using predictor.model/trainer.model in inference_on_dataset

[andreaceruti]

What is the right way of evaluating my model?

If I use predictor.model I get lower results than using trainer.model, could you explain me what is the reason?

In my case I train a MaskRCNN model saving with a hook the best weights according to the validation loss (I have implemented validation loss + early stopping). At the end of the train I create another cfg and I store there the weights of the best model found before. Then I run the evaluation on my test set but the problem is that evaluation dove with predictor.model is lower than the train.model one.
So what could be the reason, maybe it does not mean that if I have a model with the smallest validation loss I should also have the model with the better evaluation?