Assume that I have got a mini-batch of inputs with labels in {1,2, …,N}. I need the Label-N inputs have no impact on the update of BN layer parameters. Is there an easy way to reach my goal? Thanks a lot.
Btw, I tried an approach, which i finally find wrong. I pick out the Label-N inputs, set the model under evaluation mode by “model.eval()”, and let the Label-N inputs go forward through the model; Then I set the model back to training mode, input the rest data to generate the loss. But in the following back propagation stage, in which the parameters get updated, I guess the BN layer still learns from all the inputs. Is my guess correct?