bev-project/tools/analysis/model_complexity_simple.py

#!/usr/bin/env python3
"""
简化版模型复杂度分析
不依赖thop，直接统计参数
"""

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

def count_parameters(model):
    """统计模型参数"""
    total_params = 0
    trainable_params = 0
    
    for param in model.parameters():
        total_params += param.numel()
        if param.requires_grad:
            trainable_params += param.numel()
    
    return total_params, trainable_params

def analyze_by_module(model):
    """按模块统计参数"""
    module_stats = {}
    
    for name, module in model.named_children():
        params = sum(p.numel() for p in module.parameters())
        module_stats[name] = params
    
    return module_stats

def main(config_file, checkpoint_file):
    """主函数"""
    from mmcv import Config
    from mmdet3d.models import build_model
    
    print("=" * 80)
    print("BEVFusion模型复杂度分析")
    print("=" * 80)
    print(f"配置文件: {config_file}")
    print(f"Checkpoint: {checkpoint_file}")
    print()
    
    # 加载配置
    print("加载配置...")
    cfg = Config.fromfile(config_file)
    
    # 构建模型
    print("构建模型...")
    model = build_model(cfg.model)
    
    # 加载权重
    if checkpoint_file:
        print(f"加载checkpoint...")
        checkpoint = torch.load(checkpoint_file, map_location='cpu')
        if 'state_dict' in checkpoint:
            model.load_state_dict(checkpoint['state_dict'], strict=False)
        else:
            model.load_state_dict(checkpoint, strict=False)
    
    model.eval()
    
    # 统计总参数
    print("\n" + "=" * 80)
    print("总体统计")
    print("=" * 80)
    
    total_params, trainable_params = count_parameters(model)
    
    print(f"总参数量:     {total_params:,} ({total_params/1e6:.2f}M)")
    print(f"可训练参数:   {trainable_params:,} ({trainable_params/1e6:.2f}M)")
    print(f"模型大小(FP32): {total_params * 4 / 1024 / 1024:.2f} MB")
    print(f"模型大小(FP16): {total_params * 2 / 1024 / 1024:.2f} MB")
    print(f"模型大小(INT8): {total_params * 1 / 1024 / 1024:.2f} MB")
    
    # 按模块统计
    print("\n" + "=" * 80)
    print("各模块参数统计")
    print("=" * 80)
    
    module_stats = analyze_by_module(model)
    
    # 排序并显示
    sorted_modules = sorted(module_stats.items(), key=lambda x: x[1], reverse=True)
    
    for name, params in sorted_modules:
        percentage = params / total_params * 100
        print(f"{name:30s}: {params:12,} ({params/1e6:6.2f}M, {percentage:5.2f}%)")
    
    # 详细分析编码器
    print("\n" + "=" * 80)
    print("Encoders详细分析")
    print("=" * 80)
    
    if hasattr(model, 'encoders'):
        for enc_name in ['camera', 'lidar', 'radar']:
            if enc_name in model.encoders:
                encoder = model.encoders[enc_name]
                enc_params = sum(p.numel() for p in encoder.parameters())
                print(f"\n{enc_name.upper()} Encoder: {enc_params:,} ({enc_params/1e6:.2f}M)")
                
                for name, module in encoder.named_children():
                    params = sum(p.numel() for p in module.parameters())
                    if params > 0:
                        print(f"  {name:28s}: {params:12,} ({params/1e6:6.2f}M)")
    
    # 详细分析heads
    print("\n" + "=" * 80)
    print("Heads详细分析")
    print("=" * 80)
    
    if hasattr(model, 'heads'):
        for head_name, head in model.heads.items():
            head_params = sum(p.numel() for p in head.parameters())
            print(f"\n{head_name.upper()} Head: {head_params:,} ({head_params/1e6:.2f}M)")
            
            for name, module in head.named_children():
                params = sum(p.numel() for p in module.parameters())
                if params > 0:
                    print(f"  {name:28s}: {params:12,} ({params/1e6:6.2f}M)")
    
    # 优化建议
    print("\n" + "=" * 80)
    print("优化建议")
    print("=" * 80)
    
    # 找出最大的模块
    largest_module = max(sorted_modules, key=lambda x: x[1])
    print(f"\n1. 最大模块: {largest_module[0]} ({largest_module[1]/1e6:.2f}M, {largest_module[1]/total_params*100:.1f}%)")
    print(f"   建议: 优先剪枝此模块，预期可减少30-40%参数")
    
    # 剪枝潜力估算
    print(f"\n2. 剪枝潜力估算:")
    print(f"   保守剪枝(20%): {total_params*0.8/1e6:.2f}M参数, {total_params*0.8*4/1024/1024:.2f}MB")
    print(f"   中等剪枝(40%): {total_params*0.6/1e6:.2f}M参数, {total_params*0.6*4/1024/1024:.2f}MB")
    print(f"   激进剪枝(60%): {total_params*0.4/1e6:.2f}M参数, {total_params*0.4*4/1024/1024:.2f}MB")
    
    # 量化收益
    print(f"\n3. 量化收益:")
    print(f"   FP32→FP16: {total_params*4/1024/1024:.2f}MB → {total_params*2/1024/1024:.2f}MB (-50%)")
    print(f"   FP32→INT8: {total_params*4/1024/1024:.2f}MB → {total_params*1/1024/1024:.2f}MB (-75%)")
    
    # 推荐优化路线
    print(f"\n4. 推荐优化路线:")
    target_params = total_params * 0.6  # 40%剪枝
    print(f"   Step 1: 剪枝40% → {target_params/1e6:.2f}M参数")
    print(f"   Step 2: INT8量化 → {target_params*1/1024/1024:.2f}MB模型")
    print(f"   预期速度提升: 2-3倍")
    print(f"   预期精度损失: <3%")
    
    print("\n" + "=" * 80)
    print("分析完成！")
    print("=" * 80)

if __name__ == '__main__':
    if len(sys.argv) < 3:
        print("用法: python model_complexity_simple.py <config> <checkpoint>")
        print("\n示例:")
        print("  python tools/analysis/model_complexity_simple.py \\")
        print("    configs/.../multitask_enhanced_phase1_HIGHRES.yaml \\")
        print("    runs/enhanced_from_epoch19/epoch_23.pth")
        sys.exit(1)
    
    config_file = sys.argv[1]
    checkpoint_file = sys.argv[2]
    
    main(config_file, checkpoint_file)