 & 转 YOLO -----------------------
echo "==> 生成 train 5k 子集（COCO）"
$RUN python "$PROJ_ROOT/w1/scripts/make_subset_any.py" \
--labels-json "$BDD_DIR/labels/det_20/det_train.json" \
--images-root "$BDD_DIR/images/100k/train" \
--save-json "$BDD_DIR/labels/det_20/det_train_5k.json" \
--max-images 5000 \
--classes car bus truck person bicycle motorcycle "traffic light" "traffic sign"
echo "==> 转 YOLO（train 5k）"
$RUN python "$PROJ_ROOT/w1/scripts/coco_or_scalabel_to_yolo.py" \
--labels-json "$BDD_DIR/labels/det_20/det_train_5k.json" \
--images-root "$BDD_DIR/images/100k/train" \
--labels-root "$BDD_DIR/labels/train" \
--classes car bus truck person bicycle motorcycle "traffic light" "traffic sign"
echo "==> 转 YOLO（val 全量，直接从 det_val.json）"
$RUN python "$PROJ_ROOT/w1/scripts/coco_or_scalabel_to_yolo.py" \
--labels-json "$BDD_DIR/labels/det_20/det_val.json" \
--images-root "$BDD_DIR/images/100k/val" \
--labels-root "$BDD_DIR/labels/val" \
--classes car bus truck person bicycle motorcycle "traffic light" "traffic sign"
### ----------------------- 生成 train_5k.txt -----------------------
echo "==> 写入 train_5k.txt（绝对路径，最稳）"
$RUN python - <<PY
import json, os, pathlib
subset = r"$BDD_DIR/labels/det_20/det_train_5k.json"
img_root = r"$BDD_DIR/images/100k/train"
d = json.load(open(subset, "r", encoding="utf-8"))
names = {os.path.basename(x["file_name"]) for x in d["images"]}
out = r"$BDD_DIR/train_5k.txt"
with open(out, "w") as f:
for n in sorted(names):
f.write(str(pathlib.Path(img_root)/n) + "\n")
print("wrote", len(names), "lines to", out)
PY
### ----------------------- 写 data.yaml -----------------------
echo "==> 写入数据配置 data/bdd100k/bdd5k.yaml"
cat > "$BDD_DIR/bdd5k.yaml" <<YAML
# Week1 dataset config (BDD100K 5k subset + val10k)
path: $BDD_DIR
train: train_5k.txt
val: images/100k/val
names:
0: car
1: bus
2: truck
3: person
4: bicycle
5: motorcycle
6: traffic light
7: traffic sign
YAML
### ----------------------- 清缓存 & 自检 -----------------------
echo "==> 清理 .cache 并快速自检"
rm -f "$BDD_DIR/labels/100k/train.cache" "$BDD_DIR/labels/100k/val.cache" || true
$RUN python - <<'PY'
from glob import glob
print("train txt:", len(glob("data/bdd100k/labels/train/*.txt")))
print("val txt:", len(glob("data/bdd100k/labels/val/*.txt")))
PY
### ----------------------- 训练 / 验证 / 预测 -----------------------
echo "==> 开始训练（$EPOCHS epochs, imgsz=$TRAIN_IMGSZ, batch=$BATCH, device=$DEVICE）"
$RUN yolo detect train data="$BDD_DIR/bdd5k.yaml" \
model=yolov8n.pt imgsz="$TRAIN_IMGSZ" epochs="$EPOCHS" batch="$BATCH" workers=8 \
device="$DEVICE" name="$RUN_NAME"
BEST_PT="$PROJ_ROOT/runs/detect/$RUN_NAME/weights/best.pt"
echo "==> 验证 best.pt: $BEST_PT"
$RUN yolo detect val data="$BDD_DIR/bdd5k.yaml" model="$BEST_PT" imgsz="$TRAIN_IMGSZ" device="$DEVICE" plots=True
echo "==> 导出 ONNX 并做一次推理验证"
$RUN yolo export model="$BEST_PT" format=onnx opset=12
$RUN yolo detect predict source="$BDD_DIR/images/100k/val" \
model="$PROJ_ROOT/runs/detect/$RUN_NAME/weights/best.onnx" \
conf=0.25 save=True device="$DEVICE" max_det=300
echo "✅ ALL DONE. 结果目录：runs/detect/$RUN_NAME"