def run_cloudtuner()

in src/python/tensorflow_cloud/core/run.py [0:0]

• 36 lines of code
• 9 McCabe index (conditional complexity)

def run_cloudtuner(num_jobs=1, **kwargs):
    """A wrapper for tfc.run that allows for running concurrent CloudTuner jobs.

    This method takes the same parameters as tfc.run() and it allows duplicating
    a job multiple times to enable running parallel tuning jobs using
    CloudTuner. All jobs are identical except they will have a unique
    KERASTUNER_TUNER_ID environment variable set in the cluster to enable tuning
    job concurrency. This feature is only supported in Notebooks and Colab.

    Args:
        num_jobs: Number of concurrent jobs to be submitted to AI Platform
        training. Note that these are clones of the same job that are executed
        independently. Setting this value to 1 is identical to just calling
        `tfc.run()`.
        **kwargs: keyword arguments for `tfc.run()`.

    Returns:
        A dictionary with two keys.'job_ids' - a list of training job ids
        and 'docker_image'- Docker image generated for the training job.
    """
    # If code is triggered in a cloud environment, do nothing.
    if remote():
        return

    if num_jobs < 1:
        raise ValueError("num_jobs must be greater than 0.")

    run_results = run(**kwargs)

    # Setting prameters for rerun, notes paramters have already been validated
    # in tfc.run()
    docker_img_uri = run_results["docker_image"]
    chief_config = kwargs.pop("chief_config", "auto")
    worker_count = kwargs.pop("worker_count", 0)
    worker_config = kwargs.pop("worker_config", "auto")
    entry_point_args = kwargs.pop("entry_point_args", None)
    stream_logs = kwargs.pop("stream_logs", False)
    job_labels = kwargs.pop("job_labels", None)
    service_account = kwargs.pop("service_account", None)

    job_ids = [run_results["job_id"]]
    for _ in range(1, num_jobs):
        # Setting a unique default Tuner_ID for each Job,
        # This is to ensure all workers within a cluster (job) ask for the same
        # Tuner parameters, while across multiple jobs they recive different
        # HP Tuning parameters.
        default_tuner_id = f"TUNER_ID_{uuid.uuid4().hex}"
        exnteded_entry_point_args = [default_tuner_id]
        if entry_point_args:
            exnteded_entry_point_args.extend(entry_point_args)

        # Deploy Docker image on the cloud.
        job_ids.extend([
            deploy.deploy_job(
                docker_img_uri,
                chief_config,
                worker_count,
                worker_config,
                exnteded_entry_point_args,
                stream_logs,
                job_labels=job_labels,
                service_account=service_account,
            )
        ])

    return {
        "job_ids": job_ids,
        "docker_image": docker_img_uri,
    }