The confidence loss (obj_loss) scales of the training and validation sets are inconsistent.

[ayitime]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

Hi, recently I discovered an issue when using the YOLOv5 S model to train my dataset. Although the confidence loss (obj_loss) of the model has decreased in both the training and validation sets, the validation set loss is approximately twice that of the training set loss. However, the bounding box regression loss (box_loss) does not have such an issue. I carefully checked the code and confirmed that the calculation of the loss was correct. May I ask what is the reason for the different scales? Looking forward to your answer.

Additional

No response

[glenn-jocher]

@ayitime hi there! 😊

Thanks for bringing up your observation regarding the inconsistency in the confidence loss scales between training and validation. This kind of difference generally suggests that the model might be overfitting to the training data, especially if the validation loss remains significantly higher throughout the training process.

Here’s a quick suggestion:

1. Regularization: Try implementing or adjusting forms of regularization like data augmentation or dropout to see if these help generalize better.
2. Hyperparameter tuning: Particularly, you may consider tweaking the learning rate or the optimizer settings.
3. Balance the dataset: Ensure both your training and validation datasets are similarly representative of the problem space.

If you have already tuned these aspects and the issue persists, it would be helpful to experiment with a different split of the data or potentially even reassess whether the validation data is annotated consistently with the training data.

Let us know how it goes!