object_detection: Allow to provide RunConfig parameters on command line of model_main

System information

• What is the top-level directory of the model you are using: object_detection
• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): no
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 16.04.5 LTS (Xenial Xerus)
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): ('v1.10.1-0-g4dcfddc5d1', '1.10.1')
• Bazel version (if compiling from source): -
• CUDA/cuDNN version: cuda-9.0
• GPU model and memory: Tesla K80 major: 3 minor: 7 / 11.17GiB
• Exact command to reproduce: - (This is a feature request)

Describe the problem

An Estimator run is configured with a RunConfig object, however when starting model_main parameters of RunConfig can not be defined to overwrite defaults. That is why I propose to add command line parameters to model_main that allow to define RunConfig parameters. I see two options here:

1. Add specific command line parameters for specific RunConfig parameters, e.g. --save_checkpoints_steps, --save_checkpoints_secs, ...
2. Add a generic command line parameter for RunConfig parameters similar to the already existing --hparams_overrides. E.g. --runconfig_overrides save_checkpoints_steps=100,keep_checkpoint_max=10

Option 2 should result in less maintenance when RunConfig is changed. OTOH it would only allow to overwrite primitive parameters, but no structured parameters like session_config. With Option 1 changes in RunConfig might need to be reflected in code. OTOH it would for example allow to define command line switches for session_config parameters.

I would personally vote for Option 2 and if there is interest I could also open a PR for this.