I agree, if you want to maintain fidelity with the paper; later papers of course play with this and change it around.

"We adopt batch normalization (BN) [16] right after each convolution and

before activation, following [16]."