Is drop_path_rate decativated in .eval() mode?

[ahmed1996said]

I'm using efficnetNet_b4 to pretrain my dataset with drop_path_rate=0.2 . While evaluating the model (ex. for testing), does path drop gets deactivated? If not, how can I manually deactivate it?

[TorbenSDJohansen]

Dear ahmed1996said,

I believe it should be disabled so long as the Module is not in training mode. For example, see the code for drop_path. From this, note how the input is returned unaltered provided training is False, which is the case in eval mode.

Torben

[ahmed1996said]

Great, thanks for your help!