How to use dynamic arguments in SageMaker job using Step Functions?

Question

I am using AWS Step Functions to orchestrate some SageMaker jobs. However, I can't seem to dynamically retrieve the Var variable for each job. Meaning that each launched container should be run with a specific argument that is specified in the Jobs section in the following yaml file. Is there a workaround ? I have tried using Environment variable but it also doesn't support dynamic arguments.

```
Comment: "Run  SageMaker jobs in parallel with a max concurrency of 4"
StartAt: PrepareJobList
States:
  PrepareJobList:
    Type: Pass
    Result:
      Jobs: 
        - { Var: "X" }
        - { Var: "XX" }
    ResultPath: "$.JobList"
    Next: RunJobs

RunJobs:
    Type: Map
    Iterator:
      StartAt: GenerateJobName
      States:
        GenerateJobName:
          Type: Pass
          Parameters:
            "TrainingJobName.$": "States.Format('Job-{}-{}', $.Var, States.UUID())"
            "Var.$": "$.Var"
          ResultPath: "$.JobDetails"
          Next: RunSageMakerJob

RunSageMakerJob:
          Type: Task
          Resource: arn:aws:states:::sagemaker:createTrainingJob.sync
          Parameters:
            TrainingJobName.$: "$.JobDetails.TrainingJobName"
            AlgorithmSpecification:
              TrainingImage: "X"
              TrainingInputMode: "File"
              ContainerEntrypoint:
                - python
                - src/script.py
              ContainerArguments:
                - "--var"
                - "$.JobDetails.Var"
            StoppingCondition:
              MaxRuntimeInSeconds: 36000  
            RoleArn: "X"  
            ResourceConfig:
              InstanceType: "ml.m5.large"  
              InstanceCount: 1  
              VolumeSizeInGB: 50  
            OutputDataConfig:
              S3OutputPath: "x"
          Retry:
            - ErrorEquals: ["ResourceLimitExceeded"]
              IntervalSeconds: 300
              MaxAttempts: 10
              BackoffRate: 2.0
          End: true
    ItemsPath: "$.JobList.Jobs"
    MaxConcurrency: 4
    End: true

```

Alex T · Accepted Answer

It's not exactly clear from the question, but I'm guessing the error you see is that the unresolved `"$.JobDetails.Var"` string is what gets passed through to your job?

Unlike in `TrainingJobName.$`, you haven't [ended the parameter name with .$](https://docs.aws.amazon.com/step-functions/latest/dg/connect-parameters.html#connect-parameters-path) for `ContainerArguments.$`, to indicate that it's a pathspec.

However since this field is an array, I believe you'll need to use [this trick](https://stackoverflow.com/a/63713027) for something like:

```
ContainerArguments.$: "States.Array('--var', $.JobDetails.Var)"
```

Unfortunately I don't have a similar complete example to hand to easily test with.

For what it's worth, you may prefer to use the `HyperParameters` dictionary for this? Hyperparams are more broadly visible in SageMaker console & UIs than container arguments - and would be eligible for things like automatic HP tuning later if you wanted to use that. Provided hyperparameters should get automatically written into a `/opt/ml/input/config/hyperparameters.json` file in your container.

If you wanted, you could use the [sagemaker-training-toolkit](https://github.com/aws/sagemaker-training-toolkit) in your custom container to load environment variables & run your script with hyperparam CLI arguments like [SageMaker Script Mode](https://sagemaker.readthedocs.io/en/stable/overview.html#prepare-a-training-script) does. If you're already using one of the AWS-provided framework images, they'll already have something like this (might be slightly different e.g. the sagemaker-*pytorch*-training-toolkit or similar).

How to use dynamic arguments in SageMaker job using Step Functions?

新增您的答案

相關內容