Convolution with even-sized kernels and symmetric padding
paper
Shift Problem:
Asymmetric padding vs symmetric padding:
Training loss curve: