How to specify hyperparameter configuration to enable HPO in Amazon Personalize? - amazon-web-services

Question

Do we need to specify the hyperparameters both in algorithmHyperParameters and algorithmHyperParameterRanges? If yes, then should we specify a single value (string as per documentation) in algorithmHyperParameters, but a range (integer in case of integer-valued hyperparameter) in algorithmHyperParameterRanges?
For example: Similar-Items recipe has an integer-valued hyperparameter item_id_hidden_dimension. If I use the following solution_config, where item_id_hidden_dimension is specified only in algorithmHyperParameterRanges and not in algorithmHyperParameters, I get the error:
An error occurred (InvalidInputException) when calling the CreateSolution operation: Provide a hyperparameter that is used in the algorithm: arn:aws:personalize:::algorithm/aws-similar-items
"solution_config": {
"algorithmHyperParameters": {},
"hpoConfig": {
"algorithmHyperParameterRanges": {
"integerHyperParameterRanges": [
{
"name": "item_id_hidden_dimension",
"minValue": 30,
"maxValue": 200
}
],
"categoricalHyperParameterRanges": [],
"continuousHyperParameterRanges": []
},
"hpoResourceConfig": {
"maxNumberOfTrainingJobs": "4",
"maxParallelTrainingJobs": "2"
}
}
}
But if I use the following solution_config, where item_id_hidden_dimension is specified both in algorithmHyperParameterRanges and in algorithmHyperParameters, I still get the same error:
An error occurred (InvalidInputException) when calling the CreateSolution operation: Provide a hyperparameter that is used in the algorithm: arn:aws:personalize:::algorithm/aws-similar-items
"solution_config": {
"algorithmHyperParameters": {
"item_id_hidden_dimension": "100"
},
"hpoConfig": {
"algorithmHyperParameterRanges": {
"integerHyperParameterRanges": [
{
"name": "item_id_hidden_dimension",
"minValue": 30,
"maxValue": 200
}
],
"categoricalHyperParameterRanges": [],
"continuousHyperParameterRanges": []
},
"hpoResourceConfig": {
"maxNumberOfTrainingJobs": "4",
"maxParallelTrainingJobs": "2"
}
}
}

Answer 1

This is caused by an error in the documentation. The hyperparameter names should be item_id_hidden_dim and item_metadata_hidden_dim (note they are dim and not dimension as the documentation states).
This can be confirmed by calling the DescribeRecipe API to get the algorithmArn for the Similar-Items recipe and then calling the DescribeAlgorithm API to get details on the algorithm.
import boto3
import json
personalize = boto3.client('personalize')
response = personalize.describe_recipe(recipeArn = 'arn:aws:personalize:::recipe/aws-similar-items')
print(json.dumps(response['recipe'], indent=2, default=str))
{
"name": "aws-similar-items",
"recipeArn": "arn:aws:personalize:::recipe/aws-similar-items",
"algorithmArn": "arn:aws:personalize:::algorithm/aws-similar-items",
"featureTransformationArn": "arn:aws:personalize:::feature-transformation/similar-items",
"status": "ACTIVE",
"description": "Predicts items similar to a given item based on co-occurrence of items in the user-item interactions dataset and item metadata in the item dataset.",
"creationDateTime": "2019-06-10 00:00:00+00:00",
"recipeType": "RELATED_ITEMS",
"lastUpdatedDateTime": "2022-08-17 00:25:42.935000+00:00"
}
algo_arn = response['recipe']['algorithmArn']
response = personalize.describe_algorithm(algorithmArn = algo_arn)
print(json.dumps(response['algorithm'], indent=2, default=str))
{
"name": "aws-similar-items",
"algorithmArn": "arn:aws:personalize:::algorithm/aws-similar-items",
"algorithmImage": {
"name": "Item Similarity"
},
"defaultHyperParameters": {
"item_id_hidden_dim": "100",
"item_metadata_hidden_dim": "100"
},
"defaultHyperParameterRanges": {
"integerHyperParameterRanges": [
{
"name": "item_id_hidden_dim",
"minValue": 30,
"maxValue": 200,
"isTunable": true
},
{
"name": "item_metadata_hidden_dim",
"minValue": 30,
"maxValue": 200,
"isTunable": true
}
],
"continuousHyperParameterRanges": [],
"categoricalHyperParameterRanges": []
},
"defaultResourceConfig": {
"maxNumberOfTrainingJobs": "20",
"maxParallelTrainingJobs": "5"
},
"trainingInputMode": "File",
"creationDateTime": "2019-06-10 00:00:00+00:00",
"lastUpdatedDateTime": "2022-08-17 00:24:41.307000+00:00"
}
Note the hyperparameter names in the last response above.
We will get this error fixed in the documentation ASAP.