Migrate to tf gelu. Will remove activations/gelu.py after dependencies are cleaned up.

PiperOrigin-RevId: 323499265

official/modeling/activations/gelu.py

@@ -14,12 +14,6 @@
 # ==============================================================================
 """Gaussian error linear unit."""
 
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import math
-
 import tensorflow as tf
 
 
@@ -35,6 +29,4 @@ def gelu(x):
   Returns:
     `x` with the GELU activation applied.
   """
-  cdf = 0.5 * (1.0 + tf.tanh(
-      (math.sqrt(2 / math.pi) * (x + 0.044715 * tf.pow(x, 3)))))
-  return x * cdf
+  return tf.keras.activations.gelu(x, approximate=True)

official/nlp/xlnet/xlnet_modeling.py

@@ -14,32 +14,14 @@
 # ==============================================================================
 """Keras layers of XLNet model in TF 2.0."""
 
-from __future__ import absolute_import
-from __future__ import division
-# from __future__ import google_type_annotations
-from __future__ import print_function
-
 import copy
-import numpy as np
+import functools
 
 import tensorflow as tf
 from official.nlp.xlnet import data_utils
 
 
-def gelu(x):
-  """Gaussian Error Linear Unit.
-
-  This is a smoother version of the RELU.
-  Original paper: https://arxiv.org/abs/1606.08415
-  Args:
-    x: float Tensor to perform activation.
-
-  Returns:
-    `x` with the GELU activation applied.
-  """
-  cdf = 0.5 * (1.0 + tf.tanh(
-      (np.sqrt(2 / np.pi) * (x + 0.044715 * tf.pow(x, 3)))))
-  return x * cdf
+gelu = functools.partial(tf.keras.activations.gelu, approximate=True)
 
 
 def rel_shift(x, klen=-1):