Hyperparameters and HPO

Hyperparameters are used to optimize the trained model and are set before training begins. This contrasts with model parameters whose values are determined during the training process.

Hyperparameters are specified using the algorithmHyperParameters key that is part of the SolutionConfig object that is passed to the CreateSolution operation.

A condensed version of the CreateSolution request is below. The example includes the solutionConfig object. You use solutionConfig to override the default parameters of a recipe. When performAutoML is true, all parameters of the solutionConfig object are ignored except for autoMLConfig.

{
    "name": "string",
    "performAutoML": boolean,
    "recipeArn": "string",
    "performHPO": boolean,
    "eventType": "string",
    "solutionConfig": {
        "autoMLConfig": {
            "metricName": "string",
            "recipeList": [ "string" ]
        },
        "eventValueThreshold": "string",
        "featureTransformationParameters": {
            "string" : "string"
        },
        "algorithmHyperParameters": {
            "string" : "string"
        },
        "hpoConfig": {
            "algorithmHyperParameterRanges": {
                ...
            },
            "hpoResourceConfig": {
                "maxNumberOfTrainingJobs": "string",
                "maxParallelTrainingJobs": "string"
            }
        },
    },
}

Different recipes use different hyperparameters. For the available hyperparameters, see the individual recipes in Step 1: Choosing a Recipe.

Hyperparameter optimization (HPO), or tuning, is the task of choosing optimal hyperparameters for a specific learning objective. The optimal hyperparameters are determined by running many training jobs using different values from the specified ranges of possibilities. By default, Amazon Personalize does not perform HPO. To use HPO, set performHPO to true, and include the hpoConfig object.

Hyperparameters can be categorical, continuous, or integer-valued. The hpoConfig object has keys that correspond to each of these types, where you specify the hyperparameters and their ranges. Note that not all hyperparameters can be tuned (see the recipe tables).

The following is a partial example of a CreateSolution request using the HRNN recipe.

{
    "performAutoML": false,
    "recipeArn": "arn:aws:personalize:::recipe/aws-hrnn",
    "performHPO": true,
    "solutionConfig": {
        "algorithmHyperParameters": {
            "hidden_dimension": "55"
        },
        "hpoConfig": {
            "algorithmHyperParameterRanges": {
                "categoricalHyperParameterRanges": [
                    {
                        "name": "recency_mask",
                        "values": [ "true", "false" ]
                    }
                ],
                "integerHyperParameterRanges": [
                    {
                        "name": "bptt",
                        "minValue": 20,
                        "maxValue": 40
                    }
                ]
            },
            "hpoResourceConfig": {
                "maxNumberOfTrainingJobs": "4",
                "maxParallelTrainingJobs": "2"
            }
        }
    }
}

Once training is complete, you can view the hyperparameters of the best performing model by calling the DescribeSolutionVersion operation. The following sample shows a condensed DescribeSolutionVersion output with the optimized hyperparameters displayed in the tunedHPOParams object.

{
   "solutionVersion":{
      "creationDateTime":1562191944.745,
      "datasetGroupArn":"arn:aws:personalize:us-west-2:000000000000:dataset-group/hpo",
      "lastUpdatedDateTime":1562194465.075,
      "performAutoML":false,
      "performHPO":true,
      "recipeArn":"arn:aws:personalize:::recipe/aws-hrnn",
      "solutionArn":"arn:aws:personalize:us-west-2:000000000000:solution/hpo",
      "solutionVersionArn":"arn:aws:personalize:us-west-2:000000000000:solution/hpo/5a515609",
      "status":"ACTIVE",
      "tunedHPOParams":{
         "algorithmHyperParameters":{
            "hidden_dimension":"58",
            "recency_mask":"false"
         }
      }
   }
}

For more information, see Automatic Model Tuning.