Add average prediction, average label to regression head.

PiperOrigin-RevId: 204154837

5 files changed, 63 insertions, 18 deletions

diff --git a/tensorflow/python/estimator/canned/baseline_test.py b/tensorflow/python/estimator/canned/baseline_test.py
index 7bf2e62da9..e46a3a156d 100644
--- a/tensorflow/python/estimator/canned/baseline_test.py
+++ b/tensorflow/python/estimator/canned/baseline_test.py
@@ -154,6 +154,8 @@ class BaselineRegressorEvaluationTest(test.TestCase):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 9.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -176,6 +178,8 @@ class BaselineRegressorEvaluationTest(test.TestCase):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 18.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -204,6 +208,8 @@ class BaselineRegressorEvaluationTest(test.TestCase):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 27.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -229,7 +235,9 @@ class BaselineRegressorEvaluationTest(test.TestCase):
 
     self.assertItemsEqual(
         (metric_keys.MetricKeys.LOSS, metric_keys.MetricKeys.LOSS_MEAN,
-         ops.GraphKeys.GLOBAL_STEP), eval_metrics.keys())
+         metric_keys.MetricKeys.PREDICTION_MEAN,
+         metric_keys.MetricKeys.LABEL_MEAN, ops.GraphKeys.GLOBAL_STEP),
+        eval_metrics.keys())
 
     # Logit is bias which is [46, 58]
     self.assertAlmostEqual(0, eval_metrics[metric_keys.MetricKeys.LOSS])
diff --git a/tensorflow/python/estimator/canned/dnn_testing_utils.py b/tensorflow/python/estimator/canned/dnn_testing_utils.py
index ba17821259..de226ed0ef 100644
--- a/tensorflow/python/estimator/canned/dnn_testing_utils.py
+++ b/tensorflow/python/estimator/canned/dnn_testing_utils.py
@@ -1271,6 +1271,8 @@ class BaseDNNRegressorEvaluateTest(object):
     self.assertAllClose({
         metric_keys.MetricKeys.LOSS: expected_loss,
         metric_keys.MetricKeys.LOSS_MEAN: expected_loss,
+        metric_keys.MetricKeys.PREDICTION_MEAN: -2.08,
+        metric_keys.MetricKeys.LABEL_MEAN: 1.0,
         ops.GraphKeys.GLOBAL_STEP: global_step
     }, dnn_regressor.evaluate(input_fn=_input_fn, steps=1))
 
@@ -1301,6 +1303,8 @@ class BaseDNNRegressorEvaluateTest(object):
     self.assertAllClose({
         metric_keys.MetricKeys.LOSS: expected_loss,
         metric_keys.MetricKeys.LOSS_MEAN: expected_loss / label_dimension,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 0.39 / 3.0,
+        metric_keys.MetricKeys.LABEL_MEAN: 0.5 / 3.0,
         ops.GraphKeys.GLOBAL_STEP: global_step
     }, dnn_regressor.evaluate(input_fn=_input_fn, steps=1))
 
diff --git a/tensorflow/python/estimator/canned/head.py b/tensorflow/python/estimator/canned/head.py
index b74ef1015c..da9a64c2bc 100644
--- a/tensorflow/python/estimator/canned/head.py
+++ b/tensorflow/python/estimator/canned/head.py
@@ -1398,15 +1398,21 @@ class _RegressionHeadWithMeanSquaredErrorLoss(_Head):
         weights=weights,
         processed_labels=labels)
 
-  def _eval_metric_ops(self, weights, unreduced_loss, regularization_loss):
+  def _eval_metric_ops(self, predicted_value, labels, weights, unreduced_loss,
+                       regularization_loss):
     """Returns the Eval metric ops."""
     keys = metric_keys.MetricKeys
     # Estimator already adds a metric for loss.
     eval_metric_ops = {
         _summary_key(self._name, keys.LOSS_MEAN):
-            metrics_lib.mean(
-                values=unreduced_loss,
-                weights=weights)
+            metrics_lib.mean(values=unreduced_loss, weights=weights),
+        _summary_key(self._name, keys.PREDICTION_MEAN):
+            _predictions_mean(
+                predictions=predicted_value,
+                weights=weights,
+                name=keys.PREDICTION_MEAN),
+        _summary_key(self._name, keys.LABEL_MEAN):
+            metrics_lib.mean(values=labels, weights=weights)
     }
     if regularization_loss is not None:
       regularization_loss_key = _summary_key(
@@ -1489,13 +1495,13 @@ class _RegressionHeadWithMeanSquaredErrorLoss(_Head):
             predictions=predictions,
             loss=regularized_training_loss,
             eval_metrics=_create_eval_metrics_tuple(
-                self._eval_metric_ops,
-                {
+                self._eval_metric_ops, {
+                    'predicted_value': predicted_value,
+                    'labels': labels,
                     'weights': weights,
                     'unreduced_loss': unreduced_loss,
                     'regularization_loss': regularization_loss,
-                }
-            ))
+                }))
 
       # Train.
       if optimizer is not None:
diff --git a/tensorflow/python/estimator/canned/head_test.py b/tensorflow/python/estimator/canned/head_test.py
index 08ce5ca8e8..bd2e0ae943 100644
--- a/tensorflow/python/estimator/canned/head_test.py
+++ b/tensorflow/python/estimator/canned/head_test.py
@@ -3103,8 +3103,10 @@ class RegressionHead(test.TestCase):
     self.assertItemsEqual((prediction_key,), spec.predictions.keys())
     self.assertEqual(dtypes.float32, spec.predictions[prediction_key].dtype)
     self.assertEqual(dtypes.float32, spec.loss.dtype)
-    self.assertItemsEqual(
-        (metric_keys.MetricKeys.LOSS_MEAN,), spec.eval_metric_ops.keys())
+    self.assertItemsEqual((metric_keys.MetricKeys.LOSS_MEAN,
+                           metric_keys.MetricKeys.PREDICTION_MEAN,
+                           metric_keys.MetricKeys.LABEL_MEAN),
+                          spec.eval_metric_ops.keys())
     self.assertIsNone(spec.train_op)
     self.assertIsNone(spec.export_outputs)
     _assert_no_hooks(self, spec)
@@ -3140,6 +3142,9 @@ class RegressionHead(test.TestCase):
 
     expected_metric_keys = [
         '{}/some_regression_head'.format(metric_keys.MetricKeys.LOSS_MEAN),
+        '{}/some_regression_head'.format(
+            metric_keys.MetricKeys.PREDICTION_MEAN),
+        '{}/some_regression_head'.format(metric_keys.MetricKeys.LABEL_MEAN),
     ]
     self.assertItemsEqual(expected_metric_keys, spec.eval_metric_ops.keys())
 
@@ -3170,6 +3175,8 @@ class RegressionHead(test.TestCase):
     expected_metrics = {
         keys.LOSS_MEAN: expected_unregularized_loss,
         keys.LOSS_REGULARIZATION: expected_regularization_loss,
+        keys.PREDICTION_MEAN: (45 + 41) / 2.0,
+        keys.LABEL_MEAN: (43 + 44) / 2.0,
     }
 
     # Assert predictions, loss, and metrics.
@@ -3471,8 +3478,10 @@ class RegressionHead(test.TestCase):
     self.assertItemsEqual((prediction_key,), spec.predictions.keys())
     self.assertEqual(dtypes.float32, spec.predictions[prediction_key].dtype)
     self.assertEqual(dtypes.float32, spec.loss.dtype)
-    self.assertItemsEqual(
-        (metric_keys.MetricKeys.LOSS_MEAN,), spec.eval_metric_ops.keys())
+    self.assertItemsEqual((metric_keys.MetricKeys.LOSS_MEAN,
+                           metric_keys.MetricKeys.PREDICTION_MEAN,
+                           metric_keys.MetricKeys.LABEL_MEAN),
+                          spec.eval_metric_ops.keys())
     self.assertIsNone(spec.train_op)
     self.assertIsNone(spec.export_outputs)
     _assert_no_hooks(self, spec)
@@ -3700,8 +3709,10 @@ class RegressionHead(test.TestCase):
     self.assertItemsEqual((prediction_key,), spec.predictions.keys())
     self.assertEqual(dtypes.float32, spec.predictions[prediction_key].dtype)
     self.assertEqual(dtypes.float32, spec.loss.dtype)
-    self.assertItemsEqual(
-        (metric_keys.MetricKeys.LOSS_MEAN,), spec.eval_metric_ops.keys())
+    self.assertItemsEqual((metric_keys.MetricKeys.LOSS_MEAN,
+                           metric_keys.MetricKeys.PREDICTION_MEAN,
+                           metric_keys.MetricKeys.LABEL_MEAN),
+                          spec.eval_metric_ops.keys())
     self.assertIsNone(spec.train_op)
     self.assertIsNone(spec.export_outputs)
     _assert_no_hooks(self, spec)
@@ -3832,7 +3843,13 @@ class RegressionHead(test.TestCase):
     # losses = [1*(35-45)^2, .1*(42-41)^2, 1.5*(45-44)^2] = [100, .1, 1.5]
     # loss = sum(losses) = 100+.1+1.5 = 101.6
     # loss_mean = loss/(1+.1+1.5) = 101.6/2.6 = 39.076923
-    expected_metrics = {metric_keys.MetricKeys.LOSS_MEAN: 39.076923}
+    expected_metrics = {
+        metric_keys.MetricKeys.LOSS_MEAN:
+            39.076923,
+        metric_keys.MetricKeys.PREDICTION_MEAN:
+            (45 + 41 * 0.1 + 44 * 1.5) / 2.6,
+        metric_keys.MetricKeys.LABEL_MEAN: (35 + 42 * 0.1 + 45 * 1.5) / 2.6,
+    }
 
     # Assert spec contains expected tensors.
     self.assertEqual(dtypes.float32, spec.loss.dtype)
diff --git a/tensorflow/python/estimator/canned/linear_testing_utils.py b/tensorflow/python/estimator/canned/linear_testing_utils.py
index 9e9c2f7c4b..c3934c7a80 100644
--- a/tensorflow/python/estimator/canned/linear_testing_utils.py
+++ b/tensorflow/python/estimator/canned/linear_testing_utils.py
@@ -261,6 +261,8 @@ class BaseLinearRegressorEvaluationTest(object):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 9.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -286,6 +288,8 @@ class BaseLinearRegressorEvaluationTest(object):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 18.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -316,6 +320,8 @@ class BaseLinearRegressorEvaluationTest(object):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 27.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -346,7 +352,9 @@ class BaseLinearRegressorEvaluationTest(object):
 
     self.assertItemsEqual(
         (metric_keys.MetricKeys.LOSS, metric_keys.MetricKeys.LOSS_MEAN,
-         ops.GraphKeys.GLOBAL_STEP), eval_metrics.keys())
+         metric_keys.MetricKeys.PREDICTION_MEAN,
+         metric_keys.MetricKeys.LABEL_MEAN, ops.GraphKeys.GLOBAL_STEP),
+        eval_metrics.keys())
 
     # Logit is
     #   [2., 4., 5.] * [1.0, 2.0] + [7.0, 8.0] = [39, 50] + [7.0, 8.0]
@@ -383,7 +391,9 @@ class BaseLinearRegressorEvaluationTest(object):
     eval_metrics = est.evaluate(input_fn=input_fn, steps=1)
     self.assertItemsEqual(
         (metric_keys.MetricKeys.LOSS, metric_keys.MetricKeys.LOSS_MEAN,
-         ops.GraphKeys.GLOBAL_STEP), eval_metrics.keys())
+         metric_keys.MetricKeys.PREDICTION_MEAN,
+         metric_keys.MetricKeys.LABEL_MEAN, ops.GraphKeys.GLOBAL_STEP),
+        eval_metrics.keys())
 
     # Logit is [(20. * 10.0 + 4 * 2.0 + 5.0), (40. * 10.0 + 8 * 2.0 + 5.0)] =
     # [213.0, 421.0], while label is [213., 421.]. Loss = 0.