Should not generate summary during decoding in dot_product_relative_atention (#1618)

Author: phamthuonghai

tensor2tensor/layers/common_attention.py

@@ -1745,7 +1745,7 @@ def dot_product_attention_relative(q,
       save_weights_to[scope.name] = weights
       save_weights_to[scope.name + "/logits"] = logits
     weights = tf.nn.dropout(weights, 1.0 - dropout_rate)
-    if not tf.get_variable_scope().reuse and make_image_summary:
+    if not tf.get_variable_scope().reuse and common_layers.should_generate_summaries() and make_image_summary:
       attention_image_summary(weights, image_shapes)
     return _relative_attention_inner(weights, v, relations_values, False)