bev-project/archive_scripts/一键剪枝和微调.sh

#!/bin/bash
# 一键执行剪枝+微调流程

set -e

cd /workspace/bevfusion

echo "========================================================================"
echo "BEVFusion 模型剪枝+微调 完整流程"
echo "========================================================================"
echo ""
echo "Baseline: Epoch 23"
echo "  - 参数量: 45.72M"
echo "  - NDS: 0.6941, mAP: 0.6446, mIoU: 0.4130"
echo ""
echo "目标: 剪枝30% → 32M参数"
echo "预期: 精度损失<2%"
echo ""

# 配置
CHECKPOINT="runs/enhanced_from_epoch19/epoch_23.pth"
PRUNED_OUTPUT="pruning_results/bevfusion_pruned_32M.pth"
TARGET_RATIO=0.70  # 保留70%参数（剪枝30%）

TIMESTAMP=$(date +%Y%m%d_%H%M%S)
FINETUNE_DIR="runs/pruned_finetune_${TIMESTAMP}"

# 检查checkpoint
if [ ! -f "$CHECKPOINT" ]; then
    echo "错误: Checkpoint不存在: $CHECKPOINT"
    exit 1
fi

# 创建输出目录
mkdir -p pruning_results
mkdir -p "$FINETUNE_DIR"

echo "========================================================================"
echo "阶段1: 模型剪枝（预计15分钟）"
echo "========================================================================"
echo ""

# 执行剪枝
/opt/conda/bin/python tools/pruning/prune_bevfusion_builtin.py \
    --checkpoint "$CHECKPOINT" \
    --output "$PRUNED_OUTPUT" \
    --target-ratio $TARGET_RATIO \
    2>&1 | tee "pruning_results/pruning_log_${TIMESTAMP}.txt"

if [ $? -ne 0 ]; then
    echo "错误: 剪枝失败"
    exit 1
fi

echo ""
echo "✅ 剪枝完成！"
echo "   输出: $PRUNED_OUTPUT"
echo ""

# 询问是否继续微调
echo "========================================================================"
echo "阶段2: 微调训练（预计12-15小时）"
echo "========================================================================"
echo ""
echo "是否立即开始微调训练？"
echo ""
echo "微调配置:"
echo "  - Epochs: 3"
echo "  - 学习率: 5e-6 (很小)"
echo "  - GPU: 8张"
echo "  - 预计时间: 12-15小时"
echo ""
echo "选项:"
echo "  [1] 立即开始微调（后台运行）"
echo "  [2] 稍后手动启动"
echo "  [3] 查看剪枝结果后再决定"
echo ""

read -p "请选择 [1/2/3]: " choice

case $choice in
    1)
        echo ""
        echo "启动微调训练..."
        
        CONFIG="configs/nuscenes/det/transfusion/secfpn/camera+lidar/swint_v0p075/multitask_enhanced_phase1_HIGHRES.yaml"
        
        # 后台启动微调
        nohup /opt/conda/bin/torchpack dist-run -np 8 /opt/conda/bin/python tools/train.py \
            "$CONFIG" \
            --load_from "$PRUNED_OUTPUT" \
            --run-dir "$FINETUNE_DIR" \
            --cfg-options \
                max_epochs=3 \
                optimizer.lr=5.0e-6 \
                data.samples_per_gpu=2 \
                data.workers_per_gpu=0 \
            2>&1 | tee "${FINETUNE_DIR}/finetune.log" &
        
        FINETUNE_PID=$!
        echo $FINETUNE_PID > pruning_results/finetune.pid
        
        echo ""
        echo "✅ 微调训练已启动（后台运行）"
        echo "   PID: $FINETUNE_PID"
        echo "   日志: ${FINETUNE_DIR}/finetune.log"
        echo ""
        echo "监控命令:"
        echo "  tail -f ${FINETUNE_DIR}/finetune.log | grep 'Epoch'"
        echo ""
        ;;
    2)
        echo ""
        echo "稍后手动启动微调。"
        echo ""
        echo "启动命令:"
        echo "  torchpack dist-run -np 8 python tools/train.py \\"
        echo "    configs/.../multitask_enhanced_phase1_HIGHRES.yaml \\"
        echo "    --load_from $PRUNED_OUTPUT \\"
        echo "    --cfg-options max_epochs=3 optimizer.lr=5.0e-6"
        echo ""
        ;;
    3)
        echo ""
        echo "查看剪枝结果:"
        echo "  python tools/analysis/analyze_checkpoint.py $PRUNED_OUTPUT"
        echo ""
        echo "如果满意，再启动微调。"
        echo ""
        ;;
    *)
        echo "无效选择，退出。"
        exit 1
        ;;
esac

echo "========================================================================"
echo "剪枝流程完成"
echo "========================================================================"
echo ""
echo "已生成文件:"
echo "  - $PRUNED_OUTPUT"
echo "  - pruning_results/pruning_log_${TIMESTAMP}.txt"
echo ""
echo "下一步:"
echo "  1. 等待微调完成（如已启动）"
echo "  2. 评估剪枝+微调后的模型"
echo "  3. 进行INT8量化"
echo ""