Training accuracy,Validation accuracy, Test accuracy

I’m new to deep learning and image processing. I don’t really understand Training accuracy, Validation accuracy, Test accuracy properly. I really need help with this talbe.


Thank you!

  • training accuracy - the accuracy achieved by the model on training data. It’s usually the highest one, because that’s what the model has seen and been able to learn.
  • valid/validation accuracy - accuracy achieved by the model on validation data. It’s usually a subset of the data that doesn’t participate in the training, but is used during it to verify how the model behaves.
  • test accuracy - accuracy achieved by the model on test data. This is usually checked after the actual training has finished.

The confusion might arise in case of validation and test accuracy, because they both don’t participate in the training, and are used only to check how the model generalized to never seen data. The difference is that validation dataset is checked during training, while test set is left untouched until the training has finished.

There’s also a case where test set is a completely different dataset, while validation set is usually created from part of the available training data.
This causes it to have sometimes higher accuracy than test set, because it basically uses the same data that is used for training, while test set might be something slightly different.

1 Like

Oops! I reanswered this question. :sweat_smile:

1 Like