V8训练NEU-DET
Stru99le Lv2
  2024-06-05 16:41:43 2024-06-05 16:41   2024-08-15 21:48:25      2k 字  10 分钟  

第一种划分训练，1620训练180验证(9:1)

原版V8(72.2)

  • 条件

    epcoh =100,batch =-1,model=yolov8n

  • 结果

    image

  • PR_recall

    image

  • 条件2

    epoch = 300，其余一样

  • 结果

    image 
    1
    2
    metrics/precision(B)	      metrics/recall(B)	       metrics/mAP50(B)	    metrics/mAP50-95(B)
       0.73849,                0.65913,                0.72279,                0.39697,
    image

  • PR_recall

    image

    可以看到当训练轮次增多的时候map反而降低了，验证损失波动比较大，合适的eopoch应该在100-150左右，

加入SwinTransformer(76.4)

  • 条件

    epoch = 265,batch = -1,model=yolov8n

  • 结果

    image 
    1
    2
    metrics/precision(B)	      metrics/recall(B)	       metrics/mAP50(B)	    metrics/mAP50-95(B)
             0.68036	             0.72664	             0.75465	             0.39744
    image

  • PR_recall

    image

  • Confusion Matrix Normalized

    image

V9(和v8怎么一样？)

  • 条件

    epcoh=100,batch=-1,model=yolov9-c

  • 结果

    image image

    不知道为什么loss都是0？

  • PR_recall

    image

    map和v8一模一样？，不过v9模型比v8大，训练时间也更长

V10（和V8V9没什么变化）

  • 条件

    epoch=100,batch=-1,model=yolov10n

  • 结果

    image image

  • PR_Recall

    image

V8加入BRA注意力(74.4)

  • 配置文件<加入到第八层>

    image

  • 条件

    epoch=100，batch=-1，model=yolov8n

  • 结果

    image image

  • PR_Recall

    image

  • 疑问

    注意力主要是加在了backbone之中，换下位置效果会不会不一样呢？

  • 计算量

    加入BRA后计算量增大了非常多，可以看到是18.5 GFLOPS

BRA注意力换位置到第五层(降点)

  • 配置文件

    image

  • 结果

    image image

BRA注意力换到11层上采样后面(降点)

  • 配置文件

    image

  • 结果

    image image image

加入三层BRA注意力(降点)

  • 配置文件

    image

  • 结果

    image image image

  • 结论

    加入三层BRA后计算量达到了21GFLOPS,准确率依旧是降点

RT-DETR(降点，过拟合)

  • 条件

    epoch=100，batch=32，model=redetr-l

  • 结果

    image image image

  • 结论

    从结果图上看很明显出现了过拟合现象，个人觉得是数据集太小了

加入CAFM注意力(72.2)

  • 件

    和上面一致

  • 结果

    image image

第二种划分数据集训练，1297训练325验证

原版v8(74.8)

  • 条件

    和之前v8一致

  • 结果

    image image image

加入Bra注意力(73.8) +shapeiou(74.8)

  • 条件

    和之前一致

  • 结果

    image image image

加入CAFM注意力(73.5)使用预训练权重(75.3)

  • 条件

    一致

  • 结果

  • image image image

  • 加入预训练权重

    image

8:2比例划分数据集训练

原版V8，不加预训练权重(74.5)200epoch(76)

  • 条件

    一致，4060ti

  • 结果

    image image image

  • 206epoch

    image image

YOLOV10(70.8)↓

BRA注意力+ShaopeIoU(75.7)

  • 条件

    200epoch，4090ti

  • 结果

    image image image

CAA注意力(76.3)

  • 条件

    217epoch

  • 结果

    image image

FastEMA(76.6)

  • 条件

    250epoch

  • 结果

    image

CAA(76.3)

  • 条件

    200epoch

  • 结果

    image

MDCR(75.4)

  • 200epoch

    image

OSRAAtention(75.9)

  • 200epoch

  • 结果

    image

HybridTokenMixer(77.2)

  • 200epoch

  • 结果

    image image

PKIBlock(77.7)

  • 200epoch 还未收敛，300epoch降到75.5？？

  • 结果

    image image

Shift-ConvNets (76.7)

  • 200epoch

  • 结果

    image

DAF_CA(77)

  • 200epoch

  • 结果

    image image

MODSConv(76.5)

  • 200epoch

  • 结果

    image

MODSLayer(76.8)

  • 200epoch

  • 结果

    image

    详细结果在服务器exp13

DualConv(76.5)

  • 200epoch

  • 结果

    image

    详细结果在服务器exp14

MSPA(77.2)

  • 200epoch

  • 结果

    image

    详细结果在服务器epx15

msblock(73.9)↓

c2f_msblock(71.8)

SPPF_improve()↓

SPPF_UniRepLlk(77.5)↑

  • 200epoch和250epoch都是同样的效果
image image

CPMS(77.5)↑

  • 结果

    image

替换主干MobileNetV3+NWD_Loss(74.9)

  • 结果

    image

BRA注意力+PConv(76.8)↓

bra注意力来自BiFormer,动态稀疏注意力，PConv来自FasetNet旨在提高神经网络的计算速度和效率。

  • 结果

    image

替换PIOU损失函数后计算量18.2(77.8)↑

  • 结果

    image

加入3层PConv(76.3)↓

BRA+ODConv+PConv(74.7)↓

BRA+C2f_ODConv(77.2)↑

  • 结果

    image

V8加入BIFPN(75.1)↓

  • 尝试换了几种损失函数都是降点

V8加入GAM注意力backbone(77.3)

  • 结果

    image

加在neck(78)↑(采用nwd损失)77.7

  • 结果

    image image

在neck部分的所有c2f后面添加(76.3)

EMA注意力+PIoU2(75.2)↓

+WISEIOUV3(75.3)↓

我的组合

Starnet+Pkiblock(77.8)

  • 结果

    image image

MSPA+PKIBlock(77)↑

  • 运行条件

    200epoch(250epoch也是这么多)

  • 结果

    最好结果是将pkiblock放在第九层，尝试放在第十层，或者第八层都是不如第九层好，尝试使用原版ciou函数效果没有iou好

    image

MSPA+DAF_CA(77)

  • 200epoch

  • 结果

    不确定性算法警告，

    image

MSPA+slimneck+piou(76.6)

  • 200epoch
image

DWR_DRB+slimneck+piou(77.7)↑

  • 200 和 250epoch结果都是一样

  • 结果

    image

替换EIOU后(76.5)↓

  • 结果

    image

替换Focaler-Iou(76.8)↓

  • 结果

    image

替换NWD_loss<ratio=0.5>(77.3)

  • 结果

    image

rati0=0.7(77)

DWR_DRB+slimneck+SPPF_

UniRep(76.8)

  • 结果

    image

不加slimneck(77)

  • map77

c2f_DWR_DRB+CPMS+Slimneck(77)↑

  • 结果

    image image

DWR_DRB+slimneck+LADH(77.3)

  • 结果

    image

GAM+DWR_DWB+NWD(75.7)↓

GAM+DWR_DWB+PIOU2(78)↑

3.5M参数量

image image

GAM+Starnet;不替换neck(75.6)↓

GAM+Starnet;替换neck部分(78.1)↑

3.5M参数

image image

主干最后两层加starnet,neck全加(74.4)

替换WIOUv3损失(77.5)↓

image

替换neck部分所有C2f(77.7)

GAM+Slimneck+PIOU2(76.1)

  • 配置文件

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    22
    23
    24
    25
    26
    27
    28
    29
    30
    31
    32
    33
    34
    35
    36
    37
    38
    39
    40
    41
    42
    43
    44
    45
    # Ultralytics YOLO 🚀, AGPL-3.0 license
    # YOLOv8 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect
     
    # Parameters
    nc: 80  # number of classes
    scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'
      # [depth, width, max_channels]
      n: [0.33, 0.25, 1024]  # YOLOv8n summary: 225 layers,  3157200 parameters,  3157184 gradients,   8.9 GFLOPs
      s: [0.33, 0.50, 1024]  # YOLOv8s summary: 225 layers, 11166560 parameters, 11166544 gradients,  28.8 GFLOPs
      m: [0.67, 0.75, 768]   # YOLOv8m summary: 295 layers, 25902640 parameters, 25902624 gradients,  79.3 GFLOPs
      l: [1.00, 1.00, 512]   # YOLOv8l summary: 365 layers, 43691520 parameters, 43691504 gradients, 165.7 GFLOPs
      x: [1.00, 1.25, 512]   # YOLOv8x summary: 365 layers, 68229648 parameters, 68229632 gradients, 258.5 GFLOPs
     
    # YOLOv8.0n backbone
    backbone:
      # [from, repeats, module, args]
      - [-1, 1, Conv, [64, 3, 2]]  # 0-P1/2
      - [-1, 1, Conv, [128, 3, 2]]  # 1-P2/4
      - [-1, 3, C2f, [128, True]]
      - [-1, 1, Conv, [256, 3, 2]]  # 3-P3/8
      - [-1, 6, C2f, [256, True]]
      - [-1, 1, Conv, [512, 3, 2]]  # 5-P4/16
      - [-1, 6, C2f, [512, True]]
      - [-1, 1, Conv, [1024, 3, 2]]  # 7-P5/32
      - [-1, 3, C2f, [1024, True]]
      - [-1, 1, GAM_Attention, [1024]]  # 9
      - [-1, 1, SPPF, [1024, 5]]  # 10
    head:
      - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
      - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
      - [-1, 3, VoVGSCSPns, [512]]  # 13

      - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
      - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
      - [-1, 3, VoVGSCSPns, [256]]  # 16 (P3/8-small)

      - [-1, 1, GSConvns, [256, 3, 2]]
      - [[-1, 13], 1, Concat, [1]]  # cat head P4
      - [-1, 3, VoVGSCSPns, [512]]  # 19 (P4/16-medium)

      - [-1, 1, GSConvns, [512, 3, 2]]
      - [[-1, 10], 1, Concat, [1]]  # cat head P5
      - [-1, 3, VoVGSCSPns, [1024]]  # 22 (P5/32-large)

      - [[16, 19, 22], 1, Detect, [nc]]  # Detect(P3, P4, P5)

  • 结果 exp71

GAM+Starnet+Slimneck+PIOU2(75)

  • 配置文件

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    22
    23
    24
    25
    26
    27
    28
    29
    30
    31
    32
    33
    34
    35
    36
    37
    38
    39
    40
    41
    42
    43
    44
    45
    # Ultralytics YOLO 🚀, AGPL-3.0 license
    # YOLOv8 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect

    # Parameters
    nc: 80 # number of classes
    scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'
      # [depth, width, max_channels]
      n: [0.33, 0.25, 1024] # YOLOv8n summary: 225 layers,  3157200 parameters,  3157184 gradients,   8.9 GFLOPs
      s: [0.33, 0.50, 1024] # YOLOv8s summary: 225 layers, 11166560 parameters, 11166544 gradients,  28.8 GFLOPs
      m: [0.67, 0.75, 768] # YOLOv8m summary: 295 layers, 25902640 parameters, 25902624 gradients,  79.3 GFLOPs
      l: [1.00, 1.00, 512] # YOLOv8l summary: 365 layers, 43691520 parameters, 43691504 gradients, 165.7 GFLOPs
      x: [1.00, 1.25, 512] # YOLOv8x summary: 365 layers, 68229648 parameters, 68229632 gradients, 258.5 GFLOPs

    # YOLOv8.0n backbone
    backbone:
      # [from, repeats, module, args]
      - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
      - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
      - [-1, 3, C2f_StarNB, [128, True]]
      - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
      - [-1, 6, C2f_StarNB, [256, True]]
      - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
      - [-1, 6, C2f_StarNB, [512, True]]
      - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
      - [-1, 3, C2f_StarNB, [1024, True]]
      - [-1, 1, GAM_Attention, [1024]]  # 9
      - [-1, 1, SPPF, [1024, 5]] # 10
    head:
      - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
      - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
      - [-1, 3, VoVGSCSPns, [512]]  # 13

      - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
      - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
      - [-1, 3, VoVGSCSPns, [256]]  # 16 (P3/8-small)

      - [-1, 1, GSConvns, [256, 3, 2]]
      - [[-1, 13], 1, Concat, [1]]  # cat head P4
      - [-1, 3, VoVGSCSPns, [512]]  # 19 (P4/16-medium)

      - [-1, 1, GSConvns, [512, 3, 2]]
      - [[-1, 10], 1, Concat, [1]]  # cat head P5
      - [-1, 3, VoVGSCSPns, [1024]]  # 22 (P5/32-large)

      - [[16, 19, 22], 1, Detect, [nc]]  # Detect(P3, P4, P5)

  • 结果 exp72

复现WSS-YOLO()

  • 配置

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    22
    23
    24
    25
    26
    27
    28
    29
    30
    31
    # YOLOv8.0n backbone
    backbone:
      # [from, repeats, module, args]
      - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
      - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
      - [-1, 3, C2f, [128, True]]
      - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
      - [-1, 6, C2f, [256, True]]
      - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
      - [-1, 6, C2f_DySnakeConv, [512, True]]
      - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
      - [-1, 3, C2f_DySnakeConv, [1024, True]]
      - [-1, 1, SPPF, [1024, 5]] # 9
    head:
      - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
      - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
      - [-1, 3, VoVGSCSP, [512]]  # 12

      - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
      - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
      - [-1, 3, VoVGSCSP, [256]]  # 15 (P3/8-small)

      - [-1, 1, GSConv, [256, 3, 2]]
      - [[-1, 12], 1, Concat, [1]]  # cat head P4
      - [-1, 3, VoVGSCSP, [512]]  # 18 (P4/16-medium)

      - [-1, 1, GSConv, [512, 3, 2]]
      - [[-1, 9], 1, Concat, [1]]  # cat head P5
      - [-1, 3, VoVGSCSP, [1024]]  # 21 (P5/32-large)

      - [[15, 18, 21], 1, Detect, [nc]]  # Detect(P3, P4, P5)
  1. 第一种划分训练，1620训练180验证(9:1)
    1. 原版V8(72.2)
    2. 加入SwinTransformer(76.4)
    3. V9(和v8怎么一样？)
    4. V10（和V8V9没什么变化）
    5. V8加入BRA注意力(74.4)
    6. BRA注意力换位置到第五层(降点)
    7. BRA注意力换到11层上采样后面(降点)
    8. 加入三层BRA注意力(降点)
    9. RT-DETR(降点，过拟合)
    10. 加入CAFM注意力(72.2)
  2. 第二种划分数据集训练，1297训练325验证
    1. 原版v8(74.8)
    2. 加入Bra注意力(73.8) +shapeiou(74.8)
    3. 加入CAFM注意力(73.5)使用预训练权重(75.3)
  3. 8:2比例划分数据集训练
    1. 原版V8，不加预训练权重(74.5)200epoch(76)
    2. YOLOV10(70.8)↓
    3. BRA注意力+ShaopeIoU(75.7)
    4. CAA注意力(76.3)
    5. FastEMA(76.6)
    6. CAA(76.3)
    7. MDCR(75.4)
    8. OSRAAtention(75.9)
    9. HybridTokenMixer(77.2)
    10. PKIBlock(77.7)
    11. Shift-ConvNets (76.7)
    12. DAF_CA(77)
    13. MODSConv(76.5)
    14. MODSLayer(76.8)
    15. DualConv(76.5)
    16. MSPA(77.2)
    17. msblock(73.9)↓
    18. c2f_msblock(71.8)
    19. SPPF_improve()↓
    20. SPPF_UniRepLlk(77.5)↑
    21. CPMS(77.5)↑
    22. 替换主干MobileNetV3+NWD_Loss(74.9)
    23. BRA注意力+PConv(76.8)↓
      1. 替换PIOU损失函数后计算量18.2(77.8)↑
      2. 加入3层PConv(76.3)↓
    24. BRA+ODConv+PConv(74.7)↓
    25. BRA+C2f_ODConv(77.2)↑
    26. V8加入BIFPN(75.1)↓
    27. V8加入GAM注意力backbone(77.3)
      1. 加在neck(78)↑(采用nwd损失)77.7
      2. 在neck部分的所有c2f后面添加(76.3)
    28. EMA注意力+PIoU2(75.2)↓
      1. +WISEIOUV3(75.3)↓
  4. 我的组合
    1. Starnet+Pkiblock(77.8)
    2. MSPA+PKIBlock(77)↑
    3. MSPA+DAF_CA(77)
    4. MSPA+slimneck+piou(76.6)
    5. DWR_DRB+slimneck+piou(77.7)↑
      1. 替换EIOU后(76.5)↓
      2. 替换Focaler-Iou(76.8)↓
      3. 替换NWD_loss<ratio=0.5>(77.3)
      4. rati0=0.7(77)
    6. DWR_DRB+slimneck+SPPF_
    7. UniRep(76.8)
      1. 不加slimneck(77)
    8. c2f_DWR_DRB+CPMS+Slimneck(77)↑
    9. DWR_DRB+slimneck+LADH(77.3)
    10. GAM+DWR_DWB+NWD(75.7)↓
    11. GAM+DWR_DWB+PIOU2(78)↑
    12. GAM+Starnet;不替换neck(75.6)↓
      1. GAM+Starnet;替换neck部分(78.1)↑
        1. 主干最后两层加starnet,neck全加(74.4)
        2. 替换WIOUv3损失(77.5)↓
      2. 替换neck部分所有C2f(77.7)
    13. GAM+Slimneck+PIOU2(76.1)
    14. GAM+Starnet+Slimneck+PIOU2(75)
    15. 复现WSS-YOLO()