Created by: lcy-seso

The current InferShape interface cannot obtain the LoD information(the sequence information). This makes the output dimension of the CRF operator cannot be correctly set in the InferShape function.

Because the CRF operator takes sequences as its inputs and outputs a non-sequence as the result. How many sequences are there in the input is required (the LoD information) to determine the output dimension.

I have to set an arbitrarily wrong dimension to the output in InferShape and then resize the right dimension in Compute. This seems very tricky.

Have we already determined that the LoD will not pass to InferShape in operator's kernel?

@jacquesqiao @reyoung