Hi Priyanka,
unfortunately I don't know of a good, publicly available way to create a metagraph file in the same way. Basically, what we do is:
– Generate a TF graph that includes both encoder and decoder.
– Replace the TF variables with constants, then do constant folding.
– Save the resulting graph together with the function signatures for encoder and decoder (basically, which tensors are inputs and outputs for each) as a .metagraph.
The format for metagraph files are protobufs, so they are publicly accessible, but the conversion tools aren't, and they wouldn't work with the public version of TFC anyway. When we train models, we use a different codebase that only uses some parts of TFC (i.e. we don't directly use the code under models/).
Hope this helps…
Johannes