#!/bin/bash
# Phase 4A: BEV 2x分辨率提升训练
# 参考start_enhanced_training_fixed.sh的成功配置

set -e

export PATH=/opt/conda/bin:$PATH
export LD_LIBRARY_PATH=/opt/conda/lib:/opt/conda/lib/python3.8/site-packages/torch/lib:/usr/local/cuda/lib64:$LD_LIBRARY_PATH

echo "========================================================================"
echo "Phase 4A: BEV 2x分辨率提升训练"
echo "========================================================================"
echo "配置: BEV 0.15m分辨率 (2倍提升)"
echo "Decoder: 4层完整版 [256, 256, 128, 128]"
echo "从epoch_23.pth加载所有权重"
echo "========================================================================"
echo ""
echo "关键配置:"
echo "  - BEV分辨率: 0.3m → 0.15m (720×720)"
echo "  - GT标签: 0.25m → 0.125m (800×800)"
echo "  - Decoder: 2层 → 4层"
echo "  - Deep Supervision: 启用"
echo "  - Dice Loss: 启用"
echo "  - GPU数量: 6"
echo "  - Batch: 1/GPU (显存限制)"
echo "  - Workers: 4 (参考Phase 3成功配置)"
echo ""
echo "预期性能提升:"
echo "  - Stop Line IoU: 0.27 → 0.42+ (+55%)"
echo "  - Divider IoU: 0.19 → 0.35+ (+84%)"
echo "  - 整体mIoU: 0.41 → 0.54+ (+32%)"
echo ""
echo "预计训练时间: 12.5天"
echo "========================================================================"
echo ""

# 创建日志文件
LOG_FILE="phase4a_bev2x_$(date +%Y%m%d_%H%M%S).log"

echo "开始训练..."
echo "日志文件: $LOG_FILE"
echo ""

# 启动训练 - 完全参考start_enhanced_training_fixed.sh的格式
/opt/conda/bin/torchpack dist-run -np 6 /opt/conda/bin/python tools/train.py \
  configs/nuscenes/det/transfusion/secfpn/camera+lidar/swint_v0p075/multitask_BEV2X_phase4a.yaml \
  --model.encoders.camera.backbone.init_cfg.checkpoint pretrained/swint-nuimages-pretrained.pth \
  --load_from runs/enhanced_from_epoch19/epoch_23.pth \
  --data.samples_per_gpu 1 \
  --data.workers_per_gpu 4 \
  2>&1 | tee "$LOG_FILE"

echo ""
echo "========================================================================"
echo "训练完成！日志保存在: $LOG_FILE"
echo "========================================================================"