Track corrupted output_similarity.

PiperOrigin-RevId: 317342104
google · Jun 19, 2020 · 862f338 · 862f338
1 parent 990e3e7
commit 862f338
Showing 1 changed file with 14 additions and 6 deletions.
diff --git a/baselines/cifar/utils.py b/baselines/cifar/utils.py
@@ -283,7 +283,8 @@ def aggregate_corrupt_metrics(metrics,
     Dictionary of aggregated results.
 
   """
-  diversity_keys = ['disagreement', 'cosine_similarity', 'average_kl']
+  diversity_keys = ['disagreement', 'cosine_similarity', 'average_kl',
+                    'outputs_similarity']
   results = {
       'test/nll_mean_corrupted': 0.,
       'test/accuracy_mean_corrupted': 0.,
@@ -305,7 +306,7 @@ def aggregate_corrupt_metrics(metrics,
     disagreement = np.zeros(len(corruption_types))
     cosine_similarity = np.zeros(len(corruption_types))
     average_kl = np.zeros(len(corruption_types))
-
+    outputs_similarity = np.zeros(len(corruption_types))
     for i in range(len(corruption_types)):
       dataset_name = '{0}_{1}'.format(corruption_types[i], intensity)
       nll[i] = metrics['test/nll_{}'.format(dataset_name)].result()
@@ -321,17 +322,21 @@ def aggregate_corrupt_metrics(metrics,
             dataset_name)].result()
         member_ece[i] = 0.
       if corrupt_diversity is not None:
+        error = 1 - acc[i] + tf.keras.backend.epsilon()
         disagreement[i] = (
             corrupt_diversity['corrupt_diversity/disagreement_{}'.format(
-                dataset_name)].result())
+                dataset_name)].result()) / error
         # Normalize the corrupt disagreement by its error rate.
-        error = 1 - acc[i] + tf.keras.backend.epsilon()
         cosine_similarity[i] = (
             corrupt_diversity['corrupt_diversity/cosine_similarity_{}'.format(
-                dataset_name)].result()) / error
+                dataset_name)].result())
         average_kl[i] = (
             corrupt_diversity['corrupt_diversity/average_kl_{}'.format(
                 dataset_name)].result())
+        outputs_similarity[i] = (
+            corrupt_diversity['corrupt_diversity/outputs_similarity_{}'.format(
+                dataset_name)].result())
+
       if log_fine_metrics or output_dir is not None:
         fine_metrics_results['test/nll_{}'.format(dataset_name)] = nll[i]
         fine_metrics_results['test/accuracy_{}'.format(dataset_name)] = acc[i]
@@ -343,6 +348,9 @@ def aggregate_corrupt_metrics(metrics,
               dataset_name)] = cosine_similarity[i]
           fine_metrics_results['corrupt_diversity/average_kl_{}'.format(
               dataset_name)] = average_kl[i]
+          fine_metrics_results['corrupt_diversity/outputs_similarity_{}'.format(
+              dataset_name)] = outputs_similarity[i]
+
     avg_nll = np.mean(nll)
     avg_accuracy = np.mean(acc)
     avg_ece = np.mean(ece)
@@ -363,7 +371,7 @@ def aggregate_corrupt_metrics(metrics,
     results['test/member_ece_mean_corrupted'] += avg_member_ece
     if corrupt_diversity is not None:
       avg_diversity_metrics = [np.mean(disagreement), np.mean(
-          cosine_similarity), np.mean(average_kl)]
+          cosine_similarity), np.mean(average_kl), np.mean(outputs_similarity)]
       for key, avg in zip(diversity_keys, avg_diversity_metrics):
         results['corrupt_diversity/{}_mean_{}'.format(
             key, intensity)] = avg