GitLab

Hi @mbertoni!

When you want to predict using the AdanetWrapper object is rather "bothering" having to provide always a TrainTest path object (which won't be used at all for the prediction).

What about making the "TrainTest" argument optional? (obviously, if not provided for training the model I expect the AdanetWrapper to complain). Or maybe making a new Adanet object for doing predictions (that only takes as argument the "Adanet's savemodel path")?