What is the model-best output based on? #8839
-
|
How is the output in model-best folder determined? What makes it best compared to the model-last output from the final training iteration. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 2 replies
-
|
When training, there's period output as the trainer runs an evaluation step on dev data. That output includes a "score", which is used to select the best model. |
Beta Was this translation helpful? Give feedback.
-
|
in order to create my final model I use all my data (train + dev) which is evaluated on a final test set. based on what you are saying, should I swap my dev set for my final test set when I'm done hyperparameter tuning and validation phase (i.e. when ready to train the final model for production use)? |
Beta Was this translation helpful? Give feedback.
-
|
Yes, that would be better. If your train and dev set are the same you can overfit without realizing it. |
Beta Was this translation helpful? Give feedback.
When training, there's period output as the trainer runs an evaluation step on dev data. That output includes a "score", which is used to select the best model.