Under the same validation set, the evaluation results of the EMA model after loading are inconsistent with those during training.

[FriedaSmith]

Hello, I use EVA-02/asuka/run_class_finetuning.py to fine-tune an image recognition task. After training the model, I evaluated it by loading the trained model. When loading the mp_rank_00_model_states.pt under the checkpoint-best, the results on the validation set (the validation set is the same as the validation set during training) are The evaluation results during training are the same. However, when I load the model under the checkpoint-best-ema, the accuracy obtained is much smaller than the evaluation results during training using EMA. What is the problem?

[FriedaSmith]

The problem has been resolved.