add device attr for regularizer, test=develop (#24981)

python/paddle/fluid/optimizer.py

@@ -715,8 +715,8 @@ class Optimizer(object):
             params_grads = append_gradient_clip_ops(params_grads)
 
         # Add regularization if any
-        params_grads = append_regularization_ops(params_grads,
-                                                 self.regularization)
+        params_grads = append_regularization_ops(
+            params_grads, self.regularization, self._param_device_map)
 
         optimize_ops = self._create_optimization_pass(params_grads)
         if table_optimize_op is not None:
@@ -1070,7 +1070,7 @@ class MomentumOptimizer(Optimizer):
 
 class DGCMomentumOptimizer(Optimizer):
     """
-	:api_attr: Static Graph
+	:api_attr: Static Graph
 
     DGC (Deep Gradient Compression) Momentum Optimizer. Original paper is https://arxiv.org/abs/1712.01887
 
@@ -2996,7 +2996,7 @@ Lamb = LambOptimizer
 
 class ModelAverage(Optimizer):
     """
-	:api_attr: Static Graph
+	:api_attr: Static Graph
 
     The ModelAverage optimizer accumulates specific continuous historical parameters
     during training. The accumulated historical range can be controlled by the passed
@@ -3305,7 +3305,7 @@ class ModelAverage(Optimizer):
 
 class ExponentialMovingAverage(object):
     """
-	:api_attr: Static Graph
+	:api_attr: Static Graph
 
     Compute the moving average of parameters with exponential decay.
     Given a parameter :math:`\\theta`, its exponential moving average (EMA)
@@ -3555,7 +3555,7 @@ class ExponentialMovingAverage(object):
 
 class PipelineOptimizer(object):
     """
-	:api_attr: Static Graph
+	:api_attr: Static Graph
 
     Pipeline Optimizer
 
@@ -3857,7 +3857,7 @@ class PipelineOptimizer(object):
 
 class RecomputeOptimizer(Optimizer):
     """
-	:api_attr: Static Graph
+	:api_attr: Static Graph
 
     Recompute Optimizer Wrapper
 
@@ -3931,7 +3931,7 @@ class RecomputeOptimizer(Optimizer):
 
     def load(self, stat_dict):
         """
-	:api_attr: Static Graph
+	:api_attr: Static Graph
 
         load function is not supported by Recompute Optimizer for now.
         :return: None
@@ -4149,7 +4149,7 @@ class RecomputeOptimizer(Optimizer):
 
 class LookaheadOptimizer(object):
     """
-	:api_attr: Static Graph
+	:api_attr: Static Graph
 
     This implements the Lookahead optimizer of the
     paper : https://arxiv.org/abs/1907.08610.

python/paddle/fluid/regularizer.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 import logging
 
 from . import framework
-from .framework import in_dygraph_mode, _varbase_creator
+from .framework import in_dygraph_mode, _varbase_creator, device_guard
 from . import core
 
 __all__ = ['L1Decay', 'L2Decay', 'L1DecayRegularizer', 'L2DecayRegularizer']
@@ -62,7 +62,9 @@ def _create_regularization_of_grad(param, grad, regularization=None):
     return new_grad
 
 
-def append_regularization_ops(parameters_and_grads, regularization=None):
+def append_regularization_ops(parameters_and_grads,
+                              regularization=None,
+                              param_device_map=None):
     """Create and add backward regularization Operators
 
     Creates and adds backward regularization operators in the BlockDesc.
@@ -93,16 +95,19 @@ def append_regularization_ops(parameters_and_grads, regularization=None):
         repeate_regularizer = False
         with framework.name_scope('regularization'):
             for param, grad in parameters_and_grads:
+                device = param_device_map[
+                    param.name] if param_device_map else None
                 if not repeate_regularizer and param.regularizer is not None and regularization is not None:
                     repeate_regularizer = True
                     logging.info(
                         "If regularizer of a Parameter has been set by 'fluid.ParamAttr' or 'fluid.WeightNormParamAttr' already. "
                         "The Regularization[%s] in Optimizer will not take effect, and it will only be applied to other Parameters!"
                         % regularization.__str__())
-                with param.block.program._optimized_guard([param, grad]):
-                    new_grad = _create_regularization_of_grad(param, grad,
-                                                              regularization)
-                    params_and_grads.append((param, new_grad))
+                with device_guard(device):
+                    with param.block.program._optimized_guard([param, grad]):
+                        new_grad = _create_regularization_of_grad(
+                            param, grad, regularization)
+                        params_and_grads.append((param, new_grad))
     return params_and_grads