unit double (#32908)

python/paddle/fluid/incubate/fleet/parameter_server/pslib/node.py

@@ -123,7 +123,7 @@ class DownpourServer(Server):
             support_accessor_class = [
                 'DownpourFeatureValueAccessor', 'DownpourCtrAccessor',
                 'DownpourSparseValueAccessor', 'DownpourCtrDoubleAccessor',
-                'DownpourUnitAccessor'
+                'DownpourUnitAccessor', 'DownpourDoubleUnitAccessor'
             ]
             if strategy.get('sparse_accessor_class') is not None:
                 accessor_class = strategy.get('sparse_accessor_class')
@@ -254,7 +254,7 @@ class DownpourServer(Server):
                 table2.param = 2
                 table2.converter = converter
                 table2.deconverter = deconverter
-            elif accessor_class == 'DownpourUnitAccessor':
+            elif accessor_class == 'DownpourUnitAccessor' or accessor_class == 'DownpourDoubleUnitAccessor':
                 self.add_sparse_table_common_config(table, strategy)
                 self.add_sparse_optimizer(table.accessor.embed_sgd_param,
                                           strategy, "embed_")
@@ -380,7 +380,7 @@ class DownpourServer(Server):
         table.accessor.fea_dim = fea_dim
 
     def add_sparse_optimizer(self, sgd, strategy, prefix):
-        optimizer_name = strategy.get(prefix + "sparse_optimizer", "adam")
+        optimizer_name = strategy.get(prefix + "sparse_optimizer", "adagrad")
         sgd.name = optimizer_name
         if optimizer_name == "naive":
             sgd.naive.learning_rate = \
@@ -394,6 +394,19 @@ class DownpourServer(Server):
                 strategy.get(prefix + 'sparse_learning_rate', 0.05)
             sgd.adagrad.initial_range = \
                 strategy.get(prefix + 'sparse_initial_range', 1e-4)
+            if prefix == "embed_":
+                sgd.adagrad.initial_range = 0
+            sgd.adagrad.initial_g2sum = strategy.get(
+                prefix + 'sparse_initial_g2sum', 3)
+            bounds = strategy.get(prefix + 'sparse_weight_bounds', [-10, 10])
+            sgd.adagrad.weight_bounds.extend(bounds)
+        elif optimizer_name == "std_adagrad":
+            sgd.adagrad.learning_rate = \
+                strategy.get(prefix + 'sparse_learning_rate', 0.05)
+            sgd.adagrad.initial_range = \
+                strategy.get(prefix + 'sparse_initial_range', 1e-4)
+            if prefix == "embed_":
+                sgd.adagrad.initial_range = 0
             sgd.adagrad.initial_g2sum = strategy.get(
                 prefix + 'sparse_initial_g2sum', 3)
             bounds = strategy.get(prefix + 'sparse_weight_bounds', [-10, 10])

python/paddle/fluid/incubate/fleet/parameter_server/pslib/optimizer_factory.py

@@ -319,6 +319,7 @@ class DistributedAdam(DistributedOptimizerImplBase):
                 # user do not have to set it in config_fleet
                 if accessor == "DownpourFeatureValueAccessor" \
                         or accessor == "DownpourCtrAccessor" \
+                        or accessor == "DownpourDoubleUnitAccessor" \
                         or accessor == "DownpourUnitAccessor":
                     if st.get("sparse_embedx_dim") is not None \
                             and st["sparse_embedx_dim"] != emb_to_size[key] - 3:
@@ -534,7 +535,7 @@ class DistributedAdam(DistributedOptimizerImplBase):
         if server._server.downpour_server_param.downpour_table_param[
                 0].accessor.accessor_class in [
                     "DownpourCtrAccessor", "DownpourCtrDoubleAccessor",
-                    "DownpourUnitAccessor"
+                    "DownpourUnitAccessor", "DownpourDoubleUnitAccessor"
                 ]:
             opt_info["dump_slot"] = True
         elif server._server.downpour_server_param.downpour_table_param[