Use np.asarray for backend-agnostic tensor to numpy conversion.

andyl7an · The Meridian Authors · commit c62a8b0dbfb6 · 2026-02-27T16:38:33.000-08:00
PiperOrigin-RevId: 875786835
diff --git a/meridian/schema/processors/reach_frequency_optimization_processor.py b/meridian/schema/processors/reach_frequency_optimization_processor.py
@@ -447,7 +447,7 @@ def _create_frequency_outcome_grid(
         aggregated_reach = backend.einsum(
             f"{input_tensor_dims}->...{output_tensor_dims}", filtered_reach
         )
-        reach = aggregated_reach.numpy()[-1]
+        reach = np.asarray(aggregated_reach)[-1]
 
         metric_data_array = optimal_frequency_dataset[constants.ROI].sel(
             frequency=frequency, rf_channel=channel

Original file line number	Diff line number	Diff line change
`@@ -447,7 +447,7 @@ def _create_frequency_outcome_grid(`
`447`	`447`	`aggregated_reach = backend.einsum(`
`448`	`448`	`f"{input_tensor_dims}->...{output_tensor_dims}", filtered_reach`
`449`	`449`	`)`
`450`		`- reach = aggregated_reach.numpy()[-1]`
	`450`	`+ reach = np.asarray(aggregated_reach)[-1]`
`451`	`451`
`452`	`452`	`metric_data_array = optimal_frequency_dataset[constants.ROI].sel(`
`453`	`453`	`frequency=frequency, rf_channel=channel`