the accuracy in Train logs isn't the same as custom test.py

[Kegard]

Hello!
When I train the T5 large, I have encounter the question. In the train logs, the accuracy of epoch 19 is 58.1. However, when I test the accuracy use the custom_test.py, the accuracy is 38.90. The code of custom_test.py I modified as follow:

    ckpt_path = get_weight(model = model_key,data=dataset_key) # the path of the checkpoint of ckpt
    trainer.validate(lm, datamodule=data_module,ckpt_path=ckpt_path)

I want to know which score should I use? and why the score is different?

This is the score when I use the custom_test.py

And this is the score in the logs when I train the model

Thank you for your reply!