The metrics of my trainings get worse over time but training does not stop. If I understand correctly, the final model that is picked by Roboflow is worse than one of the earlier checkpoints.
As you can see the only loss that improves is the class loss, which is useless here because I only use a single class.
Unfortunately the entire dataset is relatively small, so the validation and test set are very small (around 200 images after augmentation).
Project: Sign in to Roboflow