Support TF_CONFIG environment variable in distribute cases

[xiaozhouX]

estimators support TF_CONFIG environment link We can use TF_CONFIG to specify the cluster spec in distribute tensorflow training. Will we support TF_CONFIG environment variable in benchmarks ?

[kvablack]

+1 to this issue, as it would make the benchmarks much easier to run on setups compatible with the new tf.distribute.Strategy API. For now, I was able to hack it by loading in the config manually using json.loads(os.environ["TF_CONFIG"]) and overriding the params in cnn_benchmark.py.

[gaocegege]

Looking forward to it. Is there any progress?

[reedwm]

I am not very familiar with TF_CONFIG and tf_cnn_benchmarks is unmaintained, so I don't think this will get done.

It looks like TF_CONFIG specifies the same information as --job_name, --worker_hosts, --ps_hosts, and --task_index. If that is all you want from TF_CONFIG, I might have time to add a --use_tf_config boolean flag that uses TF_CONFIG, which automatically those four flags.