07_training/serverlessml/hparam-continue.yaml

trainingInput:
  scaleTier: CUSTOM
  masterType: n1-standard-4  # or n1-highcpu-16 
  masterConfig:
    acceleratorConfig:
      count: 2
      type: NVIDIA_TESLA_T4
  hyperparameters:
    goal: MAXIMIZE
    hyperparameterMetricTag: accuracy
    algorithm: GRID_SEARCH
    maxTrials: 45  # have to calculate?
    maxParallelTrials: 2
    enableTrialEarlyStopping: True
    # resumePreviousJobId: flowers_20210101_085015_gpus_one_machine  # doesn't work because params have to be the same
    params:
    - parameterName: batch_size
      type: DISCRETE
      discreteValues:
      - 48
      - 64
      - 96
    - parameterName: num_hidden
      type: DISCRETE
      discreteValues:
      - 16
      - 24
      - 32
    - parameterName: crop_ratio
      type: DISCRETE
      discreteValues:
      - 0.65
      - 0.70
      - 0.75
      - 0.80
      - 0.85