bev-project/tools/analyze_model_params.py

#!/usr/bin/env python3
"""分析BEVFusion模型参数量和结构"""

import torch
import sys
import os

def count_parameters(model):
    """统计模型参数"""
    total_params = sum(p.numel() for p in model.parameters())
    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
    return total_params, trainable_params

def analyze_module_params(model, prefix=''):
    """分析各模块参数量"""
    results = []
    
    for name, module in model.named_children():
        full_name = f"{prefix}.{name}" if prefix else name
        
        # 统计当前模块参数
        params = sum(p.numel() for p in module.parameters())
        trainable = sum(p.numel() for p in module.parameters() if p.requires_grad)
        
        if params > 0:
            results.append({
                'name': full_name,
                'total': params,
                'trainable': trainable,
                'module_type': type(module).__name__
            })
        
        # 递归子模块（只递归一层）
        if len(list(module.children())) > 0:
            sub_results = analyze_module_params(module, full_name)
            results.extend(sub_results)
    
    return results

def format_params(num):
    """格式化参数数量"""
    if num >= 1e9:
        return f"{num/1e9:.2f}B"
    elif num >= 1e6:
        return f"{num/1e6:.2f}M"
    elif num >= 1e3:
        return f"{num/1e3:.2f}K"
    else:
        return str(num)

if __name__ == "__main__":
    print("="*90)
    print(f"{'🔍 BEVFusion 模型参数分析':^90}")
    print("="*90)
    print()
    
    # 加载checkpoint
    checkpoint_path = "/workspace/bevfusion/runs/enhanced_from_epoch19/epoch_7.pth"
    
    if not os.path.exists(checkpoint_path):
        print(f"❌ Checkpoint不存在: {checkpoint_path}")
        sys.exit(1)
    
    print(f"📂 加载checkpoint: {checkpoint_path}")
    checkpoint = torch.load(checkpoint_path, map_location='cpu')
    
    if 'state_dict' in checkpoint:
        state_dict = checkpoint['state_dict']
    else:
        state_dict = checkpoint
    
    print(f"✅ Checkpoint加载成功")
    print()
    
    # 统计参数
    print("━"*90)
    print("📊 总体参数统计")
    print("━"*90)
    
    total_params = sum(p.numel() for p in state_dict.values())
    print(f"  总参数量: {format_params(total_params)} ({total_params:,})")
    
    # 按模块分组统计
    module_stats = {}
    for key, param in state_dict.items():
        # 提取模块名（第一级）
        parts = key.split('.')
        if len(parts) > 0:
            if parts[0] == 'model':
                module = parts[1] if len(parts) > 1 else 'model'
            else:
                module = parts[0]
            
            if module not in module_stats:
                module_stats[module] = 0
            module_stats[module] += param.numel()
    
    print()
    print("━"*90)
    print("📊 各模块参数分布")
    print("━"*90)
    print()
    print(f"{'模块名':<30} {'参数量':<15} {'占比':<10} {'可视化':<30}")
    print("-"*90)
    
    # 排序显示
    sorted_modules = sorted(module_stats.items(), key=lambda x: x[1], reverse=True)
    
    for module, params in sorted_modules:
        percentage = params / total_params * 100
        bar_length = int(percentage / 2)
        bar = "█" * bar_length
        
        print(f"{module:<30} {format_params(params):<15} {percentage:>6.2f}%   {bar}")
    
    print()
    print("━"*90)
    print("🎯 剪枝潜力分析")
    print("━"*90)
    
    # 分析各模块的剪枝潜力
    pruning_potential = []
    
    for module, params in sorted_modules[:10]:
        if 'encoder' in module.lower():
            potential = "🟢 高 (30-50%)"
        elif 'decoder' in module.lower():
            potential = "🟡 中 (20-30%)"
        elif 'head' in module.lower():
            potential = "🟡 中 (15-25%)"
        elif 'neck' in module.lower() or 'fpn' in module.lower():
            potential = "🟢 高 (25-40%)"
        else:
            potential = "🔵 低 (10-20%)"
        
        pruning_potential.append((module, format_params(params), percentage, potential))
    
    print()
    print(f"{'模块':<30} {'参数量':<15} {'占比':<10} {'剪枝潜力':<20}")
    print("-"*90)
    
    for module, params, pct, potential in pruning_potential:
        print(f"{module:<30} {params:<15} {pct:>6.2f}%   {potential}")
    
    print()
    print("━"*90)
    print("💡 剪枝建议")
    print("━"*90)
    
    print(f"""
总参数量: {format_params(total_params)}

剪枝目标:
  保守目标: 110M → 70M (-36%, 保持>99%精度)
  激进目标: 110M → 50M (-55%, 保持>97%精度)

优先剪枝模块:
  1. Camera Encoder (如果参数量大)
  2. Neck/FPN层
  3. Decoder
  4. 保留Head (影响精度)

预期效果:
  - 推理速度: 1.5-2x加速
  - 显存占用: -30-40%
  - 模型大小: 515MB → 200-280MB
""")
    
    print("="*90)
    print(f"✅ 分析完成! 建议将此报告保存用于剪枝策略制定")
    print("="*90)