#!/bin/bash
# 从epoch_19开始训练Enhanced版本 (改进版)
# 使用配置文件而非命令行参数

set -e

export PATH=/opt/conda/bin:$PATH
cd /workspace/bevfusion

echo "=========================================="
echo "BEVFusion Enhanced训练 - 从Epoch 19继续"
echo "任务: 3D检测 + BEV分割（Enhanced Head）"
echo "=========================================="
echo ""
echo "配置信息："
echo "  配置文件: multitask_enhanced.yaml"
echo "  GPU数量: 4x Tesla V100S"
echo "  基础模型: epoch_19.pth (515MB)"
echo "  分割头: EnhancedBEVSegmentationHead"
echo ""
echo "Enhanced功能："
echo "  ✅ ASPP - 多尺度特征"
echo "  ✅ Channel Attention - 通道注意力"
echo "  ✅ Spatial Attention - 空间注意力"
echo "  ✅ Deep Supervision - 深度监督"
echo "  ✅ Dice Loss - 区域相似度"
echo "  ✅ Class Weights - 类别平衡"
echo ""
echo "优化措施："
echo "  - 使用4个GPU（避免8GPU同步死锁）"
echo "  - workers_per_gpu=0（避免共享内存问题）"
echo "  - samples_per_gpu=2（降低内存压力）"
echo "  - 学习率=1e-4（微调模式）"
echo "  - map loss权重=5.0（强化分割学习）"
echo ""
echo "预计训练时间: 6-8小时"
echo "预期性能提升:"
echo "  - 分割mIoU: 36% → 50-55%"
echo "  - 检测mAP: 保持65-68%"
echo ""
echo "⚠️  注意: 当前有训练正在运行！"
echo "   如需停止，请运行: bash STOP_CURRENT_TRAINING.sh"
echo ""
echo "=========================================="
echo ""

# 生成日志文件名
LOG_FILE="training_enhanced_from_epoch19_$(date +%Y%m%d_%H%M%S).log"

echo "训练日志: $LOG_FILE"
echo ""

# 确认继续
read -p "是否继续启动Enhanced训练？(y/n): " -n 1 -r
echo ""

if [[ ! $REPLY =~ ^[Yy]$ ]]; then
    echo "已取消"
    exit 0
fi

echo "开始训练..."
echo ""

# 启动训练
torchpack dist-run -np 4 python tools/train.py \
  configs/nuscenes/det/transfusion/secfpn/camera+lidar/swint_v0p075/multitask_enhanced.yaml \
  --model.encoders.camera.backbone.init_cfg.checkpoint pretrained/swint-nuimages-pretrained.pth \
  2>&1 | tee "$LOG_FILE"

echo ""
echo "=========================================="
echo "训练完成！"
echo "日志文件: $LOG_FILE"
echo "=========================================="