Adapter Override Warning During Trainer.Predict() #671
-
|
I am noticing whenever I call "trainer.predict" after training an adapter model (I am using BART for seq2seq), I get the following warning:
Is this safe to ignore? All I am doing is running Seq2SeqAdapterTrainer.train(), then immediately calling trainer.predict() |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 2 replies
-
|
Depending on your exact training setup, this can be expected. E.g., if you use |
Beta Was this translation helpful? Give feedback.
-
|
Yes that is exactly what I am doing! So this is correct then and the best adapter is whats being loaded? |
Beta Was this translation helpful? Give feedback.
-
|
Yes |
Beta Was this translation helpful? Give feedback.
Depending on your exact training setup, this can be expected. E.g., if you use
load_best_model_at_end=Trueto re-load the best checkpoint after training has completed, the existing adapter will be overwritten.