Merge remote-tracking branch 'upstream/master' into newavarecords

official/nlp/modeling/layers/dense_einsum.py

@@ -21,6 +21,8 @@ from __future__ import print_function
 
 import tensorflow as tf
 
+from tensorflow.python.util import deprecation
+
 _CHR_IDX = ["a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m"]
 
 
@@ -57,6 +59,9 @@ class DenseEinsum(tf.keras.layers.Layer):
       `(batch_size, units)`.
   """
 
+  @deprecation.deprecated(
+      None, "DenseEinsum is deprecated. Please use "
+      "tf.keras.experimental.EinsumDense layer instead.")
   def __init__(self,
                output_shape,
                num_summed_dimensions=1,

official/nlp/modeling/layers/multi_channel_attention_test.py

@@ -48,7 +48,11 @@ class MultiChannelAttentionTest(tf.test.TestCase):
     mask_data = np.random.randint(2, size=(3, num_docs, 4, 2))
     doc_probs = np.random.randint(
         2, size=(3, num_heads, 4, num_docs)).astype(float)
-    outputs = attention_layer([from_data, to_data, doc_probs], mask_data)
+    outputs = attention_layer(
+        query=from_data,
+        value=to_data,
+        context_attention_weights=doc_probs,
+        attention_mask=mask_data)
     self.assertEqual(outputs.shape, (3, 4, 8))
 
 

official/nlp/modeling/layers/position_embedding.py

@@ -160,7 +160,6 @@ class RelativePositionEmbedding(tf.keras.layers.Layer):
         "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()))

official/nlp/modeling/layers/transformer_scaffold.py

@@ -262,9 +262,8 @@ class TransformerScaffold(tf.keras.layers.Layer):
     else:
       input_tensor, attention_mask = (inputs, None)
 
-    attention_inputs = [input_tensor, input_tensor]
-
-    attention_output = self._attention_layer(attention_inputs, attention_mask)
+    attention_output = self._attention_layer(
+        query=input_tensor, value=input_tensor, attention_mask=attention_mask)
     attention_output = self._attention_dropout(attention_output)
     attention_output = self._attention_layer_norm(input_tensor +
                                                   attention_output)