Fix setting partition selection field in utility analysis result (#511)

analysis/tests/utility_analysis_test.py

@@ -243,7 +243,11 @@ def test_multi_parameters(self):
 
         multi_param = analysis.MultiParameterConfiguration(
             max_partitions_contributed=[1, 2],
-            max_contributions_per_partition=[1, 2])
+            max_contributions_per_partition=[1, 2],
+            partition_selection_strategy=[
+                pipeline_dp.PartitionSelectionStrategy.TRUNCATED_GEOMETRIC,
+                pipeline_dp.PartitionSelectionStrategy.GAUSSIAN_THRESHOLDING
+            ])
 
         # Input collection has 1 privacy id, which contributes to 2 partitions
         # 1 and 2 times correspondingly.
@@ -253,8 +257,6 @@ def test_multi_parameters(self):
             partition_extractor=lambda x: x[1],
             value_extractor=lambda x: 0)
 
-        public_partitions = ["pk0", "pk1"]
-
         output, _ = analysis.perform_utility_analysis(
             col=input,
             backend=pipeline_dp.LocalBackend(),
@@ -264,7 +266,6 @@ def test_multi_parameters(self):
                 aggregate_params=aggregate_params,
                 multi_param_configuration=multi_param),
             data_extractors=data_extractors,
-            public_partitions=public_partitions,
         )
 
         utility_reports = list(output)
@@ -273,16 +274,16 @@ def test_multi_parameters(self):
         self.assertLen(utility_reports, 2)  # one report per each configuration.
 
         # Check the parameter configuration
-        expected_noise_std = [3.02734375, 8.56262117843085]
+        expected_noise_std = [5.9765625, 16.904271487740903]
         expected_l0_error = [-0.5, 0]
-        expected_partition_info = metrics.PartitionsInfo(
-            public_partitions=True,
-            num_dataset_partitions=2,
-            num_non_public_partitions=0,
-            num_empty_partitions=0)
         for i_configuration, report in enumerate(utility_reports):
             self.assertEqual(report.configuration_index, i_configuration)
-            self.assertEqual(report.partitions_info, expected_partition_info)
+            self.assertFalse(report.partitions_info.public_partitions)
+            self.assertEqual(report.partitions_info.num_dataset_partitions, 2)
+            self.assertEqual(report.partitions_info.num_dataset_partitions, 2)
+            self.assertEqual(
+                report.partitions_info.strategy,
+                multi_param.partition_selection_strategy[i_configuration])
             self.assertLen(report.metric_errors, 1)  # metrics for COUNT
             errors = report.metric_errors[0]
             self.assertEqual(errors.metric, pipeline_dp.Metrics.COUNT)

analysis/utility_analysis.py

@@ -115,21 +115,6 @@ def perform_utility_analysis(
         "Compute cross-partition metrics")
     #  ((configuration_index, bucket), UtilityReport)
 
-    if public_partitions is None:
-        # Add partition selection strategy for private partitions.
-        strategies = data_structures.get_partition_selection_strategy(options)
-
-        def add_partition_selection_strategy(report: metrics.UtilityReport):
-            # Beam does not allow to change input arguments in map, so copy it.
-            report = copy.deepcopy(report)
-            report.partitions_info.strategy = strategies[
-                report.configuration_index]
-            return report
-
-        cross_partition_metrics = backend.map_values(
-            cross_partition_metrics, add_partition_selection_strategy,
-            "Add Partition Selection Strategy")
-
     cross_partition_metrics = backend.map_tuple(
         cross_partition_metrics, lambda key, value: (key[0], (key[1], value)),
         "Rekey")
@@ -140,6 +125,22 @@ def add_partition_selection_strategy(report: metrics.UtilityReport):
     # (configuration_index, Iterable[(bucket, UtilityReport)])
     result = backend.map_tuple(cross_partition_metrics, _group_utility_reports,
                                "Group utility reports")
+    if public_partitions is None:
+        # Add partition selection strategy for private partitions.
+        strategies = data_structures.get_partition_selection_strategy(options)
+
+        def add_partition_selection_strategy(report: metrics.UtilityReport):
+            # Beam does not allow to change input arguments in map, so copy it.
+            report = copy.deepcopy(report)
+            strategy = strategies[report.configuration_index]
+            report.partitions_info.strategy = strategy
+            for bin in report.utility_report_histogram:
+                bin.report.partitions_info.strategy = strategy
+            return report
+
+        result = backend.map(result, add_partition_selection_strategy,
+                             "Add Partition Selection Strategy")
+        # result: (UtilityReport)
     # result: (UtilityReport)
     return result, per_partition_result