Model Profiling Configuration
RIME performs some profiling of the model in order to inform which tests to run.
This can take some time, depending on the size of the dataset, so we provide some options to control this.
By default, RIME will attempt to infer an optimal value for all of these options, so only use these parameters if you think RIME is not selecting appropriate values.
Template
This configuration should be specified within the AI Stress Testing Configuration JSON file, under the model_profiling_info parameter.
{
...,
"model_profiling_info": {
"nrows_for_summary": null,
"nrows_for_feature_importance": null,
"feature_importance_config": {
"path": "path/to/feature/importance.csv",
"feature_imp_column": "featureImportance",
"feature_name_column": "featureName"
},
"impact_metric": null,
"drift_impact_metric": null,
"metric_configs": {...}
}
}
Arguments
nrows_for_summary: int ornull, default =nullThe number of rows to use for calculating summary metrics of model. You may want to specify a smaller amount if making calls to your model takes a while.
nrows_for_feature_importance: int ornull, default =nullThe number of rows to use when calculating feature importance of the model. You may want to specify a smaller amount if making calls to your model takes a while. If a feature importance config is provided, this will be ignored.
feature_importance_config: mapping ornull, default =nullIf you want to provide information about feature importance, you should specify that here. The value of this key should be another dictionary with the following key value pairs:
path: strPath to csv or parquet file containing feature importance information, should be relative to
mount_dirs/datasubdirectory.feature_imp_column: strName of the column in this csv that corresponds to feature importance values.
feature_name_column: strName of the column in this csv that corresponds to the feature name.
impact_metric:MetricNameornull, default =nullThe metric to use when computing model impact for abnormal input and transformation tests.
drift_impact_metric:MetricNameornull, default =nullThe metric to use when computing model impact for drift tests.
impact_label_threshold:int, default =0.8When the fraction of labeled rows in the evaluation data falls below this threshold, Average Prediction is used for impact_metric and drift_impact_metric.
metric_configs: mapping ornull, default =nullThe parameters to configure each metric used during testing. For instance, to configure NDCG to accumulate only to a specific rank
k=50, specify{"normalized_discounted_cumulative_gain": {"k": 50}}.