Intermediate checkpoints outperforming final checkpoint? about dinov2 HOT 3 OPEN

how bad ? can you plot the loss and eval curves ?

from dinov2.

The loss curves seem relatively normal, continuing to decrease throughout training, but here are plots of the results of 16 evaluation tasks. As you can see, very often the models at 50% do better than the models at 100%. These are ViT-S models. For some of the tasks (eg, the bottom middle two), the curves are extremely irregular, and so these might not be great tasks, but the aforementioned behavior seems to be present in many of the others.

from dinov2.

I had the same experience, not too surprised. we have plots showing similar behaviour here: https://arxiv.org/abs/2401.04720

from dinov2.