bev-project/archive_scripts/准备剪枝工具.sh

#!/bin/bash
# 准备模型剪枝工具和环境

set -e

cd /workspace/bevfusion

echo "========================================================================"
echo "准备模型剪枝工具"
echo "========================================================================"
echo ""

# 1. 检查torch-pruning
echo "========== 1. 检查torch-pruning =========="
echo ""

if /opt/conda/bin/python -c "import torch_pruning" 2>/dev/null; then
    echo "✅ torch-pruning已安装"
    VERSION=$(/opt/conda/bin/python -c "import torch_pruning as tp; print(tp.__version__)" 2>/dev/null || echo "未知")
    echo "   版本: $VERSION"
else
    echo "⏳ torch-pruning未安装，正在安装..."
    /opt/conda/bin/pip install torch-pruning -i https://pypi.tuna.tsinghua.edu.cn/simple
    
    if [ $? -eq 0 ]; then
        echo "✅ torch-pruning安装成功"
    else
        echo "⚠️ torch-pruning安装失败，尝试使用PyTorch内置剪枝"
    fi
fi

echo ""

# 2. 创建剪枝工具目录
echo "========== 2. 创建工具目录 =========="
echo ""

mkdir -p tools/pruning
mkdir -p pruning_results

echo "✅ 目录已创建:"
echo "   tools/pruning/      # 剪枝脚本"
echo "   pruning_results/    # 剪枝结果"
echo ""

# 3. 创建简单的剪枝测试脚本
echo "========== 3. 创建剪枝测试脚本 =========="
echo ""

cat > tools/pruning/test_pruning.py << 'PYTHON_EOF'
#!/usr/bin/env python3
"""
测试torch-pruning是否可用
"""

import sys
sys.path.insert(0, '/workspace/bevfusion')

def test_torch_pruning():
    """测试torch-pruning"""
    try:
        import torch_pruning as tp
        print(f"✅ torch-pruning可用")
        print(f"   版本: {tp.__version__}")
        return True
    except ImportError:
        print(f"❌ torch-pruning不可用")
        return False

def test_pytorch_builtin():
    """测试PyTorch内置剪枝"""
    try:
        import torch.nn.utils.prune as prune
        print(f"✅ PyTorch内置剪枝可用")
        return True
    except ImportError:
        print(f"❌ PyTorch内置剪枝不可用")
        return False

def test_checkpoint_load():
    """测试checkpoint加载"""
    import torch
    
    checkpoint_file = 'runs/enhanced_from_epoch19/epoch_23.pth'
    
    try:
        checkpoint = torch.load(checkpoint_file, map_location='cpu')
        
        if 'state_dict' in checkpoint:
            state_dict = checkpoint['state_dict']
            total_params = sum(v.numel() for v in state_dict.values() if isinstance(v, torch.Tensor))
            
            print(f"✅ Checkpoint加载成功")
            print(f"   参数量: {total_params/1e6:.2f}M")
            print(f"   Epoch: {checkpoint.get('epoch', 'unknown')}")
            return True
        else:
            print(f"⚠️ Checkpoint格式异常")
            return False
            
    except Exception as e:
        print(f"❌ Checkpoint加载失败: {e}")
        return False

if __name__ == '__main__':
    print("=" * 60)
    print("剪枝工具可用性测试")
    print("=" * 60)
    print()
    
    # 测试torch-pruning
    print("1. torch-pruning:")
    has_tp = test_torch_pruning()
    print()
    
    # 测试PyTorch内置
    print("2. PyTorch内置剪枝:")
    has_builtin = test_pytorch_builtin()
    print()
    
    # 测试checkpoint
    print("3. Checkpoint加载:")
    can_load = test_checkpoint_load()
    print()
    
    # 总结
    print("=" * 60)
    print("测试结果总结")
    print("=" * 60)
    
    if has_tp:
        print("✅ 推荐使用: torch-pruning (功能更强大)")
    elif has_builtin:
        print("✅ 可以使用: PyTorch内置剪枝 (基本功能)")
    else:
        print("❌ 需要安装剪枝工具")
    
    if can_load:
        print("✅ Checkpoint可正常加载")
    else:
        print("❌ Checkpoint加载有问题")
    
    print()
    print("准备就绪！可以开始剪枝。")

PYTHON_EOF

chmod +x tools/pruning/test_pruning.py

echo "✅ 测试脚本已创建: tools/pruning/test_pruning.py"
echo ""

# 4. 运行测试
echo "========== 4. 运行可用性测试 =========="
echo ""

/opt/conda/bin/python tools/pruning/test_pruning.py

echo ""

# 5. 生成剪枝方案文件
echo "========== 5. 生成剪枝方案 =========="
echo ""

cat > pruning_results/pruning_plan.md << 'PLAN_EOF'
# BEVFusion剪枝实施方案

**生成时间**: $(date)
**Baseline**: epoch_23.pth (45.72M参数)

---

## 🎯 剪枝目标

```
参数量: 45.72M → 32M (-30%)
模型大小: 174MB → 122MB (FP32)
推理加速: ~30%
精度损失: <2%
```

---

## 📋 分层剪枝计划

### Layer 1: Camera Backbone (27.55M → 20M)

**方法**: 通道剪枝
```
SwinTransformer各stage剪枝:
  Stage 1: 96 channels → 80 (-17%)
  Stage 2: 192 channels → 160 (-17%)
  Stage 3: 384 channels → 320 (-17%)
  Stage 4: 768 channels → 640 (-17%)

预期减少: ~5.5M参数
```

**实施**:
```python
# 使用torch-pruning
import torch_pruning as tp

# 对SwinTransformer剪枝
pruner = tp.pruner.MetaPruner(
    model.encoders['camera'].backbone,
    example_inputs,
    importance=tp.importance.MagnitudeImportance(p=2),
    pruning_ratio=0.17,  # 17%
)
pruner.step()
```

---

### Layer 2: ASPP模块 (4.13M → 3M)

**方法**: 通道剪枝
```
ASPP通道数: 512 → 384 (-25%)
各分支同比例剪枝

预期减少: ~1.1M参数
```

---

### Layer 3: Decoder (4.58M → 3.9M)

**方法**: 通道剪枝
```
Decoder通道: [128, 256] → [96, 192] (-25%)

预期减少: ~0.7M参数
```

---

### Layer 4: Camera VTransform (2.61M → 2.35M)

**方法**: 轻度剪枝
```
通道剪枝: 10%

预期减少: ~0.26M参数
```

---

## 📊 预期结果

```
总计减少: 7.56M参数 (-16.5%)
保守估计: 剪枝后38M参数

如果效果好，可进一步剪枝到32M
```

---

## 🚀 下一步

1. 创建剪枝脚本
2. 小规模测试剪枝
3. 评估剪枝效果
4. 全量剪枝
5. 微调3 epochs

PLAN_EOF

echo "✅ 剪枝方案已生成: pruning_results/pruning_plan.md"
echo ""

# 6. 总结
echo "========================================================================"
echo "准备工作完成！"
echo "========================================================================"
echo ""
echo "已完成:"
echo "  ✅ torch-pruning检查/安装"
echo "  ✅ 工具目录创建"
echo "  ✅ 测试脚本创建"
echo "  ✅ 剪枝方案生成"
echo ""
echo "下一步:"
echo "  1. 查看剪枝方案: cat pruning_results/pruning_plan.md"
echo "  2. 创建剪枝脚本 (明天)"
echo "  3. 开始剪枝实施 (明天)"
echo ""
echo "当前状态:"
echo "  - Stage 1训练: 继续进行中 (不受影响)"
echo "  - 优化准备: ✅ 就绪"
echo ""
echo "========================================================================"