Proper way to save RNN Models

The BucketingModule class doesn’t seem to have a very clean way to save both the symbol and the parameters (only the parameters are well supported). Whats the solution for this? Does the model need to be pickled separately?

Yeah I think you need to generate the symbols on the fly.
@zheng-da is working on control flow op, hopefully this will support checkpointing dynamic graph and no need to use bucketing module anymore?