spot_instances.py

estimator = TensorFlow(
    entry_point='model.py',               # Script that has the model
    train_instance_type='ml.p3.2xlarge',  # The instance type for training
    train_instance_count=1,               # The number of instances to be spawned
    model_dir='/opt/ml/model',            # The location for the trained model
    hyperparameters={                     # Hyperparameters for the model
      'epochs': 250, 
      'batch_size': 128, 
      'learning_rate': 0.001
    },
    role=sagemaker.get_execution_role(),  # SageMaker's execution role
    base_job_name='spot-training-test',   # Name of the training job
    framework_version='1.15',             # Tensorflow version
    py_version='py3',                     # Python version
    script_mode=True,                     # Script mode enabled
    train_use_spot_instances=True,        # Use spot instance
    train_max_run=600,                    # Max training time
    train_max_wait=3600,                  # Max training time + spot waiting time
)