Implementing TensorFlow 2.0 Compatibility

[juntai-zheng]

What changes are proposed in this pull request?

In order for MLflow to be compatible with the upcoming TensorFlow 2.0, several under the hood changes had to be made in the examples, MLflow code, and tests. The main difference is how we load back models - instead of loading back signature definitions in conjunction with tf.graphs and sessions, TF 2.0 allows us to abstract those away and load back models as callable functions for inference.

The main API change is in mlflow/tensorflow.py with the function load_model(model_uri, tf_sess). tf_sess is now optional, since TF 2.0 deprecated sessions. The code will error out if the user's TF version is < 2.0 and no argument is provided. The code will throw a warning that tf_sess is deprecated and ignore the input if the user's TF version is 2.0+.
In addition, load_model() originally returned a SignatureDef. Now, the function returns a function which can be fed data and returns inputs. See the docs for details.

The code has been modified to check for the user's TensorFlow version and be compatible with both TF 1.X and TF 2.0 (given correct arguments to APIs).

New files have been created for TF 2.0 tests and examples, with the underlying example dataset being changed to TensorFlow's iris_data example. A couple of tests testing loading a model into a tf.graph have been removed due to TF 2.0 abstracting graphs away. The Travis bash file has been modified to install the proper TF versions for each suite of tests.

StrictVersion has been changed to LooseVersion for version checks in order to allow for letters in version numbers (such as TensorFlow 2.0.0rc2) to be valid version numbers.

This PR does not address autologging, keras, or keras autologging - those will be in a future PR.

How is this patch tested?

I created a new suite of tests (mlflow/tests/tensorflow/test_tensorflow2_model_export.py), based on the original tests in order to test the changes. The changes to mlflow/tensorflow.py work with both the original test on TF 1.12 and with the new tests on TF 2.0.0rc2.

Release Notes

Is this a user-facing change?

• No. You can skip the rest of this section.
• Yes. Give a description of this change to be included in the release notes for MLflow users.

MLflow now supports TensorFlow 2.0. load_model() no longer takes a tf_sess, but still does if you are using TensorFlow 1.X. The function also returns a callable graph in the form of a tf.function rather than a SignatureDef now.

What component(s) does this PR affect?

• UI
• CLI
• API
• REST-API
• Examples
• Docs
• Tracking
• Projects
• Artifacts
• Models
• Scoring
• Serving
• R
• Java
• Python

How should the PR be classified in the release notes? Choose one:

• rn/breaking-change - The PR will be mentioned in the "Breaking Changes" section
• rn/none - No description will be included. The PR will be mentioned only by the PR number in the "Small Bugfixes and Documentation Updates" section
• rn/feature - A new user-facing feature worth mentioning in the release notes
• rn/bug-fix - A user-facing bug fix worth mentioning in the release notes
• rn/documentation - A user-facing documentation change worth mentioning in the release notes

[smurching]

BTW one comment I meant to add earlier but forgot...we ideally should have a test where we load a MLflow model persisted with 1.x using TF 2.0 (both in function form & as a pyfunc). I think it's fine if we add such a test in a follow-up PR, but seems important to have before we claim TF 2.0 compatibility, and we should make sure the follow-up work is tracked somewhere.

[smurching]

BTW, super small nit but in general it's best to avoid these sort of no-op whitespace changes if possible (did your editor make them or something?). Not a big deal, it just makes the git blame harder to trace through

[smurching]

Some small comments but otherwise looks good! Before we merge, we should manually test backwards compatibility when loading models persisted with TF 1.x using TF 2.0

[smurching]

LGTM thanks @juntai-zheng!