diff --git a/tensorflow/python/keras/losses.py b/tensorflow/python/keras/losses.py
index ed8d73747b9..164bb03b469 100644
--- a/tensorflow/python/keras/losses.py
+++ b/tensorflow/python/keras/losses.py
@@ -1056,6 +1056,9 @@ def poisson(y_true, y_pred):
   return K.mean(y_pred - y_true * math_ops.log(y_pred + K.epsilon()), axis=-1)
 
 
+# Retaining the legacy namespaces: 'cosine_proximity' and 'cosine'.
+# TODO(psv): Change name of this function to `cosine_similarity` after fixing
+# estimator test.
 @keras_export(
     'keras.losses.cosine_similarity',
     v1=[
@@ -1066,26 +1069,10 @@ def poisson(y_true, y_pred):
         'keras.losses.cosine_similarity',
     ])
 def cosine_proximity(y_true, y_pred, axis=-1):
-  """Computes the cosine similarity between labels and predictions.
-
-  Note that it is a negative quantity between -1 and 0, where 0 indicates
-  orthogonality and values closer to -1 indicate greater similarity. This makes
-  it usable as a loss function in a setting where you try to maximize the
-  proximity between predictions and targets.
-
-  `loss = -sum(y_true * y_pred)`
-
-  Args:
-    y_true: Tensor of true targets.
-    y_pred: Tensor of predicted targets.
-    axis: Axis along which to determine similarity.
-
-  Returns:
-    Cosine similarity tensor.
-  """
+  """Computes the cosine similarity between labels and predictions."""
   y_true = nn.l2_normalize(y_true, axis=axis)
   y_pred = nn.l2_normalize(y_pred, axis=axis)
-  return -math_ops.reduce_sum(y_true * y_pred, axis=axis)
+  return math_ops.reduce_sum(y_true * y_pred, axis=axis)
 
 
 @keras_export('keras.losses.CosineSimilarity')
diff --git a/tensorflow/python/keras/losses_test.py b/tensorflow/python/keras/losses_test.py
index 79b7e0190f9..f1e2a7f39fc 100644
--- a/tensorflow/python/keras/losses_test.py
+++ b/tensorflow/python/keras/losses_test.py
@@ -605,7 +605,7 @@ class CosineSimilarityTest(test.TestCase):
     self.setup()
     cosine_obj = keras.losses.CosineSimilarity()
     loss = cosine_obj(self.y_true, self.y_pred)
-    expected_loss = -np.mean(self.expected_loss)
+    expected_loss = np.mean(self.expected_loss)
     self.assertAlmostEqual(self.evaluate(loss), expected_loss, 3)
 
   def test_scalar_weighted(self):
@@ -613,7 +613,7 @@ class CosineSimilarityTest(test.TestCase):
     cosine_obj = keras.losses.CosineSimilarity()
     sample_weight = 2.3
     loss = cosine_obj(self.y_true, self.y_pred, sample_weight=sample_weight)
-    expected_loss = -np.mean(self.expected_loss * sample_weight)
+    expected_loss = np.mean(self.expected_loss * sample_weight)
     self.assertAlmostEqual(self.evaluate(loss), expected_loss, 3)
 
   def test_sample_weighted(self):
@@ -624,7 +624,7 @@ class CosineSimilarityTest(test.TestCase):
         self.y_true,
         self.y_pred,
         sample_weight=constant_op.constant(sample_weight))
-    expected_loss = -np.mean(self.expected_loss * sample_weight)
+    expected_loss = np.mean(self.expected_loss * sample_weight)
     self.assertAlmostEqual(self.evaluate(loss), expected_loss, 3)
 
   def test_timestep_weighted(self):
@@ -643,7 +643,7 @@ class CosineSimilarityTest(test.TestCase):
     loss = cosine_obj(
         y_true, y_pred, sample_weight=constant_op.constant(sample_weight))
 
-    expected_loss = -np.mean(expected_loss * sample_weight)
+    expected_loss = np.mean(expected_loss * sample_weight)
     self.assertAlmostEqual(self.evaluate(loss), expected_loss, 3)
 
   def test_zero_weighted(self):
@@ -656,7 +656,7 @@ class CosineSimilarityTest(test.TestCase):
     self.setup(axis=1)
     cosine_obj = keras.losses.CosineSimilarity(axis=1)
     loss = cosine_obj(self.y_true, self.y_pred)
-    expected_loss = -np.mean(self.expected_loss)
+    expected_loss = np.mean(self.expected_loss)
     self.assertAlmostEqual(self.evaluate(loss), expected_loss, 3)
 
 
diff --git a/tensorflow/python/keras/metrics.py b/tensorflow/python/keras/metrics.py
index 16b78946c72..cb34787270f 100644
--- a/tensorflow/python/keras/metrics.py
+++ b/tensorflow/python/keras/metrics.py
@@ -35,6 +35,7 @@ from tensorflow.python.keras.engine import base_layer_utils
 from tensorflow.python.keras.losses import binary_crossentropy
 from tensorflow.python.keras.losses import categorical_crossentropy
 from tensorflow.python.keras.losses import categorical_hinge
+from tensorflow.python.keras.losses import cosine_similarity
 from tensorflow.python.keras.losses import hinge
 from tensorflow.python.keras.losses import kullback_leibler_divergence
 from tensorflow.python.keras.losses import logcosh
@@ -2804,13 +2805,6 @@ def sparse_top_k_categorical_accuracy(y_true, y_pred, k=5):
   return math_ops.cast(
       nn.in_top_k(y_pred, math_ops.cast(y_true, 'int32'), k), K.floatx())
 
-
-def cosine_similarity(y_true, y_pred, axis=-1):
-  """Computes the cosine similarity between labels and predictions."""
-  y_true = nn.l2_normalize(y_true, axis=axis)
-  y_pred = nn.l2_normalize(y_pred, axis=axis)
-  return math_ops.reduce_sum(y_true * y_pred, axis=axis)
-
 # Aliases
 
 mse = MSE = mean_squared_error