what-if: upgrade what-if source to be tf 2.0 compatible

[stephanwlee]

The utility was using TF v1 only API.

Where necessary, we now use tf.compat.v1 instead of tf.
Also ran the migration script to add keywords to the arguments to some methods.

Report from the migration script:

INFO line 673:23: Added keywords to args of function 'tf.shape'
INFO line 713:17: Added keywords to args of function 'tf.cond'
INFO line 718:17: Added keywords to args of function 'tf.while_loop'
INFO line 749:8: Added keywords to args of function 'tf.parse_example'                              
INFO line 749:8: Renamed 'tf.parse_example' to 'tf.io.parse_example'                                                                                                                                    
TensorFlow 2.0 Upgrade Script       
-----------------------------
Converted 1 files             
Detected 0 issues that require attention                    
--------------------------------------------------------------------------------

Edit: ran the migration script to the interactive_inference source tree and it made minor changes to notebook/base.py.

Fixes #2330.

WANT_LGTM=nfelt

[nfelt]

I don't think the keyword arguments are actually necessary here since we're only using args that have stayed in the same position (but the migration script would automatically change them to kwargs since other arguments have moved).

Same goes for tf.while_loop below and tf.shape(input=...) above, and I think also for tf.io.parse_example() although passing keyword args there seems fair since the argument semantics are pretty non-obvious, but for these ones I think the argument semantics are clear (and more idiomatic) without names.

[stephanwlee]

The changes were made by the migration script. I'd revert the changes back but I want to know why TF migration script makes un-idiomatic changes. Is it that bad?

[nfelt]

The migration script migrates the arguments of any function with a changed signature to be keyword arguments, even if the function isn't called with arguments that were affected by the change. I think it's just simpler that way since then you don't have to do as much analysis of the callsite and also the signature changes (which could be argument renames or reorders) - each migration can be done as a self-contained step rather than worrying about preserving the details of which args were called with names or without names.

I agree it would be nicer if it did produce idiomatic TF 2.0 code, but this is part of why manual review of the changes is recommended anyway - the goal is for the code it produces to work but it won't necessarily be idiomatic.

[nfelt]

The migration script migrates the arguments of any function with a changed signature to be keyword arguments, even if the function isn't called with arguments that were affected by the change. I think it's just simpler that way since then you don't have to do as much analysis of the callsite and also the signature changes (which could be argument renames or reorders) - each migration can be done as a self-contained step rather than worrying about preserving the details of which args were called with names or without names.

I agree it would be nicer if it did produce idiomatic TF 2.0 code, but this is part of why manual review of the changes is recommended anyway - the goal is for the code it produces to work but it won't necessarily be idiomatic.