Hmm…in fact, like TF / TFLite / Keras / CoreML’s frontend asymmetric padding should be changed now (TF frontend is changed now), We could do attr['padding'] = (pad_t, pad_l, pad_d, pad_r), no nn.pad operator before convolution. This work only need to do one time in fact and only for convolution, we won’t touch it anymore in the future. So, I don’t think it is a problem for existing frontend. Again, when we have one unified style of padding, new frontend will follow this code style. If we have 2+ code styles, new contributors doesn’t know which one is better, they will choose one. We shouldn’t try to push everything into Relay pass / low level mechanism. We should do it in more suitable place.