bev-project/tools/analysis/analyze_checkpoint.py

#!/usr/bin/env python3
"""
直接分析checkpoint文件
不需要构建完整模型
"""

import torch
import sys
from collections import defaultdict

def analyze_checkpoint(checkpoint_file):
    """分析checkpoint文件"""
    
    print("=" * 80)
    print("BEVFusion Checkpoint分析")
    print("=" * 80)
    print(f"文件: {checkpoint_file}")
    print()
    
    # 加载checkpoint
    print("加载checkpoint...")
    checkpoint = torch.load(checkpoint_file, map_location='cpu')
    
    # 获取state_dict
    if 'state_dict' in checkpoint:
        state_dict = checkpoint['state_dict']
        print(f"Checkpoint包含: state_dict + 其他元数据")
        if 'epoch' in checkpoint:
            print(f"Epoch: {checkpoint['epoch']}")
        if 'meta' in checkpoint:
            print(f"Meta信息: {checkpoint.get('meta', {})}")
    else:
        state_dict = checkpoint
        print(f"Checkpoint类型: 纯state_dict")
    
    print()
    
    # 统计参数
    print("=" * 80)
    print("参数统计")
    print("=" * 80)
    
    total_params = 0
    layer_count = 0
    
    for key, value in state_dict.items():
        if isinstance(value, torch.Tensor):
            params = value.numel()
            total_params += params
            layer_count += 1
    
    print(f"总参数量:     {total_params:,} ({total_params/1e6:.2f}M)")
    print(f"层数:         {layer_count}")
    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")
    print()
    
    # 按模块统计
    print("=" * 80)
    print("各模块参数分布")
    print("=" * 80)
    
    module_params = defaultdict(int)
    
    for key, value in state_dict.items():
        if not isinstance(value, torch.Tensor):
            continue
        
        # 提取模块名（取第一级）
        parts = key.split('.')
        if len(parts) > 0:
            module_name = parts[0]
            module_params[module_name] += value.numel()
    
    # 排序并显示
    sorted_modules = sorted(module_params.items(), key=lambda x: x[1], reverse=True)
    
    for name, params in sorted_modules:
        percentage = params / total_params * 100
        size_mb = params * 4 / 1024 / 1024
        print(f"{name:30s}: {params:12,} ({params/1e6:6.2f}M, {percentage:5.2f}%, {size_mb:6.2f}MB)")
    
    # 详细分析encoders
    print("\n" + "=" * 80)
    print("Encoders子模块详细分析")
    print("=" * 80)
    
    encoder_submodules = defaultdict(int)
    
    for key, value in state_dict.items():
        if not isinstance(value, torch.Tensor):
            continue
        
        parts = key.split('.')
        if len(parts) >= 2 and parts[0] == 'encoders':
            submodule = f"{parts[0]}.{parts[1]}"
            if len(parts) >= 3:
                submodule = f"{parts[0]}.{parts[1]}.{parts[2]}"
            encoder_submodules[submodule] += value.numel()
    
    sorted_enc = sorted(encoder_submodules.items(), key=lambda x: x[1], reverse=True)
    for name, params in sorted_enc:
        if params > 1e5:  # 只显示>0.1M的模块
            percentage = params / total_params * 100
            print(f"{name:50s}: {params:12,} ({params/1e6:6.2f}M, {percentage:5.2f}%)")
    
    # 详细分析heads
    print("\n" + "=" * 80)
    print("Heads子模块详细分析")
    print("=" * 80)
    
    head_submodules = defaultdict(int)
    
    for key, value in state_dict.items():
        if not isinstance(value, torch.Tensor):
            continue
        
        parts = key.split('.')
        if len(parts) >= 2 and parts[0] == 'heads':
            submodule = f"{parts[0]}.{parts[1]}"
            if len(parts) >= 3:
                submodule = f"{parts[0]}.{parts[1]}.{parts[2]}"
            head_submodules[submodule] += value.numel()
    
    sorted_heads = sorted(head_submodules.items(), key=lambda x: x[1], reverse=True)
    for name, params in sorted_heads:
        if params > 1e5:  # 只显示>0.1M的模块
            percentage = params / total_params * 100
            print(f"{name:50s}: {params:12,} ({params/1e6:6.2f}M, {percentage:5.2f}%)")
    
    # 优化建议
    print("\n" + "=" * 80)
    print("优化建议")
    print("=" * 80)
    
    # 找出最大的模块
    if sorted_modules:
        largest = sorted_modules[0]
        print(f"\n1. 最大模块: {largest[0]}")
        print(f"   参数量: {largest[1]/1e6:.2f}M ({largest[1]/total_params*100:.1f}%)")
        print(f"   建议: 优先剪枝此模块")
        print(f"   如果剪枝30%: 可减少 {largest[1]*0.3/1e6:.2f}M参数")
    
    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 (FP32)")
    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 (FP32)")
    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 (FP32)")
    
    print(f"\n3. 量化收益:")
    print(f"   原始FP32: {total_params*4/1024/1024:.2f} MB")
    print(f"   FP16:     {total_params*2/1024/1024:.2f} MB (-50%)")
    print(f"   INT8:     {total_params*1/1024/1024:.2f} MB (-75%)")
    
    # 推荐路线
    pruned_params = total_params * 0.6  # 40%剪枝
    print(f"\n4. 推荐优化路线:")
    print(f"   原始模型:   {total_params/1e6:.2f}M, {total_params*4/1024/1024:.2f}MB (FP32)")
    print(f"   ↓ 剪枝40%")
    print(f"   剪枝模型:   {pruned_params/1e6:.2f}M, {pruned_params*4/1024/1024:.2f}MB (FP32)")
    print(f"   ↓ INT8量化")
    print(f"   最终模型:   {pruned_params/1e6:.2f}M, {pruned_params*1/1024/1024:.2f}MB (INT8)")
    print(f"   ")
    print(f"   总压缩比:   {(1 - pruned_params*1/1024/1024 / (total_params*4/1024/1024))*100:.1f}%")
    print(f"   预期加速:   2.5-3倍")
    print(f"   精度损失:   <3%")
    
    print("\n" + "=" * 80)
    print("分析完成！")
    print("=" * 80)
    
    return total_params, module_params

if __name__ == '__main__':
    if len(sys.argv) < 2:
        print("用法: python analyze_checkpoint.py <checkpoint_file>")
        print("\n示例:")
        print("  python tools/analysis/analyze_checkpoint.py \\")
        print("    runs/enhanced_from_epoch19/epoch_23.pth")
        sys.exit(1)
    
    checkpoint_file = sys.argv[1]
    
    total_params, module_params = analyze_checkpoint(checkpoint_file)
    
    print(f"\n✅ 总参数量: {total_params/1e6:.2f}M")