[FluidAPI]remove clip api (#48946)

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

@@ -20,11 +20,11 @@ __all__ = []
 
 import paddle
 from paddle.common_ops_import import LayerHelper
-from paddle.fluid.clip import GradientClipByNorm, append_gradient_clip_ops
 from paddle.fluid.dygraph import base as imperative_base
 from paddle.fluid.framework import in_dygraph_mode
 from paddle.fluid.optimizer import Momentum, Optimizer
 from paddle.framework import core
+from paddle.nn.clip import ClipGradByNorm, append_gradient_clip_ops
 from paddle.static import create_global_var
 
 
@@ -76,9 +76,9 @@ class DGCMomentumOptimizer(Optimizer):
 
         self._dgc_clip_norm = None
         if grad_clip is not None:
-            if not isinstance(grad_clip, GradientClipByNorm):
+            if not isinstance(grad_clip, ClipGradByNorm):
                 raise TypeError(
-                    "The type of grad_clip should be 'GradientClipByNorm', because DGCMomentumOptimizer only support GradientClipByNorm"
+                    "The type of grad_clip should be 'ClipGradByNorm', because DGCMomentumOptimizer only support ClipGradByNorm"
                 )
             assert isinstance(num_trainers, int), (
                 "The type of num_trainers should be 'int', but received %s"

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/hybrid_parallel_optimizer.py

@@ -15,9 +15,8 @@
 import paddle
 from paddle import framework
 from paddle.autograd import no_grad
-from paddle.fluid import layers
-from paddle.fluid.clip import ClipGradByGlobalNorm
 from paddle.framework import core
+from paddle.nn import ClipGradByGlobalNorm, clip
 
 from ...base.topology import ParallelMode
 from ...utils.hybrid_parallel_util import (
@@ -62,8 +61,8 @@ class HybridParallelClipGrad:
                 continue
             merge_grad = g
             if g.type == core.VarDesc.VarType.SELECTED_ROWS:
-                merge_grad = layers.merge_selected_rows(g)
-                merge_grad = layers.get_tensor_from_selected_rows(merge_grad)
+                merge_grad = clip.merge_selected_rows(g)
+                merge_grad = clip.get_tensor_from_selected_rows(merge_grad)
             square = paddle.square(merge_grad)
             sum_square = paddle.sum(square)
 

python/paddle/distributed/fleet/meta_parallel/sharding/group_sharded_optimizer_stage2.py

@@ -30,7 +30,7 @@ import paddle
 import paddle.distributed as dist
 from paddle.distributed import ParallelMode, fleet
 from paddle.fluid import core
-from paddle.fluid.clip import ClipGradByGlobalNorm
+from paddle.nn import ClipGradByGlobalNorm
 from paddle.optimizer import Optimizer
 
 HybridParallelClipGrad = (

python/paddle/distributed/fleet/meta_parallel/sharding/group_sharded_stage3.py

@@ -25,8 +25,8 @@ import paddle.fluid.framework as framework
 from paddle import nn
 from paddle.autograd import PyLayer
 from paddle.distributed import collective
-from paddle.fluid.clip import ClipGradByGlobalNorm
 from paddle.fluid.framework import EagerParamBase
+from paddle.nn import ClipGradByGlobalNorm
 
 from .group_sharded_storage import GradStorage
 from .group_sharded_utils import GroupShardedClipGrad, Type, device_guard

python/paddle/distributed/fleet/meta_parallel/sharding/group_sharded_utils.py

@@ -23,6 +23,7 @@ from paddle import _legacy_C_ops
 from paddle.fluid import core, layers
 from paddle.fluid.dygraph import to_variable
 from paddle.fluid.framework import dygraph_only
+from paddle.nn import clip
 
 
 class Taskflow:
@@ -65,8 +66,8 @@ class GroupShardedClipGrad:
 
             merge_grad = g
             if g.type == core.VarDesc.VarType.SELECTED_ROWS:
-                merge_grad = layers.get_tensor_from_selected_rows(
-                    layers.merge_selected_rows(g)
+                merge_grad = clip.get_tensor_from_selected_rows(
+                    clip.merge_selected_rows(g)
                 )
             square = paddle.square(merge_grad)
             sum_square = paddle.sum(square)

python/paddle/distributed/fleet/metrics/metric.py

@@ -159,7 +159,7 @@ def auc(stat_pos, stat_neg, scope=None, util=None):
         .. code-block:: python
 
           # in model.py
-          similarity_norm = fluid.layers.sigmoid(fluid.layers.clip(output, min=-15.0, max=15.0))
+          similarity_norm = fluid.layers.sigmoid(paddle.clip(output, min=-15.0, max=15.0))
           binary_predict = fluid.layers.concat(
               input=[paddle.subtract(fluid.layers.ceil(similarity_norm), similarity_norm), similarity_norm], axis=1)
           self.auc, batch_auc, [batch_stat_pos, batch_stat_neg, stat_pos, stat_neg] =

python/paddle/fluid/__init__.py

@@ -90,7 +90,6 @@ from .transpiler import (
     DistributeTranspilerConfig,
 )
 from .lod_tensor import create_lod_tensor, create_random_int_lodtensor
-from . import clip
 from . import profiler
 from . import unique_name
 from . import parallel_executor
@@ -164,7 +163,6 @@ __all__ = (
         'ParamAttr',
         'WeightNormParamAttr',
         'DataFeeder',
-        'clip',
         'profiler',
         'unique_name',
         'Scope',

python/paddle/fluid/incubate/fleet/utils/fleet_util.py

@@ -185,7 +185,7 @@ class FleetUtil:
 
               # below is part of model
               emb = my_slot_net(slots, label) # emb can be fc layer of size 1
-              similarity_norm = fluid.layers.sigmoid(fluid.layers.clip(\
+              similarity_norm = fluid.layers.sigmoid(paddle.clip(\
                   emb, min=-15.0, max=15.0), name="similarity_norm")\
               binary_predict = fluid.layers.concat(input=[\
                   paddle.subtract(\
@@ -1374,7 +1374,7 @@ class FleetUtil:
               label = fluid.layers.data(name="click", shape=[-1, 1],\
                   dtype="int64", lod_level=0, append_batch_size=False)
               emb = my_slot_net(slots, label) # emb can be fc layer of size 1
-              similarity_norm = fluid.layers.sigmoid(fluid.layers.clip(\
+              similarity_norm = fluid.layers.sigmoid(paddle.clip(\
                   emb, min=-15.0, max=15.0), name="similarity_norm")\
               binary_predict = fluid.layers.concat(input=[\
                   paddle.subtract(\
@@ -1574,7 +1574,7 @@ class FleetUtil:
               label = fluid.layers.data(name="click", shape=[-1, 1],\
                   dtype="int64", lod_level=0, append_batch_size=False)
               emb = my_slot_net(slots, label) # emb can be fc layer of size 1
-              similarity_norm = fluid.layers.sigmoid(fluid.layers.clip(\
+              similarity_norm = fluid.layers.sigmoid(paddle.clip(\
                   emb, min=-15.0, max=15.0), name="similarity_norm")\
               binary_predict = fluid.layers.concat(input=[\
                   paddle.subtract(\

python/paddle/fluid/layers/nn.py

@@ -63,10 +63,6 @@ __all__ = [
     'fc',
     'embedding',
     'autoincreased_step_counter',
-    'clip',
-    'clip_by_norm',
-    'merge_selected_rows',
-    'get_tensor_from_selected_rows',
 ]
 
 OP_NAMEMAPPING = {
@@ -997,199 +993,3 @@ def _logical_op(op_name, x, y, out=None, name=None, binary_op=True):
             )
 
         return out
-
-
-@templatedoc()
-def clip(x, min, max, name=None):
-    """
-        :old_api: paddle.fluid.layers.clip
-
-    ${comment}
-
-    Args:
-        x(${x_type}): ${x_comment}
-        min(float): ${min_comment}
-        max(float): ${max_comment}
-        name(str, optional): The default value is None.
-                             Normally there is no need for user to set this property.
-                             For more information, please refer to :ref:`api_guide_Name`
-
-    Returns:
-        ${out_comment}
-
-    Return Type:
-        ${out_type}
-
-    Examples:
-        .. code-block:: python
-
-            import paddle.fluid as fluid
-            input = fluid.data(
-                name='data', shape=[1], dtype='float32')
-            reward = fluid.layers.clip(x=input, min=-1.0, max=1.0)
-    """
-
-    helper = LayerHelper("clip", **locals())
-    check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'], 'clip')
-
-    if name is None:
-        name = unique_name.generate_with_ignorable_key(
-            ".".join([helper.name, 'tmp'])
-        )
-
-    out = helper.create_variable(
-        type=x.type, name=name, dtype=x.dtype, persistable=False
-    )
-
-    helper.append_op(
-        type="clip",
-        inputs={"X": x},
-        attrs={"min": min, "max": max},
-        outputs={"Out": out},
-    )
-
-    return out
-
-
-@templatedoc()
-def clip_by_norm(x, max_norm, name=None):
-    """
-    ${comment}
-
-    Args:
-        x(${x_type}): ${x_comment}
-        max_norm(${max_norm_type}): ${max_norm_comment}
-        name(str, optional): For detailed information, please refer
-            to :ref:`api_guide_Name`. Usually name is no need to set and
-            None by default.
-
-    Returns:
-        Tensor:
-
-        out(${out_type}): ${out_comment}
-
-
-    Examples:
-        .. code-block:: python
-
-            import paddle
-            import paddle.fluid as fluid
-
-            input = paddle.to_tensor([[2.0, 2.0], [2.0, 2.0]], dtype='float32')
-            reward = fluid.layers.clip_by_norm(x=input, max_norm=1.0)
-            # [[0.5, 0.5], [0.5, 0.5]]
-    """
-
-    if in_dygraph_mode():
-        return _C_ops.clip_by_norm(x, max_norm)
-    else:
-        helper = LayerHelper("clip_by_norm", **locals())
-        check_variable_and_dtype(x, 'X', ['float32', 'float16'], 'clip_by_norm')
-        check_type(max_norm, 'max_norm', (float), 'clip_by_norm')
-
-        if name is None:
-            name = unique_name.generate_with_ignorable_key(
-                ".".join([helper.name, 'tmp'])
-            )
-
-        out = helper.create_variable(
-            type=x.type, name=name, dtype=x.dtype, persistable=False
-        )
-
-        helper.append_op(
-            type="clip_by_norm",
-            inputs={"X": x},
-            attrs={"max_norm": max_norm},
-            outputs={"Out": out},
-        )
-
-        return out
-
-
-@templatedoc()
-def merge_selected_rows(x, name=None):
-    """
-    ${comment}
-
-    Args:
-        x(${x_type}): ${x_comment}
-        name(basestring|None): Name of the output.
-
-    Returns:
-        out(${out_type}): ${out_comment}
-
-    Examples:
-        .. code-block:: python
-
-            import paddle.fluid as fluid
-            b = fluid.default_main_program().global_block()
-            var = b.create_var(
-                name="X", dtype="float32", persistable=True,
-                type=fluid.core.VarDesc.VarType.SELECTED_ROWS)
-            y = fluid.layers.merge_selected_rows(var)
-    """
-    if in_dygraph_mode():
-        return _C_ops.merge_selected_rows(x)
-    else:
-        helper = LayerHelper("merge_selected_rows", **locals())
-        out = helper.create_variable_for_type_inference(dtype=x.dtype)
-        helper.append_op(
-            type="merge_selected_rows",
-            inputs={"X": x},
-            attrs={},
-            outputs={"Out": out},
-        )
-        return out
-
-
-@templatedoc()
-def get_tensor_from_selected_rows(x, name=None):
-    """
-    This operator gets tensor data from input with SelectedRows type, and outputs a LoDTensor.
-
-    .. code-block:: text
-
-        input x is SelectedRows:
-           x.rows = [0, 5, 5, 4, 19]
-           x.height = 20
-           x.value = [[1, 1] [2, 2] [2, 2] [3, 3] [6, 6]]
-
-        Output is LoDTensor:
-           out.shape = [5, 2]
-           out.data = [[1, 1],
-                       [2, 2],
-                       [2, 2],
-                       [3, 3],
-                       [6, 6]]
-
-    Args:
-        x(SelectedRows): Input with SelectedRows type. The data type is float32, float64, int32 or int64.
-        name(str, optional): The default value is None.  Normally there is no need for user to set this property.
-            For more information, please refer to :ref:`api_guide_Name` .
-
-    Returns:
-        Variable: LoDTensor transformed from SelectedRows. The data type is same with input.
-
-    Examples:
-        .. code-block:: python
-
-            import paddle.fluid as fluid
-            b = fluid.default_main_program().global_block()
-            input = b.create_var(name="X", dtype="float32", persistable=True, type=fluid.core.VarDesc.VarType.SELECTED_ROWS)
-            out = fluid.layers.get_tensor_from_selected_rows(input)
-    """
-
-    check_type(x, 'x', Variable, 'get_tensor_from_selected_rows')
-    if x.type != core.VarDesc.VarType.SELECTED_ROWS:
-        raise TypeError(
-            "The type of 'x' in get_tensor_from_selected_rows must be SELECTED_ROWS."
-        )
-    helper = LayerHelper('get_tensor_from_selected_rows', **locals())
-    out = helper.create_variable_for_type_inference(dtype=x.dtype)
-    helper.append_op(
-        type='get_tensor_from_selected_rows',
-        inputs={'X': x},
-        outputs={'Out': out},
-        attrs={},
-    )
-    return out

python/paddle/fluid/optimizer.py

@@ -38,13 +38,6 @@ from .backward import (
     _append_grad_suffix_,
     _get_no_grad_set_name,
 )
-from .clip import (
-    GradientClipBase,
-    GradientClipByNorm,
-    error_clip_callback,
-    append_gradient_clip_ops,
-    ClipGradByGlobalNorm,
-)
 from .framework import program_guard
 from .initializer import Constant
 from .layer_helper import LayerHelper
@@ -160,7 +153,7 @@ class Optimizer:
                 )
 
         if grad_clip is not None:
-            if not isinstance(grad_clip, GradientClipBase):
+            if not isinstance(grad_clip, paddle.nn.clip.GradientClipBase):
                 raise TypeError(
                     "'grad_clip' should be an instance of GradientClipBase's derived class"
                 )
@@ -1030,7 +1023,7 @@ class Optimizer:
                     params_grads.append((param, grad_var))
         else:
             if callbacks is None:
-                callbacks = [error_clip_callback]
+                callbacks = [paddle.nn.clip.error_clip_callback]
             else:
                 assert isinstance(callbacks, list)
             program = loss.block.program
@@ -1260,7 +1253,7 @@ class Optimizer:
         # NOTE(zhiqiu): currently, only support ClipGradByGlobalNorm and without regularization.
         if self._flatten_param_grads and self.regularization is None:
             if self._grad_clip is None or isinstance(
-                self._grad_clip, ClipGradByGlobalNorm
+                self._grad_clip, paddle.nn.ClipGradByGlobalNorm
             ):
                 params_grads = self.flatten_param_grads(params_grads)
 
@@ -1268,7 +1261,7 @@ class Optimizer:
         if self._grad_clip is not None:
             params_grads = self._grad_clip(params_grads)
         else:
-            params_grads = append_gradient_clip_ops(params_grads)
+            params_grads = paddle.nn.clip.append_gradient_clip_ops(params_grads)
 
         # Add regularization if any
         params_grads = self.append_regularization_ops(

python/paddle/fluid/tests/test_error_clip.py

@@ -38,13 +38,13 @@ with fluid.program_guard(main_program=prog):
 
 prog_clip = prog.clone()
 prog_clip.block(0).var(hidden1.name)._set_error_clip(
-    fluid.clip.ErrorClipByValue(max=CLIP_MAX, min=CLIP_MIN)
+    paddle.nn.clip.ErrorClipByValue(max=CLIP_MAX, min=CLIP_MIN)
 )
 
 avg_cost_clip = prog_clip.block(0).var(avg_cost.name)
 fluid.backward.append_backward(loss=avg_cost)
 fluid.backward.append_backward(
-    loss=avg_cost_clip, callbacks=[fluid.clip.error_clip_callback]
+    loss=avg_cost_clip, callbacks=[paddle.nn.clip.error_clip_callback]
 )
 
 hidden1_grad = prog.block(0).var(hidden1.name + "@GRAD")

python/paddle/fluid/tests/unittests/collective/fleet/pipeline_mnist.py

@@ -122,7 +122,7 @@ class TestDistMnist2x2(TestDistRunnerBase):
 
         opt = paddle.optimizer.AdamW(
             learning_rate=lr_val,
-            grad_clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0),
+            grad_clip=paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0),
         )
 
         acc_steps = 2  # accumulated steps for pipeline

python/paddle/fluid/tests/unittests/collective/fleet/pipeline_mnist_multi_device.py

@@ -122,7 +122,7 @@ class TestDistMnist2x2(TestDistRunnerBase):
         opt = fluid.optimizer.Momentum(
             learning_rate=lr_val,
             momentum=0.9,
-            grad_clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0),
+            grad_clip=paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0),
         )
 
         acc_steps = 2  # accumulated steps for pipeline

python/paddle/fluid/tests/unittests/collective/fleet/test_dgc_optimizer.py

@@ -15,10 +15,10 @@
 import unittest
 
 import paddle
-import paddle.fluid.clip as clip
 import paddle.fluid.framework as framework
 import paddle.fluid.optimizer as optimizer
 import paddle.fluid.regularizer as regularizer
+import paddle.nn.clip as clip
 
 paddle.enable_static()
 
@@ -76,7 +76,7 @@ class TestDGCMomentumOptimizer(unittest.TestCase):
             rampup_begin_step=0,
             num_trainers=2,
             regularization=regularization,
-            grad_clip=clip.GradientClipByNorm(1.0),
+            grad_clip=clip.ClipGradByNorm(1.0),
         )
 
         if use_recompute:
@@ -144,14 +144,14 @@ class TestDGCMomentumOptimizer(unittest.TestCase):
                 print("dgc regular_coeff=" + str(coeff))
 
     def test_tpyeError(self):
-        # the type of DGCMomentumOptimizer(grad_clip=) must be 'GradientClipByNorm'
+        # the type of DGCMomentumOptimizer(grad_clip=) must be 'ClipGradByNorm'
         with self.assertRaises(TypeError):
             dgc_momentum_optimizer = self.MockDGCMomentum(
                 learning_rate=0.01,
                 momentum=0.2,
                 rampup_begin_step=0,
                 num_trainers=2,
-                grad_clip=clip.GradientClipByGlobalNorm(1.0),
+                grad_clip=clip.ClipGradByGlobalNorm(1.0),
             )
 
     def test_momentum_without_dgc(self):

python/paddle/fluid/tests/unittests/collective/fleet/test_fleet_hybrid_meta_optimizer.py

@@ -354,7 +354,7 @@ class TestFleetHybridOptimizer(TestFleetMetaOptimizer):
         }
         strategy.fuse_all_reduce_ops = True
         strategy.fuse_grad_size_in_MB = 32
-        clip = paddle.fluid.clip.GradientClipByGlobalNorm(1.0)
+        clip = paddle.nn.ClipGradByGlobalNorm(1.0)
 
         self.optimizer(
             avg_cost, strategy, train_prog, startup_prog, grad_clip=clip
@@ -552,7 +552,7 @@ class TestFleetHybridOptimizer(TestFleetMetaOptimizer):
         strategy.fuse_all_reduce_ops = True
         strategy.fuse_grad_size_in_MB = 32
         strategy.fuse_grad_merge = True
-        clip = paddle.fluid.clip.GradientClipByGlobalNorm(1.0)
+        clip = paddle.nn.ClipGradByGlobalNorm(1.0)
 
         self.optimizer(
             avg_cost, strategy, train_prog, startup_prog, grad_clip=clip
@@ -940,7 +940,7 @@ class TestFleetHybridOptimizerBoundary(TestFleetMetaOptimizer):
         }
         strategy.fuse_all_reduce_ops = True
         strategy.fuse_grad_size_in_MB = 32
-        clip = paddle.fluid.clip.GradientClipByGlobalNorm(1.0)
+        clip = paddle.nn.ClipGradByGlobalNorm(1.0)
 
         self.optimizer(
             avg_cost, strategy, train_prog, startup_prog, grad_clip=clip
@@ -1044,7 +1044,7 @@ class TestFleetHybridOptimizerBoundary(TestFleetMetaOptimizer):
         }
         strategy.fuse_all_reduce_ops = True
         strategy.fuse_grad_size_in_MB = 32
-        clip = paddle.fluid.clip.GradientClipByGlobalNorm(1.0)
+        clip = paddle.nn.ClipGradByGlobalNorm(1.0)
 
         self.optimizer(
             avg_cost, strategy, train_prog, startup_prog, grad_clip=clip

python/paddle/fluid/tests/unittests/collective/fleet/test_fleet_sharding_meta_optimizer.py

@@ -640,7 +640,7 @@ class TestFleetShardingMetaOptimizer(TestFleetMetaOptimizer):
         )
         avg_cost, strategy = self.net(train_prog, startup_prog)
         self.set_strategy(strategy, 'sharding')
-        clip = paddle.fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0)
+        clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0)
         self.optimizer(
             avg_cost, strategy, train_prog, startup_prog, grad_clip=clip
         )
@@ -1309,7 +1309,7 @@ class TestFleetShardingHybridOptimizer(TestFleetMetaOptimizer):
             "micro_batch_size": 2,
             "accumulate_steps": 4,
         }
-        clip = paddle.fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0)
+        clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0)
         self.optimizer(
             avg_cost, strategy, train_prog, startup_prog, grad_clip=clip
         )
@@ -1547,7 +1547,7 @@ class TestFleetShardingHybridOptimizer(TestFleetMetaOptimizer):
             "micro_batch_size": 2,
             "accumulate_steps": 4,
         }
-        clip = paddle.fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0)
+        clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0)
         self.optimizer(
             avg_cost,
             strategy,

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

@@ -22,8 +22,8 @@ import paddle
 import paddle.distributed.fleet as fleet
 import paddle.fluid.core as core
 from paddle.distributed.fleet.meta_optimizers.common import CollectiveHelper
-from paddle.fluid.clip import ClipGradBase, _clip_by_global_norm_using_mp_type
 from paddle.incubate import DistributedFusedLamb
+from paddle.nn.clip import ClipGradBase, _clip_by_global_norm_using_mp_type
 from paddle.vision.models import resnet18 as resnet
 
 

python/paddle/fluid/tests/unittests/dygraph_to_static/test_len.py

@@ -19,6 +19,7 @@ import numpy as np
 import paddle
 import paddle.fluid as fluid
 from paddle.jit.dy2static import Call
+from paddle.nn import clip
 
 SEED = 2020
 np.random.seed(SEED)
@@ -89,11 +90,11 @@ def len_with_selected_rows(place):
         type=fluid.core.VarDesc.VarType.SELECTED_ROWS,
     )
     # y is Variable(SelectedRows)
-    y = fluid.layers.merge_selected_rows(var)
+    y = clip.merge_selected_rows(var)
     y_len = Call(len)(y)
 
     # z is inner tensor with shape [4, 2]
-    z = fluid.layers.get_tensor_from_selected_rows(y)
+    z = clip.get_tensor_from_selected_rows(y)
     z_len = Call(len)(z)
 
     # set data for selected_rows

python/paddle/fluid/tests/unittests/dygraph_to_static/test_seq2seq.py

@@ -22,8 +22,8 @@ from seq2seq_dygraph_model import AttentionModel, BaseModel
 from seq2seq_utils import Seq2SeqModelHyperParams, get_data_iter
 
 import paddle.fluid as fluid
-from paddle.fluid.clip import GradientClipByGlobalNorm
 from paddle.jit import ProgramTranslator
+from paddle.nn import ClipGradByGlobalNorm
 
 place = (
     fluid.CUDAPlace(0) if fluid.is_compiled_with_cuda() else fluid.CPUPlace()
@@ -71,7 +71,7 @@ def train(args, attn_model=False):
                 dropout=args.dropout,
             )
 
-        gloabl_norm_clip = GradientClipByGlobalNorm(args.max_grad_norm)
+        gloabl_norm_clip = ClipGradByGlobalNorm(args.max_grad_norm)
         optimizer = fluid.optimizer.SGD(
             args.learning_rate,
             parameter_list=model.parameters(),

python/paddle/fluid/tests/unittests/ir/inference/test_mkldnn_elt_act_fuse_pass.py

@@ -127,7 +127,7 @@ class ElementwiseActivationMkldnnFusePassTest_Add_Clip(
 ):
     def set_params(self):
         self.operand = paddle.add
-        self.act = fluid.layers.clip
+        self.act = paddle.clip
         self.act_alpha = 0.0
         self.act_beta = 10.0
 
@@ -219,7 +219,7 @@ class ElementwiseActivationMkldnnFusePassTest_Sub_Clip(
 ):
     def set_params(self):
         self.operand = paddle.subtract
-        self.act = fluid.layers.clip
+        self.act = paddle.clip
         self.act_alpha = 0.0
         self.act_beta = 10.0
 
@@ -319,7 +319,7 @@ class ElementwiseActivationMkldnnFusePassTest_Mul_Clip(
 ):
     def set_params(self):
         self.operand = paddle.multiply
-        self.act = fluid.layers.clip
+        self.act = paddle.clip
         self.act_alpha = 0.0
         self.act_beta = 10.0
 

python/paddle/fluid/tests/unittests/ir/inference/test_trt_activation_pass.py

@@ -106,7 +106,7 @@ class TensorRTSubgraphPassHardSwishPluginTest(
 
 class TensorRTSubgraphPassClipTest(TensorRTSubgraphPassActivationTest):
     def append_act(self, x):
-        return fluid.layers.clip(x, 0, 1)
+        return paddle.clip(x, 0, 1)
 
 
 class TensorRTSubgraphPassTanhTest(TensorRTSubgraphPassActivationTest):

python/paddle/fluid/tests/unittests/npu/test_clip_op_npu.py

@@ -117,13 +117,13 @@ class TestClipOpError(unittest.TestCase):
             input_data = np.random.random((2, 4)).astype("float32")
 
             def test_Variable():
-                fluid.layers.clip(x=input_data, min=-1.0, max=1.0)
+                paddle.clip(x=input_data, min=-1.0, max=1.0)
 
             self.assertRaises(TypeError, test_Variable)
 
             def test_dtype():
                 x2 = fluid.layers.data(name='x2', shape=[1], dtype='int32')
-                fluid.layers.clip(x=x2, min=-1.0, max=1.0)
+                paddle.clip(x=x2, min=-1.0, max=1.0)
 
             self.assertRaises(TypeError, test_dtype)
         paddle.disable_static()

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

@@ -686,7 +686,7 @@ class TestAdamOpV2(unittest.TestCase):
         value = np.arange(26).reshape(2, 13).astype("float32")
         a = fluid.dygraph.to_variable(value)
         linear = paddle.nn.Linear(13, 5)
-        clip = fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0)
+        clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0)
         adam = paddle.optimizer.Adam(
             0.1, parameters=linear.parameters(), grad_clip=clip
         )

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

@@ -20,12 +20,13 @@ from op_test import OpTest
 import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
+from paddle.nn import clip
 
 
 class TestClipByNormOp(OpTest):
     def setUp(self):
         self.max_relative_error = 0.006
-        self.python_api = fluid.layers.clip_by_norm
+        self.python_api = clip.clip_by_norm
         self.init_dtype()
         self.initTestCase()
         input = np.random.random(self.shape).astype(self.dtype)

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

@@ -128,15 +128,9 @@ class TestClipOpError(unittest.TestCase):
             input_data = np.random.random((2, 4)).astype("float32")
 
             def test_Variable():
-                fluid.layers.clip(x=input_data, min=-1.0, max=1.0)
+                paddle.clip(x=input_data, min=-1.0, max=1.0)
 
             self.assertRaises(TypeError, test_Variable)
-
-            def test_dtype():
-                x2 = fluid.layers.data(name='x2', shape=[1], dtype='int32')
-                fluid.layers.clip(x=x2, min=-1.0, max=1.0)
-
-            self.assertRaises(TypeError, test_dtype)
         paddle.disable_static()
 
 

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

@@ -584,7 +584,7 @@ class TestL2Decay(TranspilerTest):
         def filter(param):
             return param.name == "fc_w"
 
-        clip = fluid.clip.GradientClipByValue(0.1, need_clip=filter)
+        clip = paddle.nn.ClipGradByValue(0.1, need_clip=filter)
         sgd_optimizer.minimize(avg_cost, grad_clip=clip)
 
     def transpiler_test_impl(self):

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

@@ -504,8 +504,8 @@ class PaddingRNNTestBase(unittest.TestCase):
                     self.feed_order,
                 ) = res_vars
 
-                fluid.clip.set_gradient_clip(
-                    clip=fluid.clip.GradientClipByGlobalNorm(
+                paddle.nn.clip.set_gradient_clip(
+                    clip=paddle.nn.ClipGradByGlobalNorm(
                         clip_norm=config.max_grad_norm
                     )
                 )

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

@@ -64,7 +64,7 @@ class TestFleetExecutor(unittest.TestCase):
             )
             opt = paddle.optimizer.AdamW(
                 learning_rate=lr_val,
-                grad_clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0),
+                grad_clip=paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0),
             )
             opt.minimize(loss)
         # TODO: section_program will be removed in the future

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

@@ -64,7 +64,7 @@ class TestFleetExecutor(unittest.TestCase):
             )
             opt = paddle.optimizer.AdamW(
                 learning_rate=lr_val,
-                grad_clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0),
+                grad_clip=paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0),
             )
             opt.minimize(loss)
         # TODO: section_program will be removed in the future

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

@@ -47,7 +47,7 @@ class TestFleetExecutor(unittest.TestCase):
             )
             opt = paddle.optimizer.AdamW(
                 learning_rate=lr_val,
-                grad_clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0),
+                grad_clip=paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0),
             )
             opt.minimize(loss)
         # TODO: section_program will be removed in the future

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

@@ -20,6 +20,7 @@ import paddle.fluid as fluid
 import paddle.fluid.core as core
 from paddle.fluid import Program, program_guard
 from paddle.fluid.op import Operator
+from paddle.nn import clip
 
 
 class TestGetTensorFromSelectedRowsError(unittest.TestCase):
@@ -31,12 +32,12 @@ class TestGetTensorFromSelectedRowsError(unittest.TestCase):
             x_data = np.random.random((2, 4)).astype("float32")
 
             def test_Variable():
-                fluid.layers.get_tensor_from_selected_rows(x=x_data)
+                clip.get_tensor_from_selected_rows(x=x_data)
 
             self.assertRaises(TypeError, test_Variable)
 
             def test_SELECTED_ROWS():
-                fluid.layers.get_tensor_from_selected_rows(x=x_var)
+                clip.get_tensor_from_selected_rows(x=x_var)
 
             self.assertRaises(TypeError, test_SELECTED_ROWS)
 

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

@@ -17,12 +17,8 @@ import unittest
 import numpy as np
 
 import paddle.fluid as fluid
-from paddle.fluid.clip import (
-    GradientClipByGlobalNorm,
-    GradientClipByNorm,
-    GradientClipByValue,
-)
 from paddle.fluid.dygraph.base import to_variable
+from paddle.nn import ClipGradByGlobalNorm, ClipGradByNorm, ClipGradByValue
 
 
 class TestGradClipByGlobalNorm(unittest.TestCase):
@@ -67,7 +63,7 @@ class TestGradClipByGlobalNorm(unittest.TestCase):
     def get_dygrap_global_norm_result(self):
         with fluid.dygraph.guard():
 
-            gloabl_norm_clip = GradientClipByGlobalNorm(self.max_global_norm)
+            gloabl_norm_clip = ClipGradByGlobalNorm(self.max_global_norm)
             p_g_var = []
             for p, g in self.para_and_grad:
                 new_p = to_variable(p)
@@ -142,7 +138,7 @@ class TestGradClipByNorm(unittest.TestCase):
     def get_dygrap_norm_result(self):
         with fluid.dygraph.guard():
 
-            norm_clip = GradientClipByNorm(self.max_norm)
+            norm_clip = ClipGradByNorm(self.max_norm)
             p_g_var = []
             for p, g in self.para_and_grad:
                 new_p = to_variable(p)
@@ -212,9 +208,7 @@ class TestGradClipByValue(unittest.TestCase):
 
     def get_dygrap_clip_result(self):
         with fluid.dygraph.guard():
-            value_clip = GradientClipByValue(
-                max=self.max_value, min=self.min_value
-            )
+            value_clip = ClipGradByValue(max=self.max_value, min=self.min_value)
             p_g_var = []
             for p, g in self.para_and_grad:
                 new_p = to_variable(p)

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

@@ -20,7 +20,7 @@ from fake_reader import fake_imdb_reader
 import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
-from paddle.fluid.clip import _allow_pure_fp16_global_norm_clip
+from paddle.nn.clip import _allow_pure_fp16_global_norm_clip
 
 paddle.enable_static()
 
@@ -173,9 +173,9 @@ class TestGradientClipByGlobalNorm(TestGradientClip):
     # test whether the output is right when use 'set_gradient_clip'
     def test_old_gradient_clip(self):
         def func(params_grads):
-            clip = fluid.clip.GradientClipByGlobalNorm(clip_norm=self.clip_norm)
-            fluid.clip.set_gradient_clip(clip)
-            return fluid.clip.append_gradient_clip_ops(params_grads)
+            clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=self.clip_norm)
+            paddle.nn.clip.set_gradient_clip(clip)
+            return paddle.nn.clip.append_gradient_clip_ops(params_grads)
 
         self.clip_gradient = func
         self.check_gradient_clip(fluid.CPUPlace())
@@ -183,7 +183,7 @@ class TestGradientClipByGlobalNorm(TestGradientClip):
     # test whether the output is right when use grad_clip
     def test_new_gradient_clip(self):
         def func(params_grads):
-            clip = fluid.clip.GradientClipByGlobalNorm(clip_norm=self.clip_norm)
+            clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=self.clip_norm)
             return clip(params_grads)
 
         self.clip_gradient = func
@@ -192,7 +192,7 @@ class TestGradientClipByGlobalNorm(TestGradientClip):
     # test whether the output is right when use grad_clip under float64
     def test_new_gradient_clip_fp64(self):
         def func(params_grads):
-            clip = fluid.clip.GradientClipByGlobalNorm(clip_norm=self.clip_norm)
+            clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=self.clip_norm)
             return clip(params_grads)
 
         self.clip_gradient = func
@@ -201,15 +201,15 @@ class TestGradientClipByGlobalNorm(TestGradientClip):
     # invoke 'set_gradient_clip' in a wrong order
     def test_wrong_API_order(self):
         def backward_func(cost):
-            clip = fluid.clip.GradientClipByGlobalNorm(clip_norm=5.0)
-            fluid.clip.set_gradient_clip(clip)
+            clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=5.0)
+            paddle.nn.clip.set_gradient_clip(clip)
             sgd_optimizer = fluid.optimizer.SGD(
                 learning_rate=0.01, grad_clip=clip
             )
             # if 'set_gradient_clip' and 'optimize(grad_clip)' together, 'set_gradient_clip' will be ineffective
             sgd_optimizer.minimize(cost)
             # 'set_gradient_clip' must before 'minimize', otherwise, 'set_gradient_clip' will be ineffective
-            fluid.clip.set_gradient_clip(clip)
+            paddle.nn.clip.set_gradient_clip(clip)
 
         self.backward_and_optimize = backward_func
         for place in self.get_places():
@@ -269,7 +269,7 @@ class TestGradientClipByGlobalNorm(TestGradientClip):
         with fluid.program_guard(
             main_program=prog, startup_program=startup_program
         ):
-            clip = fluid.clip.GradientClipByGlobalNorm(self.clip_norm)
+            clip = paddle.nn.ClipGradByGlobalNorm(self.clip_norm)
             x = (
                 fluid.default_main_program()
                 .global_block()
@@ -313,7 +313,7 @@ class TestGradientClipByNorm(TestGradientClip):
     # test whether the output is right when use grad_clip
     def test_gradient_clip(self):
         def func(params_grads):
-            clip = fluid.clip.GradientClipByNorm(clip_norm=self.clip_norm)
+            clip = paddle.nn.ClipGradByNorm(clip_norm=self.clip_norm)
             return clip(params_grads)
 
         self.clip_gradient = func
@@ -321,7 +321,7 @@ class TestGradientClipByNorm(TestGradientClip):
 
     # if grad is None or not need clip
     def test_none_grad(self):
-        clip = fluid.clip.GradientClipByNorm(self.clip_norm)
+        clip = paddle.nn.ClipGradByNorm(self.clip_norm)
         x = (
             fluid.default_main_program()
             .global_block()
@@ -371,7 +371,7 @@ class TestGradientClipByValue(TestGradientClip):
     # test whether the output is right when use grad_clip
     def test_gradient_clip(self):
         def func(params_grads):
-            clip = fluid.clip.GradientClipByValue(max=self.max, min=self.min)
+            clip = paddle.nn.ClipGradByValue(max=self.max, min=self.min)
             return clip(params_grads)
 
         self.clip_gradient = func
@@ -379,7 +379,7 @@ class TestGradientClipByValue(TestGradientClip):
 
     # if grad is None or not need clip
     def test_none_grad(self):
-        clip = fluid.clip.GradientClipByValue(self.max, self.min)
+        clip = paddle.nn.ClipGradByValue(self.max, self.min)
         x = (
             fluid.default_main_program()
             .global_block()
@@ -419,7 +419,7 @@ class TestDygraphGradientClip(unittest.TestCase):
             sgd_optimizer = fluid.optimizer.SGD(
                 learning_rate=0.0,
                 parameter_list=linear.parameters(),
-                grad_clip=fluid.clip.GradientClipByGlobalNorm(0.1),
+                grad_clip=paddle.nn.ClipGradByGlobalNorm(0.1),
             )
             self.check_clip_result(loss, sgd_optimizer)
 
@@ -430,12 +430,8 @@ class TestDygraphGradientClip(unittest.TestCase):
 class TestDygraphGradientClipByGlobalNorm(TestDygraphGradientClip):
     def setUp(self):
         self.clip_norm = 0.8
-        self.clip1 = fluid.clip.GradientClipByGlobalNorm(
-            clip_norm=self.clip_norm
-        )
-        self.clip2 = fluid.clip.GradientClipByGlobalNorm(
-            clip_norm=self.clip_norm
-        )
+        self.clip1 = paddle.nn.ClipGradByGlobalNorm(clip_norm=self.clip_norm)
+        self.clip2 = paddle.nn.ClipGradByGlobalNorm(clip_norm=self.clip_norm)
 
     def check_clip_result(self, loss, optimizer):
         # if grad is None
@@ -476,7 +472,7 @@ class TestDygraphGradientClipByGlobalNorm(TestDygraphGradientClip):
 class TestDygraphGradientClipByNorm(TestDygraphGradientClip):
     def setUp(self):
         self.clip_norm = 0.8
-        self.clip = fluid.clip.GradientClipByNorm(clip_norm=self.clip_norm)
+        self.clip = paddle.nn.ClipGradByNorm(clip_norm=self.clip_norm)
 
     def check_clip_result(self, loss, optimizer):
         # if grad is None
@@ -506,7 +502,7 @@ class TestDygraphGradientClipByValue(TestDygraphGradientClip):
     def setUp(self):
         self.max = 0.2
         self.min = 0.1
-        self.clip = fluid.clip.GradientClipByValue(max=self.max, min=self.min)
+        self.clip = paddle.nn.ClipGradByValue(max=self.max, min=self.min)
 
     def check_clip_result(self, loss, optimizer):
         # if grad is None
@@ -572,7 +568,7 @@ class TestDygraphGradientClipFP16(unittest.TestCase):
                         params_grads.append((param, param._grad_ivar()))
                 _, grads = zip(*params_grads)
                 # clip grads
-                clip = fluid.clip.GradientClipByGlobalNorm(clip_norm=0.8)
+                clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=0.8)
                 params_grads = clip(params_grads)
                 _, grads_clip = zip(*params_grads)
                 # param update
@@ -616,7 +612,7 @@ class TestDygraphGradientClipFP64(unittest.TestCase):
                     params_grads.append((param, param._grad_ivar()))
             _, grads = zip(*params_grads)
             # clip grads
-            clip = fluid.clip.GradientClipByGlobalNorm(clip_norm=0.1)
+            clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=0.1)
             params_grads = clip(params_grads)
             _, grads_clip = zip(*params_grads)
 

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

@@ -361,7 +361,7 @@ class TestImperativeAutoPrune(unittest.TestCase):
         place = fluid.CPUPlace()
         with fluid.dygraph.guard(place):
             model = MyLayer(size, vocab_size, size)
-            grad_clip = fluid.clip.GradientClipByGlobalNorm(0.001)
+            grad_clip = paddle.nn.ClipGradByGlobalNorm(0.001)
             optimizer = fluid.optimizer.AdamOptimizer(
                 0.001, parameter_list=model.parameters(), grad_clip=grad_clip
             )
@@ -380,7 +380,7 @@ class TestImperativeAutoPrune(unittest.TestCase):
 
         with fluid.dygraph.guard(place):
             model = MyLayer2(size, vocab_size, size)
-            grad_clip = fluid.clip.GradientClipByGlobalNorm(0.001)
+            grad_clip = paddle.nn.ClipGradByGlobalNorm(0.001)
             optimizer = fluid.optimizer.AdamOptimizer(
                 0.001, parameter_list=model.parameters(), grad_clip=grad_clip
             )

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

@@ -52,7 +52,7 @@ class TestSimpleNet(unittest.TestCase):
                     fluid.set_flags(
                         {'FLAGS_sort_sum_gradient': sort_sum_gradient}
                     )
-                    # grad_clip = fluid.clip.GradientClipByGlobalNorm(5.0)
+                    # grad_clip = paddle.nn.ClipGradByGlobalNorm(5.0)
 
                     input_word = np.array([[1, 2], [2, 1]]).astype('int64')
                     input = paddle.to_tensor(input_word)
@@ -91,7 +91,7 @@ class TestSimpleNet(unittest.TestCase):
                     fluid.set_flags(
                         {'FLAGS_sort_sum_gradient': sort_sum_gradient}
                     )
-                    grad_clip = fluid.clip.GradientClipByGlobalNorm(5.0)
+                    grad_clip = paddle.nn.ClipGradByGlobalNorm(5.0)
 
                     input_word = np.array([[1, 2], [2, 1]]).astype('int64')
                     input = to_variable(input_word)

python/paddle/fluid/tests/unittests/xpu/test_clip_op_xpu.py

@@ -131,13 +131,13 @@ class TestClipOpError(unittest.TestCase):
             input_data = np.random.random((2, 4)).astype("float32")
 
             def test_Variable():
-                fluid.layers.clip(x=input_data, min=-1.0, max=1.0)
+                paddle.clip(x=input_data, min=-1.0, max=1.0)
 
             self.assertRaises(TypeError, test_Variable)
 
             def test_dtype():
                 x2 = fluid.layers.data(name='x2', shape=[1], dtype='int32')
-                fluid.layers.clip(x=x2, min=-1.0, max=1.0)
+                paddle.clip(x=x2, min=-1.0, max=1.0)
 
             self.assertRaises(TypeError, test_dtype)
         paddle.disable_static()

python/paddle/hapi/model.py

@@ -1535,7 +1535,7 @@ class Model:
                 assert isinstance(
                     self._optimizer._grad_clip,
                     (paddle.nn.ClipGradByGlobalNorm, paddle.nn.ClipGradByNorm),
-                ), "Only GradientClipByNorm and GradientClipByGlobalNorm are supported in amp training with level=O2 currently."
+                ), "Only ClipGradByNorm and ClipGradByGlobalNorm are supported in amp training with level=O2 currently."
 
         self._adapter._amp_custom_lists = {}
         self._adapter._amp_configs = {}

python/paddle/incubate/distributed/models/moe/grad_clip.py

@@ -15,13 +15,14 @@
 import paddle
 import paddle.distributed as dist
 from paddle.fluid import core, layers
-from paddle.fluid.clip import ClipGradBase, _squared_l2_norm
 from paddle.fluid.dygraph import base as imperative_base
+from paddle.nn import clip
+from paddle.nn.clip import ClipGradBase, _squared_l2_norm
 
 
 class ClipGradForMOEByGlobalNorm(ClipGradBase):
     r"""
-    The Algrithm is the same as paddle.fluid.clip.ClipGradByGlobalNorm
+    The Algrithm is the same as paddle.nn.ClipGradByGlobalNorm
     Given a list of Tensor :math:`t\_list` , calculate the global norm for the elements of all tensors in
     :math:`t\_list` , and limit it to ``clip_norm`` .
 
@@ -113,8 +114,8 @@ class ClipGradForMOEByGlobalNorm(ClipGradBase):
                 continue
             merge_grad = g
             if g.type == core.VarDesc.VarType.SELECTED_ROWS:
-                merge_grad = layers.merge_selected_rows(g)
-                merge_grad = layers.get_tensor_from_selected_rows(merge_grad)
+                merge_grad = clip.merge_selected_rows(g)
+                merge_grad = clip.get_tensor_from_selected_rows(merge_grad)
             sum_square = _squared_l2_norm(merge_grad)
             if sum_square.dtype == core.VarDesc.VarType.FP16:
                 sum_square_list_fp16.append(sum_square)

python/paddle/incubate/optimizer/distributed_fused_lamb.py

@@ -16,11 +16,11 @@ import os
 
 import paddle
 from paddle.fluid import core, framework, unique_name
-from paddle.fluid.clip import ClipGradByGlobalNorm
 from paddle.fluid.executor import global_scope
 from paddle.fluid.framework import Variable, name_scope
 from paddle.fluid.layer_helper import LayerHelper
 from paddle.fluid.optimizer import Optimizer
+from paddle.nn import ClipGradByGlobalNorm
 
 
 def init_communicator(block, rank, ranks, ring_id):

python/paddle/optimizer/adamw.py

@@ -20,10 +20,10 @@ import paddle
 
 from .. import _C_ops
 from ..fluid import core, framework, unique_name
-from ..fluid.clip import GradientClipBase
 from ..fluid.dygraph import base as imperative_base
 from ..fluid.framework import Parameter, Variable
 from ..fluid.layer_helper import LayerHelper
+from ..nn.clip import GradientClipBase
 from .lr import LRScheduler
 from .optimizer import Optimizer
 

python/paddle/optimizer/optimizer.py

@@ -18,6 +18,7 @@ from collections import defaultdict
 import numpy as np
 
 import paddle
+import paddle.autograd as imperative_base
 from paddle import _C_ops
 from paddle.fluid import core
 from paddle.fluid.framework import (
@@ -32,12 +33,6 @@ from paddle.fluid.framework import (
 
 from ..fluid import framework, unique_name
 from ..fluid.backward import _get_no_grad_set_name, append_backward
-from ..fluid.clip import (
-    GradientClipBase,
-    append_gradient_clip_ops,
-    error_clip_callback,
-)
-from ..fluid.dygraph import base as imperative_base
 from ..fluid.framework import Parameter, program_guard
 from ..fluid.initializer import Constant
 from ..fluid.layer_helper import LayerHelper
@@ -168,7 +163,7 @@ class Optimizer:
 
     """
 
-    @imperative_base.no_grad
+    @imperative_base.no_grad()
     def __init__(
         self,
         learning_rate,
@@ -225,7 +220,7 @@ class Optimizer:
                 % type(learning_rate)
             )
         if grad_clip is not None:
-            if not isinstance(grad_clip, GradientClipBase):
+            if not isinstance(grad_clip, paddle.nn.clip.GradientClipBase):
                 raise TypeError(
                     "'grad_clip' should be an instance of GradientClipBase's derived class"
                 )
@@ -1042,7 +1037,7 @@ class Optimizer:
                     params_grads.append((parameter_list[index], grad))
         else:
             if callbacks is None:
-                callbacks = [error_clip_callback]
+                callbacks = [paddle.nn.clip.error_clip_callback]
             else:
                 assert isinstance(callbacks, list)
             program = loss.block.program
@@ -1103,7 +1098,7 @@ class Optimizer:
             params_grads = self._grad_clip(params_grads)
         else:
 
-            params_grads = append_gradient_clip_ops(params_grads)
+            params_grads = paddle.nn.clip.append_gradient_clip_ops(params_grads)
 
         # Add regularization if any
         params_grads = self.append_regularization_ops(
@@ -1317,7 +1312,7 @@ class Optimizer:
         else:
             core.clear_gradients(param_list, set_to_zero)
 
-    @imperative_base.no_grad
+    @imperative_base.no_grad()
     def minimize(
         self, loss, startup_program=None, parameters=None, no_grad_set=None
     ):
@@ -1380,7 +1375,7 @@ class Optimizer:
 
         return optimize_ops, params_grads
 
-    @imperative_base.no_grad
+    @imperative_base.no_grad()
     @framework.dygraph_only
     def step(self):
         """