We have discussed many things about extending conv2d padding to 4 dims padding. (top, bottom, left, right) like this issue: https://github.com/dmlc/nnvm/pull/472
We have done this support completely and run models very well (TF MobilenetV1 / Mobilenet V2 / Deeplab V3 / Resnet50 and so on). We want to contribute back to community. However, this affects many places, we want to make sure community want to accept this change or not. Then we can pull request it.