#!/bin/bash
# 启动Phase 2: 添加Dice Loss
# 前提：Phase 1 (epoch_23.pth) 已完成

set -e

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

echo "=========================================="
echo "Phase 2: 添加Dice Loss"
echo "=========================================="
echo ""
echo "增强内容："
echo "  ✅ ASPP (Phase 1)"
echo "  ✅ Dice Loss - 改善小目标"
echo "  ❌ Deep Supervision (Phase 3)"
echo "  ❌ 高权重 (Phase 4)"
echo ""
echo "训练设置："
echo "  基础模型: epoch_23.pth"
echo "  目标: epoch_26.pth"
echo "  学习率: 5e-5"
echo "  预期提升: mIoU +3-5% (累计)"
echo ""
echo "=========================================="
echo ""

# 查找最新的epoch_23.pth
EPOCH_23_PATH=$(find runs/ -name "epoch_23.pth" -type f | sort -r | head -1)

if [ -z "$EPOCH_23_PATH" ]; then
    echo "错误: 找不到epoch_23.pth"
    echo "请先完成Phase 1训练"
    exit 1
fi

echo "找到Phase 1模型: $EPOCH_23_PATH"
echo ""
echo "开始Phase 2训练..."
echo ""

LOG_FILE="training_phase2_$(date +%Y%m%d_%H%M%S).log"

torchpack dist-run -np 8 python tools/train.py \
  configs/nuscenes/det/transfusion/secfpn/camera+lidar/swint_v0p075/multitask_enhanced_phase2.yaml \
  --model.encoders.camera.backbone.init_cfg.checkpoint pretrained/swint-nuimages-pretrained.pth \
  --load_from "$EPOCH_23_PATH" \
  --data.workers_per_gpu 0 \
  2>&1 | tee "$LOG_FILE"

echo ""
echo "=========================================="
echo "Phase 2训练完成！"
echo "日志: $LOG_FILE"
echo "下一步: 运行 bash scripts/start_phase3.sh"
echo "=========================================="