80 lines
3.6 KiB
Plaintext
80 lines
3.6 KiB
Plaintext
================================================================================
|
|
BEVFusion Checkpoint分析
|
|
================================================================================
|
|
文件: runs/enhanced_from_epoch19/epoch_23.pth
|
|
|
|
加载checkpoint...
|
|
Checkpoint包含: state_dict + 其他元数据
|
|
Meta信息: {'fp16': {'loss_scaler': {'scale': 1024.0, 'growth_factor': 2.0, 'backoff_factor': 0.5, 'growth_interval': 2000, '_growth_tracker': 903}}, 'hook_msgs': {'last_ckpt': '/workspace/bevfusion/runs/enhanced_from_epoch19/epoch_22.pth'}, 'epoch': 23, 'iter': 236877, 'mmcv_version': '1.4.0', 'time': 'Wed Oct 29 23:21:47 2025'}
|
|
|
|
================================================================================
|
|
参数统计
|
|
================================================================================
|
|
总参数量: 45,721,876 (45.72M)
|
|
层数: 636
|
|
模型大小(FP32): 174.42 MB
|
|
模型大小(FP16): 87.21 MB
|
|
模型大小(INT8): 43.60 MB
|
|
|
|
================================================================================
|
|
各模块参数分布
|
|
================================================================================
|
|
encoders : 34,449,014 ( 34.45M, 75.34%, 131.41MB)
|
|
heads : 5,915,279 ( 5.92M, 12.94%, 22.56MB)
|
|
decoder : 4,582,414 ( 4.58M, 10.02%, 17.48MB)
|
|
fuser : 775,169 ( 0.78M, 1.70%, 2.96MB)
|
|
|
|
================================================================================
|
|
Encoders子模块详细分析
|
|
================================================================================
|
|
encoders.camera.backbone : 27,549,318 ( 27.55M, 60.25%)
|
|
encoders.lidar.backbone : 2,697,029 ( 2.70M, 5.90%)
|
|
encoders.camera.vtransform : 2,609,319 ( 2.61M, 5.71%)
|
|
encoders.camera.neck : 1,593,348 ( 1.59M, 3.48%)
|
|
|
|
================================================================================
|
|
Heads子模块详细分析
|
|
================================================================================
|
|
heads.map.aspp : 4,131,840 ( 4.13M, 9.04%)
|
|
heads.object.shared_conv : 589,952 ( 0.59M, 1.29%)
|
|
heads.map.classifiers : 443,526 ( 0.44M, 0.97%)
|
|
heads.map.decoder : 295,168 ( 0.30M, 0.65%)
|
|
heads.object.decoder : 233,602 ( 0.23M, 0.51%)
|
|
heads.object.heatmap_head : 159,499 ( 0.16M, 0.35%)
|
|
|
|
================================================================================
|
|
优化建议
|
|
================================================================================
|
|
|
|
1. 最大模块: encoders
|
|
参数量: 34.45M (75.3%)
|
|
建议: 优先剪枝此模块
|
|
如果剪枝30%: 可减少 10.33M参数
|
|
|
|
2. 剪枝潜力估算:
|
|
保守剪枝(20%): 36.58M参数 → 139.53MB (FP32)
|
|
中等剪枝(40%): 27.43M参数 → 104.65MB (FP32)
|
|
激进剪枝(60%): 18.29M参数 → 69.77MB (FP32)
|
|
|
|
3. 量化收益:
|
|
原始FP32: 174.42 MB
|
|
FP16: 87.21 MB (-50%)
|
|
INT8: 43.60 MB (-75%)
|
|
|
|
4. 推荐优化路线:
|
|
原始模型: 45.72M, 174.42MB (FP32)
|
|
↓ 剪枝40%
|
|
剪枝模型: 27.43M, 104.65MB (FP32)
|
|
↓ INT8量化
|
|
最终模型: 27.43M, 26.16MB (INT8)
|
|
|
|
总压缩比: 85.0%
|
|
预期加速: 2.5-3倍
|
|
精度损失: <3%
|
|
|
|
================================================================================
|
|
分析完成！
|
|
================================================================================
|
|
|
|
✅ 总参数量: 45.72M
|