transformer, attention layers

official/nlp/modeling/layers/attention.py

@@ -523,9 +523,8 @@ class CachedAttention(MultiHeadAttention):
 
     # Take the dot product between "query" and "key" to get the raw
     # attention scores.
+    query = tf.multiply(query,1.0 / math.sqrt(float(self._key_size)))
     attention_scores = tf.einsum(self._dot_product_equation, key, query)
-    attention_scores = tf.multiply(attention_scores,
-                                   1.0 / math.sqrt(float(self._key_size)))
 
     # Normalize the attention scores to probabilities.
     # `attention_scores` = [B, N, F, T]

official/nlp/modeling/layers/transformer.py

@@ -65,6 +65,9 @@ class Transformer(tf.keras.layers.Layer):
                activity_regularizer=None,
                kernel_constraint=None,
                bias_constraint=None,
+               use_bias=True,
+               norm_first=False,
+               norm_epsilon=1e-12,
                **kwargs):
     super(Transformer, self).__init__(**kwargs)
 
@@ -81,6 +84,9 @@ class Transformer(tf.keras.layers.Layer):
     self._activity_regularizer = tf.keras.regularizers.get(activity_regularizer)
     self._kernel_constraint = tf.keras.constraints.get(kernel_constraint)
     self._bias_constraint = tf.keras.constraints.get(bias_constraint)
+    self._use_bias = use_bias
+    self._norm_first = norm_first
+    self._norm_epsilon = norm_epsilon
 
   def build(self, input_shape):
     input_tensor = input_shape[0] if len(input_shape) == 2 else input_shape
@@ -117,6 +123,7 @@ class Transformer(tf.keras.layers.Layer):
         num_heads=self._num_heads,
         key_size=self._attention_head_size,
         dropout=self._attention_dropout_rate,
+        use_bias=self._use_bias,
         name="self_attention",
         **common_kwargs)
     # pylint: disable=protected-access
@@ -132,7 +139,7 @@ class Transformer(tf.keras.layers.Layer):
         tf.keras.layers.LayerNormalization(
             name="self_attention_layer_norm",
             axis=-1,
-            epsilon=1e-12,
+            epsilon=self._norm_epsilon,
             dtype=tf.float32))
     self._intermediate_dense = tf.keras.layers.experimental.EinsumDense(
         "abc,cd->abd",
@@ -157,7 +164,8 @@ class Transformer(tf.keras.layers.Layer):
     self._output_dropout = tf.keras.layers.Dropout(rate=self._dropout_rate)
     # Use float32 in layernorm for numeric stability.
     self._output_layer_norm = tf.keras.layers.LayerNormalization(
-        name="output_layer_norm", axis=-1, epsilon=1e-12, dtype=tf.float32)
+        name="output_layer_norm", axis=-1, epsilon=self._norm_epsilon,
+        dtype=tf.float32)
 
     super(Transformer, self).build(input_shape)
 
@@ -203,13 +211,22 @@ class Transformer(tf.keras.layers.Layer):
       target_tensor = input_tensor[:, 0:self._output_range, :]
       attention_mask = attention_mask[:, 0:self._output_range, :]
     else:
+      if self._norm_first:
+        source_tensor = input_tensor
+        input_tensor = self._attention_layer_norm(input_tensor)
       target_tensor = input_tensor
 
     attention_output = self._attention_layer(
         query=target_tensor, value=input_tensor, attention_mask=attention_mask)
     attention_output = self._attention_dropout(attention_output)
-    attention_output = self._attention_layer_norm(target_tensor +
-                                                  attention_output)
+    if self._norm_first:
+      attention_output = source_tensor + attention_output
+    else:
+      attention_output = self._attention_layer_norm(target_tensor +
+                                                    attention_output)
+    if self._norm_first:
+      source_attention_output = attention_output
+      attention_output = self._output_layer_norm(attention_output)
     intermediate_output = self._intermediate_dense(attention_output)
     intermediate_output = self._intermediate_activation_layer(
         intermediate_output)
@@ -219,7 +236,10 @@ class Transformer(tf.keras.layers.Layer):
     # is always fp32 for now. Cast layer_output to fp32 for the subsequent
     # add.
     layer_output = tf.cast(layer_output, tf.float32)
-    layer_output = self._output_layer_norm(layer_output + attention_output)
+    if self._norm_first:
+      layer_output = source_attention_output + layer_output
+    else:
+      layer_output = self._output_layer_norm(layer_output + attention_output)
 
     return layer_output
 
@@ -273,6 +293,9 @@ class TransformerDecoderLayer(tf.keras.layers.Layer):
                activity_regularizer=None,
                kernel_constraint=None,
                bias_constraint=None,
+               use_bias=True,
+               norm_first=False,
+               norm_epsilon=1e-12,
                **kwargs):
     super(TransformerDecoderLayer, self).__init__(**kwargs)
     self.num_attention_heads = num_attention_heads
@@ -289,6 +312,9 @@ class TransformerDecoderLayer(tf.keras.layers.Layer):
     self._activity_regularizer = tf.keras.regularizers.get(activity_regularizer)
     self._kernel_constraint = tf.keras.constraints.get(kernel_constraint)
     self._bias_constraint = tf.keras.constraints.get(bias_constraint)
+    self._use_bias = use_bias
+    self._norm_first = norm_first
+    self._norm_epsilon = norm_epsilon
     if self.multi_channel_cross_attention:
       self._cross_attention_cls = multi_channel_attention.MultiChannelAttention
     else:
@@ -318,6 +344,7 @@ class TransformerDecoderLayer(tf.keras.layers.Layer):
         num_heads=self.num_attention_heads,
         key_size=self.attention_head_size,
         dropout=self.attention_dropout_rate,
+        use_bias=self._use_bias,
         name="self_attention",
         **common_kwargs)
     self.self_attention_output_dense = tf.keras.layers.experimental.EinsumDense(
@@ -330,13 +357,15 @@ class TransformerDecoderLayer(tf.keras.layers.Layer):
         rate=self.dropout_rate)
     self.self_attention_layer_norm = (
         tf.keras.layers.LayerNormalization(
-            name="self_attention_layer_norm", axis=-1, epsilon=1e-12))
+            name="self_attention_layer_norm",
+            axis=-1, epsilon=self._norm_epsilon))
     # Encoder-decoder attention.
     self.encdec_attention = self._cross_attention_cls(
         num_heads=self.num_attention_heads,
         key_size=self.attention_head_size,
         dropout=self.attention_dropout_rate,
         output_shape=hidden_size,
+        use_bias=self._use_bias,
         name="attention/encdec",
         **common_kwargs)
 
@@ -344,7 +373,8 @@ class TransformerDecoderLayer(tf.keras.layers.Layer):
         rate=self.dropout_rate)
     self.encdec_attention_layer_norm = (
         tf.keras.layers.LayerNormalization(
-            name="attention/encdec_output_layer_norm", axis=-1, epsilon=1e-12))
+            name="attention/encdec_output_layer_norm",
+            axis=-1, epsilon=self._norm_epsilon))
 
     # Feed-forward projection.
     self.intermediate_dense = tf.keras.layers.experimental.EinsumDense(
@@ -363,7 +393,7 @@ class TransformerDecoderLayer(tf.keras.layers.Layer):
         **common_kwargs)
     self.output_dropout = tf.keras.layers.Dropout(rate=self.dropout_rate)
     self.output_layer_norm = tf.keras.layers.LayerNormalization(
-        name="output_layer_norm", axis=-1, epsilon=1e-12)
+        name="output_layer_norm", axis=-1, epsilon=self._norm_epsilon)
     super(TransformerDecoderLayer, self).build(input_shape)
 
   def common_layers_with_encoder(self):
@@ -384,6 +414,9 @@ class TransformerDecoderLayer(tf.keras.layers.Layer):
           "TransformerDecoderLayer must have 4 inputs, but it got: %d" %
           len(inputs))
     input_tensor, memory, attention_mask, self_attention_mask = inputs[:4]
+    source_tensor = input_tensor
+    if self._norm_first:
+      input_tensor = self.self_attention_layer_norm(input_tensor)
     self_attention_output, cache = self.self_attention(
         query=input_tensor,
         value=input_tensor,
@@ -391,8 +424,15 @@ class TransformerDecoderLayer(tf.keras.layers.Layer):
         cache=cache,
         decode_loop_step=decode_loop_step)
     self_attention_output = self.self_attention_dropout(self_attention_output)
-    self_attention_output = self.self_attention_layer_norm(
-        input_tensor + self_attention_output)
+    if self._norm_first:
+      self_attention_output = source_tensor + self_attention_output
+    else:
+      self_attention_output = self.self_attention_layer_norm(
+          input_tensor + self_attention_output)
+    if self._norm_first:
+      source_self_attention_output = self_attention_output
+      self_attention_output = self.encdec_attention_layer_norm(
+          self_attention_output)
     cross_attn_inputs = dict(
         query=self_attention_output,
         value=memory,
@@ -402,13 +442,23 @@ class TransformerDecoderLayer(tf.keras.layers.Layer):
       cross_attn_inputs["context_attention_weights"] = inputs[-1]
     attention_output = self.encdec_attention(**cross_attn_inputs)
     attention_output = self.encdec_attention_dropout(attention_output)
-    attention_output = self.encdec_attention_layer_norm(self_attention_output +
-                                                        attention_output)
+    if self._norm_first:
+      attention_output = source_self_attention_output + attention_output
+    else:
+      attention_output = self.encdec_attention_layer_norm(
+          self_attention_output +
+          attention_output)
+    if self._norm_first:
+      source_attention_output = attention_output
+      attention_output = self.output_layer_norm(attention_output)
 
     intermediate_output = self.intermediate_dense(attention_output)
     intermediate_output = self.intermediate_activation_layer(
         intermediate_output)
     layer_output = self.output_dense(intermediate_output)
     layer_output = self.output_dropout(layer_output)
-    layer_output = self.output_layer_norm(layer_output + attention_output)
+    if self._norm_first:
+      layer_output = source_attention_output + layer_output
+    else:
+      layer_output = self.output_layer_norm(layer_output + attention_output)
     return layer_output, cache