# Phase 4A Stage 1 - 8卡训练快速参考

**版本**: v1.0  
**更新**: 2025-11-01  

---

## 🚀 快速启动

### 启动训练
```bash
cd /workspace/bevfusion
nohup bash START_FROM_EPOCH1.sh > /tmp/train_8gpu.log 2>&1 &
```

### 监控状态
```bash
# GPU状态
nvidia-smi

# 训练进度
tail -f $(ls -t phase4a_stage1_new_*.log | head -1) | grep "Epoch"

# 磁盘空间
df -h | grep -E "/workspace|/data"
```

---

## 📊 关键配置

```yaml
GPU: 8×Tesla V100S-32GB
Batch: 1/GPU (总batch=8)
分辨率: 600×600 GT
Epochs: 10
预计耗时: 9.5天
输出: /data/runs/phase4a_stage1/
```

---

## 🎯 预期性能

**Epoch 1 (11/2完成)**
- Stop Line: 0.27 → 0.30+
- Divider: 0.19 → 0.22+
- mIoU: 0.41 → 0.43+

**Epoch 10 (11/10完成)**
- Stop Line: 0.27 → 0.35+
- Divider: 0.19 → 0.28+
- mIoU: 0.41 → 0.48+

---

## 🔧 常用命令

```bash
# 停止训练
pkill -f "train.py"

# 清理评估文件
bash cleanup_eval_hook.sh

# 查看checkpoints
ls -lh /data/runs/phase4a_stage1/

# 查看进程
ps aux | grep "train.py" | grep -v grep
```

---

## 📝 完整文档

详见: `project/docs/Phase4A_Stage1_8GPU配置_20251101.md`