Validate the test

[Maiko7]

I obtained best.pt by using python start_train.py --model ./ultralytics/cfg/models/v8/yolov8m_ECA.yaml --data_dir ./GRAZPEDWRI-DX/data/meta.yaml. Now, what code should I use to train my test set with best.pt?

[RuiyangJu]

Hello,
''start_train.py'' contains the training and validation.

For example, you will get:

Optimizer stripped from runs\detect\train\weights\last.pt, 22.9MB
Optimizer stripped from runs\detect\train\weights\best.pt, 22.9MB

Validating runs\detect\train\weights\best.pt...

[Maiko7]

Hello, thank you for your response. I would like to know how I can use code to test my test dataset with best.pt obtained from start_train.py. How should I proceed with the test dataset? Should I use test? Or how can I use best.pt to validate the test dataset?

[ls-ale]

您好，为什么我训练的map值达不到这么高呢，差了十几个点

[Arch1o]

您的yaml文件中设置的neck部分连接方式与论文图片不相符，论文中图片显示中间两个注意力机制位置都在检测头仅作为输出，而yaml文件中设置的却连接了后面的conv模块。

[RuiyangJu]

@ls-ale
你和我所使用的數據集是不一樣的，所以實驗結果會不一樣。

[RuiyangJu]

@Arch1o
具体你可以看Figure 3.

[Arch1o]

感谢回复，如果按照您论文中注意力机制的位置，yaml文件中第15和20层Conv模块的输入层是否应该填-2而不是-1呢？

• [-2, 1, Conv, [256, 3, 2]] # 16

• [[-1, 12], 1, Concat, [1]] # cat head P4

• [-1, 3, C2f, [512]] # 18 (P4/16-medium)

• [-1, 1, ResBlock_CBAM, [512]] # 19

• [-2, 1, Conv, [512, 3, 2]]

• [[-1, 9], 1, Concat, [1]] # cat head P5

• [-1, 3, C2f, [1024]] # 22 (P5/32-large)

• [-1, 1, ResBlock_CBAM, [1024]] # 23

[Dipan112233]

感谢回复，如果按照您论文中注意力机制的位置，yaml文件中第15和20层Conv模块的输入层是否应该填-2而不是-1呢？

• [-2, 1, Conv, [256, 3, 2]] # 16
• [[-1, 12], 1, Concat, [1]] # cat head P4
• [-1, 3, C2f, [512]] # 18 (P4/16-medium)
• [-1, 1, ResBlock_CBAM, [512]] # 19
• [-2, 1, Conv, [512, 3, 2]]
• [[-1, 9], 1, Concat, [1]] # cat head P5
• [-1, 3, C2f, [1024]] # 22 (P5/32-large)
• [-1, 1, ResBlock_CBAM, [1024]] # 23

我感觉你说的对，对于论文中出现的Attention Module，第2个出现在c2f后面，然后就是卷积，但图中卷积也是在c2f后面

[Dipan112233]

@Arch1o 具体你可以看Figure 3.

但朋友你的yaml文件确实如上面所说，yolov8_ECA中确实是第二个注意力块后面跟的卷积，而不是卷积和注意力层数相同

[Salmankm93]

Hello;
I appreciate your efforts in this astonishing repo which truly helped me well. However, I want to validate the custom dataset on this repo, so I have changed the default.yaml file, so I have changed the task to val instead of detect what I have to do next is it acceptable to code like this :
!python /content/Fracture_Detection_Improved_YOLOv8/start_train.py --model /content/Fracture_Detection_Improved_YOLOv8/runs/detect/train/weights/best.pt --data_dir /content/PCB-Dataset-Defect-1/data.yaml
Is this enough and accepted?
I appreciate any help you can provide.