Refine multi-head attention

Author: ranqiu

python/paddle/trainer_config_helpers/networks.py

@@ -1557,15 +1557,15 @@ def multi_head_attention(query,
     for i in range(head_num):
         with mixed_layer(size=key_proj_size) as sub_query_proj:
             sub_query_proj += identity_projection(
-                query_proj, offset=key_proj_size * i)
+                query_proj, offset=key_proj_size * i, size=key_proj_size)
 
         with mixed_layer(size=key_proj_size) as sub_key_proj:
             sub_key_proj += identity_projection(
-                key_proj, offset=key_proj_size * i)
+                key_proj, offset=key_proj_size * i, size=key_proj_size)
 
         with mixed_layer(size=value_proj_size) as sub_value_proj:
             sub_value_proj += identity_projection(
-                value_proj, offset=value_proj_size * i)
+                value_proj, offset=value_proj_size * i, size=value_proj_size)
 
         if attention_type == 'dot-product attention':
             m = linear_comb_layer(
@@ -1603,11 +1603,7 @@ def multi_head_attention(query,
 
         head_list.append(head)
 
-    multi_head = concat_layer(head_list)
-
-    with mixed_layer(
-            size=value_proj_size * head_num, name='%s_proj' % name) as attended:
-        attended += full_matrix_projection(multi_head)
+    attended = concat_layer(head_list)
 
     return attended