Support Clip in param_attr (#6729)

* Support Clip in param_attr

* Fix the order of clip & regular

Regular is not need to be clipped

python/paddle/v2/fluid/__init__.py

@@ -16,12 +16,13 @@ import regularizer
 from param_attr import ParamAttr
 from data_feeder import DataFeeder
 from core import LoDTensor, CPUPlace, GPUPlace
+import clip
 
 Tensor = LoDTensor
 __all__ = framework.__all__ + executor.__all__ + [
     'io', 'initializer', 'layers', 'nets', 'optimizer', 'backward',
     'regularizer', 'LoDTensor', 'CPUPlace', 'GPUPlace', 'Tensor', 'ParamAttr'
-    'DataFeeder'
+    'DataFeeder', 'clip'
 ]
 
 

python/paddle/v2/fluid/clip.py

+import functools
+import layers
+
+__all__ = ['GradientClipByValue', 'append_gradient_clip_ops']
+
+
+class BaseGradientClipAttr(object):
+    def process_context(self, context, p_g):
+        raise NotImplementedError()
+
+    def create_operators(self, param, grad):
+        raise NotImplementedError()
+
+
+class NullGradientClipAttr(BaseGradientClipAttr):
+    def process_context(self, context, p_g):
+        pass
+
+    def create_operators(self, param, grad):
+        return param, grad
+
+
+class GradientClipByValue(BaseGradientClipAttr):
+    def __init__(self, max, min=None):
+        max = float(max)
+        if min is None:
+            min = -max
+        else:
+            min = float(min)
+        self.max = max
+        self.min = min
+
+    def process_context(self, context, p_g):
+        pass
+
+    def create_operators(self, param, grad):
+        new_grad = layers.clip(x=grad, min=self.min, max=self.max)
+        return param, new_grad
+
+
+def append_gradient_clip_ops(param_grad):
+    context = dict()
+    create_op_callbacks = []
+    for p, g in param_grad:
+        clip_attr = getattr(p, 'clip_attr', NullGradientClipAttr())
+        if clip_attr is None:
+            clip_attr = NullGradientClipAttr()
+        if not isinstance(clip_attr, BaseGradientClipAttr):
+            raise TypeError(
+                "clip attribute should be an instance of BaseGradientClippingAttr"
+            )
+
+        clip_attr.process_context(context=context, p_g=param_grad)
+        create_op_callbacks.append(
+            functools.partial(
+                clip_attr.create_operators, param=p, grad=g))
+
+    return [each_callback() for each_callback in create_op_callbacks]
+
+
+ClipByValue = GradientClipByValue

python/paddle/v2/fluid/framework.py

@@ -704,6 +704,7 @@ class Block(object):
                 trainable=p.trainable,
                 optimize_attr=p.optimize_attr,
                 regularizer=p.regularizer,
+                clip_attr=p.clip_attr,
                 name=v.name)
             self.vars[new_p.name] = new_p
 
@@ -866,6 +867,8 @@ class Parameter(Variable):
 
         self.regularizer = kwargs.get('regularizer', None)
 
+        self.clip_attr = kwargs.get('clip_attr', None)
+
 
 # program is a global instance.
 _main_program_ = Program()

python/paddle/v2/fluid/optimizer.py

@@ -6,6 +6,7 @@ from framework import unique_name, program_guard
 from initializer import Constant
 from layer_helper import LayerHelper
 from regularizer import append_regularization_ops
+from clip import append_gradient_clip_ops
 
 __all__ = ['SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad']
 
@@ -197,9 +198,13 @@ class Optimizer(object):
         `create_optimization_pass()` into one.
         """
         params_grads = append_backward_ops(loss, parameter_list, no_grad_set)
+
+        params_grads = append_gradient_clip_ops(params_grads)
+
         # Add regularization if any
         params_grads = append_regularization_ops(params_grads,
                                                  self.regularization)
+
         optimize_ops = self.create_optimization_pass(params_grads, loss,
                                                      startup_program)
         return optimize_ops

python/paddle/v2/fluid/param_attr.py

 from initializer import Initializer, Xavier, Constant
 from regularizer import WeightDecayRegularizer
 
+__all__ = ['ParamAttr']
+
 
 class ParamAttr(object):
     def __init__(self,
@@ -8,12 +10,14 @@ class ParamAttr(object):
                  initializer=None,
                  learning_rate=1.0,
                  regularizer=None,
-                 trainable=True):
+                 trainable=True,
+                 clip=None):
         self.name = name
         self.initializer = initializer
         self.learning_rate = learning_rate
         self.regularizer = regularizer
         self.trainable = trainable
+        self.clip = clip
 
     def set_default_initializer(self, initializer):
         if initializer is None:
@@ -56,7 +60,8 @@ class ParamAttr(object):
             'name': self.name,
             'learning_rate': self.learning_rate,
             'regularizer': self.regularizer,
-            'trainable': self.trainable
+            'trainable': self.trainable,
+            'clip_attr': self.clip
         }
         if with_initializer:
             kwargs['initializer'] = self.initializer

python/paddle/v2/fluid/tests/book/test_recognize_digits_mlp.py

@@ -11,7 +11,9 @@ regularizer = fluid.regularizer.L2Decay(0.0005 * BATCH_SIZE)
 hidden1 = fluid.layers.fc(input=image,
                           size=128,
                           act='relu',
-                          param_attr=regularizer)
+                          param_attr=fluid.ParamAttr(
+                              regularizer=regularizer,
+                              clip=fluid.clip.ClipByValue(10)))
 hidden2 = fluid.layers.fc(input=hidden1,
                           size=64,
                           act='relu',