Merge pull request #13638 from typhoonzero/cherrypick_13602

add decayed_adagrad support for dist train

Author: typhoonzero

python/paddle/fluid/tests/unittests/test_dist_transpiler.py

@@ -264,6 +264,25 @@ def transpiler_test_impl(self):
         ])
 
 
+class TestDecayedAdagrad(TranspilerTest):
+    def net_conf(self):
+        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
+        y_predict = fluid.layers.fc(input=x,
+                                    size=1000,
+                                    act=None,
+                                    param_attr=fluid.ParamAttr(name='fc_w'),
+                                    bias_attr=fluid.ParamAttr(name='fc_b'))
+        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+        avg_cost = fluid.layers.mean(cost)
+        opt = fluid.optimizer.DecayedAdagrad(learning_rate=0.1)
+        opt.minimize(avg_cost)
+
+    def transpiler_test_impl(self):
+        pserver, startup = self.get_pserver(self.pserver1_ep)
+        trainer, _ = self.get_trainer()
+
+
 class TestLRDecayConditional(TranspilerTest):
     def net_conf(self):
         x = fluid.layers.data(name='x', shape=[1000], dtype='float32')

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -1433,6 +1433,9 @@ def _get_optimizer_input_shape(self, op_type, varkey, orig_shape,
         elif op_type == "rmsprop":
             if varkey in ["Moment", "MeanSquare"]:
                 return param_shape
+        elif op_type == "decayed_adagrad":
+            if varkey == "Moment":
+                return param_shape
         elif op_type == "sgd":
             pass
         return orig_shape