Created by: lcy-seso

the infer interface in V2 API has some problems which are VERY UNREASONABLE in use:

• One problem as reported in this issue: https://github.com/PaddlePaddle/Paddle/issues/2170.

  • the infer interface always flattens the prediction results given by one or multiple layer (if multiple layers are specified as outputs of the network) by concatenating them: https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/inference.py#L86
  • Such a concatenation will fail if multiple output layers have a different height which is a very common situation, for example, If I want to output both a sequence layer and a non-sequence layer. The former one has a variable length, then the concatenation will fail.

• by using the infer interface, users have to choose a field to output, the available choices are value, field, ids.

  • this line https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/inference.py#L70 iterate over the specified fields for all the chosen output layers.
  • for one layer user may want multiple fields as outputs, such as in NMT.
  • but if multiple layers are specified as outputs of the network, a user may want different fields for different layers.
  • This will cause a bug in the infer interface because the current implementation of infer interface constraints different output layers must output the same fields.

@reyoung I will fix this because I find I cannot bypass it, but I just want you to know the problem!