I have a little question about the BatchNorm layer in the model.
In ./colorization/models/colorization_train_val_v2.prototxt, I noticed that you are using BatchNorm layers without Scale layers following behind, so does this mean that you are not using the original definition of BatchNorm that has additional scaling factors and biases?
If my understanding is correct, have you tried BN layers with scaling factors and biases? I am wondering to what degree will this difference affects the training results.
Hi,
I have a little question about the BatchNorm layer in the model.
In ./colorization/models/colorization_train_val_v2.prototxt, I noticed that you are using
BatchNorm
layers withoutScale
layers following behind, so does this mean that you are not using the original definition of BatchNorm that has additional scaling factors and biases?If my understanding is correct, have you tried BN layers with scaling factors and biases? I am wondering to what degree will this difference affects the training results.