fly6516s_Education.test/isolators-transfer
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

train:调整训练参数以优化绝缘子检测性能

Browse Source 
- 增加训练轮数和批量大小，提高模型收敛速度- 调整学习率范围，加快训练进程
-增强数据增强技术，提高模型泛化能力
- 引入类别权重，解决类别不平衡问题- 优化优化器参数，提高模型训练效果
This commit is contained in:
fly6516 2025-05-29 20:32:57 +08:00
parent 13bd165ff5
commit c208f07e28
3 changed files with 23 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

BIN
best_isolator_model.pt Normal file
View File

Binary file not shown.

0
model.pt Normal file
View File

43
train.py
View File

@ -1,3 +1,4 @@
"""
YOLOv12 绝缘子检测优化训练脚本（最终稳定版）
主要改进：
@ -17,28 +18,30 @@ import numpy as np
# +++ 有效参数配置区块 +++
TRAIN_CONFIG = {
"stage1": {
"freeze": [0,1,2,3,4], # 冻结前5层
"epochs": 50,
"batch": 16,
"lr0": 0.005, # 初始学习率
"lrf": 0.05, # 最终学习率
"hsv_h": 0.01, # 色相增强幅度
"degrees": 5.0, # 旋转角度范围
"perspective": 0.001, # 透视变换
"flipud": 0.1, # 上下翻转概率
"optimizer": "SGD", # 优化器类型
"weight_decay": 0.001, # 权重衰减
"label_smoothing": 0.1 # 标签平滑
"freeze": [0, 1, 2, 3, 4], # 冻结前5层
"epochs": 75, # 增加训练轮数
"batch": 32, # 增加批量大小
"lr0": 0.001, # 降低初始学习率
"lrf": 0.1, # 提高最终学习率
"hsv_h": 0.01, # 色相增强幅度
"degrees": 10.0, # 增加旋转角度范围
"perspective": 0.005, # 增加透视变换
"flipud": 0.2, # 增加上下翻转概率
"optimizer": "SGD", # 优化器类型
"weight_decay": 0.001, # 权重衰减
"label_smoothing": 0.1, # 标签平滑
"mosaic": True, # 启用Mosaic增强
"class_weights": [1.0, 2.0] # 设置类别权重，insulator权重更高
},
"stage2": {
"epochs": 100,
"batch": 8,
"lr0": 0.0005,
"lrf": 0.005,
"mixup": 0.15, # MixUp增强系数
"close_mosaic": 10, # 最后10epoch关闭Mosaic
"optimizer": "AdamW", # 优化器切换
"warmup_epochs": 5 # 学习率预热
"epochs": 150, # 增加训练轮数
"batch": 16, # 增加批量大小
"lr0": 0.0001, # 降低初始学习率
"lrf": 0.01, # 提高最终学习率
"mixup": 0.2, # 增加MixUp增强系数
"close_mosaic": 10, # 最后10epoch关闭Mosaic
"optimizer": "AdamW", # 切换优化器为AdamW
"warmup_epochs": 10 # 增大学习率预热轮数
}
}