Hyperparameters and HPO

You specify hyperparameters before training to optimize the trained model for your particular use case. 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.

{
  "name": "string",
  "recipeArn": "string",
  "eventType": "string",
  "solutionConfig": {
      "optimizationObjective": {
          "itemAttribute": "string",
          "objectiveSensitivity": "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.

Enabling hyperparameter optimization

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. You must provide each type in your request, but if a recipe doesn't have a parameter of a type, you can leave it empty. For example, User-Personalization does not have a tunable hyperparameter of continuous type. So for the continousHyperParameterRange, you would pass an empty array.

The following code shows how to create a solution with HPO enabled using the SDK for Python (Boto3). The solution in the example uses the User-Personalization recipe recipe and has HPO set to true. The code provides a value for hidden_dimension and the categoricalHyperParameterRanges and integerHyperParameterRanges. The continousHyperParameterRange is empty and the hpoResourceConfig sets the maxNumberOfTrainingJobs and maxParallelTrainingJobs.

create_solution_response = personalize.create_solution(
    name = solutionName,
    datasetGroupArn = 'arn:aws:personalize:region:accountId:dataset-group/datasetGroupName',
    recipeArn = 'arn:aws:personalize:::recipe/aws-user-personalization',
    performHPO = True,
    solutionConfig = {
        "algorithmHyperParameters": {
          "hidden_dimension": "55"
        },
        "hpoConfig": {
          "algorithmHyperParameterRanges": {
              "categoricalHyperParameterRanges": [
                  {
                      "name": "recency_mask",
                      "values": [ "true", "false"]
                  }
              ],
              "integerHyperParameterRanges": [
                  {
                      "name": "bptt",
                      "minValue": 2,
                      "maxValue": 22
                  }
              ],
              "continuousHyperParameterRanges": [

              ]
          },
          "hpoResourceConfig": {
              "maxNumberOfTrainingJobs": "4",
              "maxParallelTrainingJobs": "2"
          }
        }
    }
)

For more information about HPO, see Automatic model tuning.

Viewing hyperparameters

You can view the hyperparameters of the solution by calling the DescribeSolution operation. The following sample shows a DescribeSolution output. After creating a solution version (training a model), you can also view hyperparameters with the DescribeSolutionVersion operation.

{
  "solution": {
    "name": "hpo_coonfig_solution",
    "solutionArn": "arn:aws:personalize:region:accountId:solution/solutionName",
    "performHPO": true,
    "performAutoML": false,
    "recipeArn": "arn:aws:personalize:::recipe/aws-user-personalization",
    "datasetGroupArn": "arn:aws:personalize:region:accountId:dataset-group/datasetGroupName",
    "eventType": "click",
    "solutionConfig": {
      "hpoConfig": {
        "hpoResourceConfig": {
          "maxNumberOfTrainingJobs": "4",
          "maxParallelTrainingJobs": "2"
        },
        "algorithmHyperParameterRanges": {
          "integerHyperParameterRanges": [
            {
              "name": "training.bptt",
              "minValue": 2,
              "maxValue": 22
            }
          ],
          "continuousHyperParameterRanges": [],
          "categoricalHyperParameterRanges": [
            {
              "name": "data.recency_mask",
              "values": [
                "true",
                "false"
              ]
            }
          ]
        }
      },
      "algorithmHyperParameters": {
        "hidden_dimension": "55"
      }
    },
    "status": "ACTIVE",
    "creationDateTime": "2022-07-08T12:12:48.565000-07:00",
    "lastUpdatedDateTime": "2022-07-08T12:12:48.565000-07:00"
  }
}