TF 1.15.3 Compatibility

[Mistobaan]

The code is interleaved with TF 1.x features and TF 2.x features making it a nightmare to develop upon. On top of that, there is no explicit dependency set on the setup.py installation to TensorFlow 2.x TPUs tool (ie.e the profiler) only reliable work still with the 1.x version and the TF enterprise is only for 1.15.x (2.1 exists but is not usable by the majority of the codebase).

• Can we have a TF 1.15x / v1 compatibility layer for tf.ranking?
• can you split keras into its own tf-ranking-keras so that the library does not constrain us with the library version?

forking the repository for now.

[ramakumar1729]

You can use the latest release built against TF v1.15, release v0.1.6.

We leave it to the user to pick TF version and whether they want CPU/GPU compatibility. Hence we do not have an explicit dependency, but the releases are tested against specific TF versions. See README and releases.

FYI TF 2.x supports many of TF1.x features via tf.compat.v1.* APIs.

[Mistobaan]

There are many good primitives for ranking in the repo, so I just decided to hyper specialize it for the only large scale real production use case I have:

• TensorFlow 1.15.3 graph mode (NO EAGER)
• TPU v3-8 / 128 / 256
• Apache Beam / Google Dataflow
• Google Cloud Storage
• Kubeflow/Kubernetes
• Python 3.7

Overall only the multiple nested functions are giving me a bit of mind-bending to unroll them.