WongKinYiu/YOLO

Training on a custom dataset

johannes-tum opened this issue · 2 comments

johannes-tum commented

Issue Description
I loaded the COCO pretrained weights and tried to train on a custom dataset.

  • I got overfitting to work on the custom dataset, but full training does not work yet. For some reason, BCELoss is extremely high after a few (maybe 30 iterations). After that, it is in the range of roughly 50-250.
  • Maybe for others, it would also be great to have a guide how to modify the code for a custom dataset.
  • Finally, I would greatly appreciate if you could conduct an experiment where you retrain the model for YOLO and show reproduced performances.
pavisworkshop commented

Hi @johannes-tum.

Could I ask how you organised your data and the type of annotation format that you used ? I never quite get the expected behaviour when trying to train on my data (I also tried on an old dataset organised for yolov5).

yjmm10 commented

Hi @johannes-tum ，

I try to train a model with a custom dataset, but failed in the model validation phase. With debugging, I find the result of the model is questionable, because after NMS , the predicted box is none. It means that all results are have been filtered. Maybe the NMS also have questions.

If you have new findings, welcome to talk with me.