Clearer interaction of TensorFlowModel with new framework versions

[athewsey]

Is your feature request related to a problem? Please describe.
We can create a sagemaker.tensorflow.model.TensorFlowModel for new versions of TensorFlow (e.g. 2.0, 2.1), but get a "container not found" error when trying to deploy() it.

Presumably this is because newer framework versions should make use of sagemaker.tensorflow.serving.Model instead, for the new-style TFServing based container instead of the old-style inference container?

Describe the solution you'd like
For these new TF versions where the old-style container isn't supported and there's no "choice", it would be best to make the core TensorFlowModel class produce a TFServing-based model.

Describe alternatives you've considered
Alternatively could raise errors on TensorFlowModel init with a new/unsupported framework version, and consider adding docs deprecation warnings to the old class suggesting the new serving-based class instead for modern framework versions.

Additional context
Clear, centralized documentation of SageMaker-provided framework container image URIs would also help, as it might be clearer what the SDK is trying to do wrong.

[laurenyu]

thanks for the feedback!

For these new TF versions where the old-style container isn't supported and there's no "choice", it would be best to make the core TensorFlowModel class produce a TFServing-based model.

We're actually planning on taking this route. It hasn't happened yet because we've thus far tried to avoid making breaking changes to the SDK, but I will (hopefully soon) be posting an issue/announcement with more information about this :)