DCN V2 #28
Description
Recently MSRA published their paper about DCN v2 and released their code about v2 op, would you like to change it to tf version? What's more, when I init the DCN deeplab model with parameter learned with origin deeplab model, the result is different which is abnormal due to the initized 0 offset values. I only replace the conv2d with DCN, and I wonder if the implementation of dcn op is different from the origin in mxnet.