当前位置: 首页 > news >正文

实验记录2025/10/14

我现在是把picsize从640变化到了960,而且把原先7000张的训练数据集精简成了3600张的数据集

下面是跑出来的结果:

Validating runs/detect/yolo11-tea-yolo11s36/weights/best.pt...
Ultralytics 8.3.182 🚀 Python-3.10.12 torch-2.4.0a0+07cecf4168.nv24.05 CUDA:0 (NVIDIA A100-SXM4-80GB, 32768MiB)
YOLO11s summary (fused): 100 layers, 9,413,961 parameters, 0 gradientsClass     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 6/6 [00:00<00:00,  6.25it/s]all        181        368      0.663      0.531      0.616      0.437algal leaf spot         69        203       0.73      0.562      0.699      0.517brown blight         77         89      0.688      0.517      0.633       0.45grey blight         66         76      0.572      0.513      0.515      0.342
Speed: 0.2ms preprocess, 1.9ms inference, 0.0ms loss, 1.3ms postprocess per image
Results saved to runs/detect/yolo11-tea-yolo11s36
Ultralytics 8.3.182 🚀 Python-3.10.12 torch-2.4.0a0+07cecf4168.nv24.05 CUDA:0 (NVIDIA A100-SXM4-80GB, 32768MiB)
YOLO11s summary (fused): 100 layers, 9,413,961 parameters, 0 gradients
val: Fast image access ✅ (ping: 0.0±0.0 ms, read: 2.9±0.9 MB/s, size: 47.8 KB)
val: Scanning /home/share/priv/yolo_new/ultralytics-main/ultralytics-main/datasets/teaDiseases/val_10_13/labels.cache... 181 images, 0 backgrounds, 0 corrupt: 100
WARNING ⚠️ cache='ram' may produce non-deterministic training results. Consider cache='disk' as a deterministic alternative if your disk space allows.
val: Caching images (0.5GB RAM): 100%|██████████| 181/181 [00:00<00:00, 193.63it/s]Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 12/12 [00:01<00:00,  6.28it/s]all        181        368      0.671      0.531      0.619      0.466algal leaf spot         69        203      0.733      0.562      0.671      0.525brown blight         77         89        0.7      0.517      0.656      0.498grey blight         66         76      0.582      0.513      0.529      0.375
Speed: 1.6ms preprocess, 3.6ms inference, 0.0ms loss, 1.4ms postprocess per image
Results saved to runs/detect/yolo11-tea-yolo11s362
验证结果: ultralytics.utils.metrics.DetMetrics object with attributes:

现在尝试一下改为yolo11m.yaml+ imgsz=640,看看效果:

Validating runs/detect/yolo11-tea-yolo11s39/weights/best.pt...
Ultralytics 8.3.182 🚀 Python-3.10.12 torch-2.4.0a0+07cecf4168.nv24.05 CUDA:0 (NVIDIA A100-SXM4-80GB, 32768MiB)
YOLO11m summary (fused): 125 layers, 20,032,345 parameters, 0 gradientsClass     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 3/3 [00:08<00:00,  2.91s/it]all        181        368      0.561      0.538      0.574      0.412algal leaf spot         69        203      0.718      0.571      0.736      0.553brown blight         77         89       0.56      0.449      0.514      0.361grey blight         66         76      0.405      0.592      0.473      0.322
Speed: 0.3ms preprocess, 4.4ms inference, 0.0ms loss, 10.2ms postprocess per image
Results saved to runs/detect/yolo11-tea-yolo11s39
Ultralytics 8.3.182 🚀 Python-3.10.12 torch-2.4.0a0+07cecf4168.nv24.05 CUDA:0 (NVIDIA A100-SXM4-80GB, 32768MiB)
YOLO11m summary (fused): 125 layers, 20,032,345 parameters, 0 gradients
val: Fast image access ✅ (ping: 0.0±0.0 ms, read: 2.6±0.8 MB/s, size: 47.8 KB)
val: Scanning /home/share/priv/yolo_new/ultralytics-main/ultralytics-main/datasets/teaDiseases/val_10_13/labels.cache... 181 images, 0 backgrounds, 0 corrupt: 100%|██████████| 181/181
WARNING ⚠️ cache='ram' may produce non-deterministic training results. Consider cache='disk' as a deterministic alternative if your disk space allows.
val: Caching images (0.2GB RAM): 100%|██████████| 181/181 [00:01<00:00, 170.08it/s]Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 6/6 [00:08<00:00,  1.48s/it]all        181        368      0.568      0.538      0.569      0.432algal leaf spot         69        203      0.719      0.571      0.705      0.554brown blight         77         89      0.571      0.449      0.524      0.398grey blight         66         76      0.414      0.592      0.479      0.345
Speed: 1.1ms preprocess, 5.7ms inference, 0.0ms loss, 11.1ms postprocess per image
Results saved to runs/detect/yolo11-tea-yolo11s392
验证结果: ultralytics.utils.metrics.DetMetrics object with attributes:

感觉m对于精度的提高没有imgsz的提高要好的多。

目前看来是yolo11s.yaml+imgsz=960时跑出的效果最好,使用的数据集是train_aug_10_14

现在尝试使用增强grey_light的图像使用960+yolo11s.yaml+train_aug_10_15进行测试,看看效果如何。

300 epochs completed in 4.279 hours.
Optimizer stripped from runs/detect/yolo11-tea-yolo11s56/weights/last.pt, 19.3MB
Optimizer stripped from runs/detect/yolo11-tea-yolo11s56/weights/best.pt, 19.3MBValidating runs/detect/yolo11-tea-yolo11s56/weights/best.pt...
Ultralytics 8.3.182 🚀 Python-3.10.12 torch-2.4.0a0+07cecf4168.nv24.05 CUDA:0 (NVIDIA A100-SXM4-80GB, 32768MiB)
YOLO11s summary (fused): 100 layers, 9,413,961 parameters, 0 gradientsClass     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 6/6 [00:02<00:00,  2.30it/s]all        181        368       0.56      0.596      0.589      0.416algal leaf spot         69        203      0.667      0.709      0.724       0.54brown blight         77         89      0.529      0.539      0.556       0.38grey blight         66         76      0.483      0.539      0.488      0.329
Speed: 0.2ms preprocess, 4.0ms inference, 0.0ms loss, 2.4ms postprocess per image
Results saved to runs/detect/yolo11-tea-yolo11s56
Ultralytics 8.3.182 🚀 Python-3.10.12 torch-2.4.0a0+07cecf4168.nv24.05 CUDA:0 (NVIDIA A100-SXM4-80GB, 32768MiB)
YOLO11s summary (fused): 100 layers, 9,413,961 parameters, 0 gradients
val: Fast image access ✅ (ping: 0.0±0.0 ms, read: 4.3±1.2 MB/s, size: 46.9 KB)
val: Scanning /home/share/priv/yolo_new/ultralytics-main/ultralytics-main/datasets/teaDiseases/val_10_13/labels.cache... 181 images, 0 backgrounds, 0 corrupt: 100%|██████████| 181/181 
WARNING ⚠️ cache='ram' may produce non-deterministic training results. Consider cache='disk' as a deterministic alternative if your disk space allows.
val: Caching images (0.5GB RAM): 100%|██████████| 181/181 [00:01<00:00, 115.39it/s]Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 12/12 [00:03<00:00,  3.93it/s]all        181        368      0.564      0.596      0.588      0.441algal leaf spot         69        203      0.668      0.709        0.7      0.553brown blight         77         89       0.54      0.539       0.56      0.409grey blight         66         76      0.484      0.539      0.503       0.36
Speed: 1.3ms preprocess, 5.1ms inference, 0.0ms loss, 2.3ms postprocess per image
Results saved to runs/detect/yolo11-tea-yolo11s562
验证结果: ultralytics.utils.metrics.DetMetrics object with attributes:

结果依然没有上次的好。

现在在最好的基础上加入无参数的SimAM模块,并且更改val的验证集,查看变化。

 

http://www.hskmm.com/?act=detail&tid=32350

相关文章:

  • 个人微信开发框架
  • 投资指标技术分析
  • linux源码编译python
  • uni-app x开发商城系统,Swiper 轮播图
  • 昂瑞微OM6651A:国产车规级蓝牙芯片的破局者
  • 2025年中央空调/锅炉房/机房运维服务厂家最新权威推荐榜:专业托管与维修外包一体化解决方案精选
  • 【终极解决方案】api-ms-win-core-path-l1-1-0.dll 缺失?Win7/Win10/Win11完整修复教程
  • 2025 年最新推荐分切机实力厂家权威榜单:覆盖全自动高速、铝箔、薄膜、高精度等机型,为软包装企业精选优质设备
  • 打破应用跳转流失困局,提升推广链接转化率
  • 性能测试进阶秘籍:如何用JMeter分布式压测挖掘系统极限潜
  • Codeforces Round 1058 (Div. 2) A~E
  • 2025 年生料带厂家最新推荐排行榜:解析优质品牌优势,涵盖新型、彩色、液态等多类型生料带厂家企业推荐
  • openresty开发lua-resty-openssl之对称加密解密 - liuxm
  • 哈希乱搞:CF1418G Three Occurrences
  • 2025 年废旧轮胎裂解加热生产厂家最新推荐榜单:优质企业专利技术、产能规模与口碑实力全景解析锂化工焚烧炉/氟化热风系统/煤化工热风炉厂家推荐
  • 悲伤 自卑 乖戾 独自哭泣 陪伴空虚 kill my memory 让我将痛苦全忘记
  • 日志 | 2025.10
  • 工程师的 “指尖实验室”!正点原子 LT1 电桥镊子深度测评:同价位竞品谁能打?
  • 【ACM出版|EI检索稳定】2025年AI驱动下:业务转型和数据科学创新国际学术会议(ICBTDS 2025)
  • 破解跨域监控难题:国标GB28181算法算力平台EasyGBS视频调阅技术在跨域安防监控中的核心应用
  • 2025 年电缆桥架源头厂家最新推荐排行榜:聚焦优质供应商核心竞争力,助力工程采购精准选型
  • 2025 年厂房出售公司服务推荐排行榜:珠三角/广州/深圳/东莞/佛山/珠海等城市优质厂房出售公司全面测评解析
  • 构建智能视觉中枢:国标GB28181算法算力平台EasyGBS的全域感知与播放方案
  • 别再乱排查了!Kafka 消息积压、重复、丢失,根源基本都是 Rebalance!
  • 2025年交通杯-爆破题wp
  • 挖象浏览器下载安装教程|支持淘宝、拼多多、抖音多平台账号分区管理
  • 2025 年国内活性炭回收交易公司最新推荐排行榜:实力厂商深度解析,助力企业精准选合作方回收果壳活性炭/回收煤质柱状活性炭/库存各种活性炭公司推荐
  • 2025-10-15 CSP-J 模拟赛 赛后总结【ZROI】
  • 辐射检测仪哪家好?CT剂量模体哪家好?
  • 2025 木饰面源头厂家最新推荐榜单:21 年深耕企业领衔,背景墙 / 全屋 / 碳晶板 / 岩板全场景适配品牌解析