Revert "Add relative positional embedding to KerasBERT (#8617)" (#8619)

This reverts commit c3c2386c.

official/nlp/modeling/layers/position_embedding.py

@@ -19,8 +19,6 @@ from __future__ import division
 # from __future__ import google_type_annotations
 from __future__ import print_function
 
-import math
-
 import tensorflow as tf
 
 from official.modeling import tf_utils
@@ -120,81 +118,3 @@ class PositionEmbedding(tf.keras.layers.Layer):
       position_embeddings = self._position_embeddings
 
     return tf.broadcast_to(position_embeddings, input_shape)
-
-@tf.keras.utils.register_keras_serializable(package="Text")
-class RelativePositionEmbedding(tf.keras.layers.Layer):
-  """Creates a positional embedding.
-  This layer calculates the position encoding as a mix of sine and cosine
-  functions with geometrically increasing wavelengths. Defined and formulized in
-   "Attention is All You Need", section 3.5.
-  (https://arxiv.org/abs/1706.03762).
-  Arguments:
-    hidden_size: Size of the hidden layer.
-    min_timescale: Minimum scale that will be applied at each position
-    max_timescale: Maximum scale that will be applied at each position.
-    length: Number of positions. Should be specified if `inputs` is None at
-      `call(self, inputs)`
-  """
-
-  def __init__(self,
-               hidden_size,
-               min_timescale=1.0,
-               max_timescale=1.0e4,
-               length=None,
-               **kwargs):
-    # We need to have a default dtype of float32, since the inputs (which Keras
-    # usually uses to infer the dtype) will always be int32.
-    # We compute the positional encoding in float32 even if the model uses
-    # float16, as many of the ops used, like log and exp, are numerically
-    # unstable in float16.
-    if "dtype" not in kwargs:
-      kwargs["dtype"] = "float32"
-
-    super(RelativePositionEmbedding, self).__init__(**kwargs)
-    self._hidden_size = hidden_size
-    self._min_timescale = min_timescale
-    self._max_timescale = max_timescale
-    self._length = length
-
-  def get_config(self):
-    config = {
-        "hidden_size": self._hidden_size,
-        "min_timescale": self._min_timescale,
-        "max_timescale": self._max_timescale,
-        "length": self._length,
-    }
-    base_config = super(RelativePositionEmbedding, self).get_config()
-    return dict(list(base_config.items()) + list(config.items()))
-
-  def build(self, input_shape):
-    """Implements build() for the layer."""
-    super(RelativePositionEmbedding, self).build(input_shape)
-
-  def call(self, inputs):
-    """Implements call() for the layer."""
-    length = self._length
-    if inputs is None and length is None:
-      raise ValueError(
-          "If inputs is None, `length` must be set in "
-          "RelativePositionEmbedding().")
-    if inputs is not None:
-      input_shape = tf_utils.get_shape_list(inputs)
-      if length is not None and length != input_shape[1]:
-        raise ValueError(
-            "If inputs is not None, `length` must equal to input_shape[1]."
-        )
-      length = input_shape[1]
-    position = tf.cast(tf.range(length), tf.float32)
-    num_timescales = self._hidden_size // 2
-    min_timescale, max_timescale = self._min_timescale, self._max_timescale
-    log_timescale_increment = (
-        math.log(float(max_timescale) / float(min_timescale)) /
-        (tf.cast(num_timescales, tf.float32) - 1))
-    inv_timescales = min_timescale * tf.exp(
-        tf.cast(tf.range(num_timescales), tf.float32) *
-        -log_timescale_increment)
-    scaled_time = tf.expand_dims(position, 1) * tf.expand_dims(inv_timescales,
-                                                               0)
-    position_embeddings = tf.concat([tf.sin(scaled_time), tf.cos(scaled_time)],
-                                    axis=1)
-    return position_embeddings

official/nlp/modeling/layers/position_embedding_test.py

@@ -36,7 +36,7 @@ class PositionEmbeddingLayerTest(keras_parameterized.TestCase):
     sequence_length = 21
     width = 30
     input_tensor = tf.keras.Input(shape=(sequence_length, width))
-    output_tensor = test_layer(input_tensor) # pylint: disable=not-callable
+    output_tensor = test_layer(input_tensor)
 
     # When using static positional embedding shapes, the output is expected
     # to be the same as the input shape in all dimensions save batch.
@@ -51,7 +51,7 @@ class PositionEmbeddingLayerTest(keras_parameterized.TestCase):
     sequence_length = 21
     width = 30
     input_tensor = tf.keras.Input(shape=(sequence_length, width))
-    output_tensor = test_layer(input_tensor) # pylint: disable=not-callable
+    output_tensor = test_layer(input_tensor)
 
     # When using static positional embedding shapes, the output is expected
     # to be the same as the input shape in all dimensions save batch.
@@ -67,7 +67,7 @@ class PositionEmbeddingLayerTest(keras_parameterized.TestCase):
     # Create a 3-dimensional input (the first dimension is implicit).
     width = 30
     input_tensor = tf.keras.Input(shape=(None, width))
-    output_tensor = test_layer(input_tensor) # pylint: disable=not-callable
+    output_tensor = test_layer(input_tensor)
 
     # When using dynamic positional embedding shapes, the output is expected
     # to be the same as the input shape in all dimensions - but may be None if
@@ -82,7 +82,7 @@ class PositionEmbeddingLayerTest(keras_parameterized.TestCase):
     # Create a 3-dimensional input (the first dimension is implicit).
     width = 30
     input_tensor = tf.keras.Input(shape=(None, width))
-    output_tensor = test_layer(input_tensor) # pylint: disable=not-callable
+    output_tensor = test_layer(input_tensor)
 
     model = tf.keras.Model(input_tensor, output_tensor)
 
@@ -98,34 +98,6 @@ class PositionEmbeddingLayerTest(keras_parameterized.TestCase):
 
     self.assertAllEqual([1, input_length, width], output_data.shape)
 
-  def test_relative_tensor_input(self):
-    hidden_size = 8
-    test_layer = position_embedding.RelativePositionEmbedding(
-        hidden_size=hidden_size)
-
-    # create a 3-dimensional input for test_layer to infer length as 1.
-    input_tensor = tf.constant([[[0] * hidden_size]])
-    output_tensor = test_layer(input_tensor) # pylint: disable=not-callable
-
-    # expected output is the theoretical result of the input based on
-    # sine cosine relative position embedding formula.
-    expected_output_tensor = tf.constant([[0, 0, 0, 0, 1, 1, 1, 1]])
-    self.assertAllEqual(output_tensor, expected_output_tensor)
-
-  def test_relative_length_input(self):
-    hidden_size = 8
-
-    # When we do not have tensor as input, we explicitly specify length
-    # value when initializing test_layer.
-    test_layer = position_embedding.RelativePositionEmbedding(
-        hidden_size=hidden_size, length=1)
-    input_tensor = None
-    output_tensor = test_layer(input_tensor) # pylint: disable=not-callable
-
-    # expected output is the theoretical result of the input based on
-    # sine cosine relative position embedding formula.
-    expected_output_tensor = tf.constant([[0, 0, 0, 0, 1, 1, 1, 1]])
-    self.assertAllEqual(output_tensor, expected_output_tensor)
 
 if __name__ == "__main__":
   tf.test.main()

official/nlp/transformer/transformer.py

@@ -22,7 +22,6 @@ from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
-from official.nlp.modeling.layers import position_embedding
 from official.nlp.transformer import attention_layer
 from official.nlp.transformer import beam_search
 from official.nlp.transformer import embedding_layer
@@ -171,9 +170,9 @@ class Transformer(tf.keras.Model):
       attention_bias = tf.cast(attention_bias, self.params["dtype"])
 
       with tf.name_scope("add_pos_encoding"):
-        pos_layer = position_embedding.RelativePositionEmbedding(
-            hidden_size=self.params["hidden_size"])
-        pos_encoding = pos_layer(embedded_inputs)
+        length = tf.shape(embedded_inputs)[1]
+        pos_encoding = model_utils.get_position_encoding(
+            length, self.params["hidden_size"])
         pos_encoding = tf.cast(pos_encoding, self.params["dtype"])
         encoder_inputs = embedded_inputs + pos_encoding
 
@@ -210,9 +209,8 @@ class Transformer(tf.keras.Model):
                                 [[0, 0], [1, 0], [0, 0]])[:, :-1, :]
       with tf.name_scope("add_pos_encoding"):
         length = tf.shape(decoder_inputs)[1]
-        pos_layer = position_embedding.RelativePositionEmbedding(
-            hidden_size=self.params["hidden_size"])
-        pos_encoding = pos_layer(decoder_inputs)
+        pos_encoding = model_utils.get_position_encoding(
+            length, self.params["hidden_size"])
         pos_encoding = tf.cast(pos_encoding, self.params["dtype"])
         decoder_inputs += pos_encoding
       if training:
@@ -235,10 +233,8 @@ class Transformer(tf.keras.Model):
   def _get_symbols_to_logits_fn(self, max_decode_length, training):
     """Returns a decoding function that calculates logits of the next tokens."""
 
-    pos_layer = position_embedding.RelativePositionEmbedding(
-        hidden_size=self.params["hidden_size"],
-        length=max_decode_length + 1)
-    timing_signal = pos_layer(None)
+    timing_signal = model_utils.get_position_encoding(
+        max_decode_length + 1, self.params["hidden_size"])
     timing_signal = tf.cast(timing_signal, self.params["dtype"])
     decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
         max_decode_length, dtype=self.params["dtype"])