Is the any reason for not having a name argument in the Relay ops? This was present in NNVM.
Absence of this makes debugging extremely hard. It also makes it difficult to connect TVM to frameworks that transfer the information through tensor names.