remove the optimizer base and learning rate base (#56099)

* remove the optimizer base and learning rate base

* fix bug

* fix bug

python/paddle/amp/auto_cast.py

@@ -250,7 +250,6 @@ def _is_valid_optimizer(optimizer):
         optimizer,
         (
             paddle.optimizer.Optimizer,
-            paddle.fluid.optimizer.Optimizer,
             DygraphShardingOptimizer,
         ),
     )
@@ -260,7 +259,7 @@ def check_optimizers(optimizers):
     for optimizer in optimizers:
         if not _is_valid_optimizer(optimizer):
             raise RuntimeError(
-                "Current train mode is pure fp16, optimizers should be paddle.optimizer.Optimizer or paddle.fluid.optimizer.Optimizer or DygraphShardingOptimizer, but receive {}.".format(
+                "Current train mode is pure fp16, optimizers should be paddle.optimizer.Optimizer or DygraphShardingOptimizer, but receive {}.".format(
                     type(optimizer)
                 )
             )

python/paddle/distributed/auto_parallel/static/engine.py

@@ -146,11 +146,10 @@ def __init__(
 
         if optimizer and not isinstance(
             optimizer,
-            (paddle.optimizer.Optimizer, paddle.static.Optimizer),
+            (paddle.optimizer.Optimizer),
         ):
             raise TypeError(
                 "'optimizer' must be object of class `paddle.optimizer.Optimizer`"
-                " or `paddle.static.Optimizer`."
             )
         self._optimizer = auto_utils.validate_opt(optimizer)
 

python/paddle/distributed/fleet/meta_optimizers/dgc_optimizer.py

@@ -20,11 +20,11 @@
 
 import paddle
 from paddle.common_ops_import import LayerHelper
+from paddle.fluid import framework
 from paddle.fluid.dygraph import base as imperative_base
-from paddle.fluid.optimizer import Optimizer
 from paddle.framework import core, in_dynamic_mode
 from paddle.nn.clip import ClipGradByNorm, append_gradient_clip_ops
-from paddle.optimizer import Momentum
+from paddle.optimizer import Momentum, Optimizer
 from paddle.regularizer import L1Decay, L2Decay
 from paddle.static import create_global_var
 
@@ -58,8 +58,8 @@ def __init__(
         assert momentum is not None
         super().__init__(
             learning_rate=learning_rate,
-            parameter_list=parameter_list,
-            regularization=regularization,
+            parameters=parameter_list,
+            weight_decay=regularization,
             grad_clip=grad_clip,
             name=name,
         )
@@ -396,6 +396,55 @@ def _dgc_op(
             op_maker.kOpRoleVarAttrName(), [param_var.name, grad_var.name]
         )
 
+    def _process_distribute_lookuptable(self, param_grads):
+        """
+        Because distribute lookup table only support SGD optimizer for now, not support
+        other optimizer and regularization, so we should find the table parameter out,
+        and avoid to add regularization and other op for it, and add sgd optimize op
+        for it independently.
+        :param param_grads(list((Var, Var))): list of (param, grad) pair.
+        :param loss: the loss variable.
+        :param startup_program: the startup program
+        """
+        from paddle.distributed.distribute_lookup_table import (
+            find_distributed_lookup_table,
+        )
+
+        program = framework.default_main_program()
+        global_block = framework.default_main_program().global_block()
+        table_name = find_distributed_lookup_table(program)
+        table_param = None
+        table_grad = None
+        new_param_grads = []
+        for p, g in param_grads:
+            if p.name == table_name:
+                if table_param is not None:
+                    raise RuntimeError(
+                        "multi dist table var found, only support one now!"
+                    )
+                table_param = p
+                table_grad = g
+            else:
+                new_param_grads.append((p, g))
+        sgd_op = None
+        if table_param is not None:
+            param_and_grad = [table_param, table_grad]
+            with table_param.block.program._optimized_guard(
+                param_and_grad
+            ), framework.name_scope("optimizer"):
+                self._create_global_learning_rate()
+                # create the optimize op
+                sgd_op = global_block.append_op(
+                    type='sgd',
+                    inputs={
+                        "Param": table_param,
+                        "Grad": table_grad,
+                        "LearningRate": self._create_param_lr(param_and_grad),
+                    },
+                    outputs={"ParamOut": param_and_grad[0]},
+                )
+        return new_param_grads, (table_param, table_grad), sgd_op
+
     @imperative_base.no_grad()
     def apply_gradients(self, params_grads):
         # Note: since we can't use all_reduce_op now,
@@ -532,7 +581,7 @@ def apply_gradients(self, params_grads):
 
     def apply_optimize(self, loss, startup_program, params_grads):
         self._init_dgc_opt()
-        return self.dgc_opt.apply_optimize(
+        return self.dgc_opt._apply_optimize(
             loss, startup_program=startup_program, params_grads=params_grads
         )
 

python/paddle/fluid/__init__.py

@@ -53,7 +53,6 @@
 from .initializer import set_global_initializer
 from . import layers
 from . import dygraph
-from . import optimizer
 from . import backward
 from .backward import gradients
 from . import incubate
@@ -109,7 +108,6 @@
         'disable_dygraph',
         'enable_imperative',
         'disable_imperative',
-        'optimizer',
         'backward',
         'LoDTensor',
         'LoDTensorArray',

python/paddle/fluid/dygraph/__init__.py

@@ -18,9 +18,6 @@
 from . import tracer
 from .tracer import *
 
-from . import learning_rate_scheduler
-from .learning_rate_scheduler import *
 
 __all__ = []
 __all__ += base.__all__
-__all__ += learning_rate_scheduler.__all__