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.
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.