On bn_final in cnn_learner/create_cnn

Hi all,

in function cnn_learner/create_cnn, there is an argument bn_final, with default False. What is the effect for normal image classification tasks if we set it to True?

Anyone has any idea?

It creates a latch BatchNorm layer for the output. We have found it sometimes help, so don’t hesitate to try it for yourself!


Did you mean add a BatchNorm layer at the last?

Yes, just before the output.