Merge branch 'develop' of github.com:baidu/Paddle into feature/add_sum_cost_in_args

python/paddle/trainer_config_helpers/__init__.py

@@ -20,6 +20,4 @@ from layers import *
 from networks import *
 from optimizers import *
 from attrs import *
-
-# This will enable operator overload for LayerOutput
-import math as layer_math
+import layer_math

python/paddle/trainer_config_helpers/attrs.py

@@ -88,6 +88,10 @@ class ParameterAttribute(object):
     :type learning_rate: float or None
     :param momentum: The parameter momentum. None means use global value.
     :type momentum: float or None
+    :param gradient_clipping_threshold: gradient clipping threshold. If gradient
+                                        value larger than some value, will be
+                                        clipped.
+    :type gradient_clipping_threshold: float
     :param sparse_update: Enable sparse update for this parameter. It will
                           enable both local and remote sparse update.
     :type sparse_update: bool
@@ -104,6 +108,7 @@ class ParameterAttribute(object):
                  l2_rate=None,
                  learning_rate=None,
                  momentum=None,
+                 gradient_clipping_threshold=None,
                  sparse_update=False):
         # initialize strategy.
         if is_static:
@@ -152,6 +157,11 @@ class ParameterAttribute(object):
             self.attr['sparse_update'] = True
             self.attr['sparse_remote_update'] = True
 
+        if gradient_clipping_threshold is not None and \
+                is_compatible_with(gradient_clipping_threshold, float):
+            self.attr['gradient_clipping_threshold'] = \
+                gradient_clipping_threshold
+
     def set_default_parameter_name(self, name):
         """
         Set default parameter name. If parameter not set, then will use default