[bug] Cannot initialize the training job with TF Estimator when the user uses 1 worker and 0 PS

[gaocegege]

There are some users want to use TFJob to run local training jobs with Estimator. They will have such a config:

 spec:
    tfReplicaSpecs:
      Worker:
        replicas: 1
        template:
          spec:
            containers:
            - command:
              - /bin/bash
              - -c
              - python cifar10_main.py --data-dir=/clever/input/datasets/cifar10 
                --job-dir=/tmp/cifar10 --num-gpus=0 --train-steps=1000

One worker will be created with TF_CONFIG:

{
    "cluster": {
        "worker": [
            "tensorflow-190911-rr84s-worker-0.tsk.svc:2222"
        ]
    },
    "task": {
        "type": "worker",
        "index": 0
    },
    "environment": "cloud"
}

The estimator will try to load TF_CONFIG from the env var. then it will get the config.

But we do not expect this. We should not set TF_CONFIG for this scenario.

[issue-label-bot]

Issue-Label Bot is automatically applying the label kind/bug to this issue, with a confidence of 0.90. Please mark this comment with 👍 or 👎 to give our bot feedback!

Links: app homepage, dashboard and code for this bot.

[gaocegege]

Here is an example: https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10_estimator

We cannot use 1 worker to run the example with tf-operator. After deleting the env var manually, it can be run with tf-operator.

[gaocegege]

/cc @johnugeorge @richardsliu

I can fix it

/assign @gaocegege

[johnugeorge]

In Pytorch, it is validated to have 1 Master with 0 or more workers

[gaocegege]

@johnugeorge Will pytorch-operator set MASTER_ADDR in this master pod?

[johnugeorge]

I think so. But, it would not matter even if it does it

[gaocegege]

Gotcha. TF_CONFIG has a side effect in this case.