Yolov5: Yolov5l and Yolov5x results on custom dataset results

[jaqub-manuel]

❔Question

Hello Dear @glenn-jocher

I have 12000 images with single class. I have got best results with yolov5l via transfer learning and multi-scale training.

I wonder why I get lower results with yolo5x. Normally I should get higher results, right?
Also, I train the network first with transfer learning, multi-scale and default values. Then I fine-tune the model with coco finetune.yaml and the results are getting better. I evolve the model with the best weight I got. Do you think this order makes sense?
Also, python train.py --img 640 --batch 16 --epochs 3 --data coco128.yaml --cfg yolov5s.yaml --weights yolov5s.pt is transfer learning or strach training ? I think it is strach training because training takes longer than transfer training.

Thanks in advance for your help

Additional context

[glenn-jocher]

@jaqub-manuel there seem to be various questions merged togethor in your issue. I can't comment on your workflow other than to say that smaller models will overfit faster than larger models, so if you think you have a better result from a smaller model this is a likely indicator that you have not trained the larger model long enough.

[glenn-jocher]

@jaqub-manuel also, python train.py --weights yolov5s.pt by definition starts from the weights in yolov5s.pt. Training from scratch is done by python train.py --weights ''

You should start from the custom training tutorial which explains all of this.
https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data

[jaqub-manuel]

@glenn-jocher, I have trained yolov5l with 1000 epochs, and yolov5x with 2000 epochs. I am still getting best result with yolo5l.
In yolov5x, generally after 400 epochs, preciison increases, but recall decreases. What should be wrong ?

[jaqub-manuel]

@jaqub-manuel also, python train.py --weights yolov5s.pt by definition starts from the weights in yolov5s.pt. Training from scratch is done by python train.py --weights ''

You should start from the custom training tutorial which explains all of this.
https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data

python train.py --img 640 --batch 16 --epochs 3 --data coco128.yaml --cfg yolov5s.yaml --weights yolov5s.pt (In custom dataset trainig), I did not see any changes adding or removing --cfg file when use --weights yolov5s or m,l,x.. I have tried many times.

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[kazaros91]

@glenn-jocher, I have trained yolov5l with 1000 epochs, and yolov5x with 2000 epochs. I am still getting best result with yolo5l.
In yolov5x, generally after 400 epochs, preciison increases, but recall decreases. What should be wrong ?

This is probably because both False Positives and True Positives decrease, but False Positives decrease faster.