Intellindust-AI-Lab
Why is my dataset converging so slowly
Why is my dataset converging so slowly, even slower than the YOLO model? After running for 300 epochs, DEIM-fines has an AP of 74.4, while YOLOv5s has an AP of 87.1.
Roughly how big is your dataset? My observation is that yolo tends to perform better on smaller dataset and detr based model is very good when the problem is complex and dataset is large. Just a personal anecdote.
Hello, thank you very much for your answer! The training and validation sets together contain a total of 9,460 images.
I see, any chance that there is a similar dataset like yours that I can try out? Which model are you training on?
Do you use Baidu Netdisk? I have uploaded my dataset to Baidu Netdisk, and the link is provided below. If you do not have access, please let me know promptly. Thank you!
通过网盘分享的文件：cup_coco.zip
链接: https://pan.baidu.com/s/11b08UR78CZKvTDFV_iqhyg?pwd=r5ks 提取码: r5ks
Dear dnth，here is the link to my Google Drive. I hope you can help me understand why the convergence is so slow. If you are able to improve the convergence speed by adjusting the hyperparameters, I would be greatly appreciative.
https://drive.google.com/file/d/1oA26W_KQbdJwfigSV2kEuZDZr92lJn-j/
Hey @xiarencunzhang I don't use Baidu Netdisk. Is there alternative?
@xiarencunzhang would you mind using this wrapper I made on your dataset?
It's a fork of DEIM and I added a few additional features to make training easier https://github.com/dnth/DEIMKit
I also noticed you don't have a background class in your dataset. Not sure if that is a problem. All dataset i train on so far has a background class with id 0.
Hello, @dnth .Thank you for your response. I believe that the absence of a background class in the dataset should not significantly impact the convergence of the training. On the other hand, I am also quite interested in your project. I have looked at your work and can sense that your wrapper makes DEIM and inference more convenient. If our goal is to improve DEIM, I am not sure if your project would be helpful.