move `decayed_adagrad_op` to phi (#55995)

* move decayed_adagrad_op to phi

* fix bug

paddle/fluid/operators/optimizers/decayed_adagrad_op.cc

@@ -12,7 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/optimizers/decayed_adagrad_op.h"
+#include "paddle/fluid/framework/infershape_utils.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+#include "paddle/phi/infermeta/multiary.h"
 
 namespace paddle {
 namespace operators {
@@ -127,9 +130,12 @@ stability to avoid the division by zero error.
 }  // namespace paddle
 
 namespace ops = paddle::operators;
+
+DECLARE_INFER_SHAPE_FUNCTOR(decayed_adagrad,
+                            DecayedAdagradShapeFunctor,
+                            PD_INFER_META(phi::DecayedAdagradInferMeta));
+
 REGISTER_OP_WITHOUT_GRADIENT(decayed_adagrad,
                              ops::DecayedAdagradOp,
-                             ops::DecayedAdagradOpMaker);
-
-PD_REGISTER_STRUCT_KERNEL(
-    decayed_adagrad, CPU, ALL_LAYOUT, ops::DecayedAdagradOpKernel, float) {}
+                             ops::DecayedAdagradOpMaker,
+                             DecayedAdagradShapeFunctor);

paddle/fluid/operators/optimizers/decayed_adagrad_op.h

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#pragma once
-#include "paddle/fluid/framework/eigen.h"
-#include "paddle/fluid/framework/op_registry.h"
-
-namespace paddle {
-namespace operators {
-
-template <typename T, typename DeviceContext>
-class DecayedAdagradOpKernel : public framework::OpKernel<T> {
- public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    const auto* param_var = ctx.InputVar("Param");
-    PADDLE_ENFORCE_EQ(param_var->IsType<phi::DenseTensor>(),
-                      true,
-                      platform::errors::InvalidArgument(
-                          "The Var(%s)'s type should be phi::DenseTensor, "
-                          "but the received is %s",
-                          ctx.InputNames("Param").front(),
-                          framework::ToTypeName(param_var->Type())));
-    const auto* grad_var = ctx.InputVar("Grad");
-    PADDLE_ENFORCE_EQ(grad_var->IsType<phi::DenseTensor>(),
-                      true,
-                      platform::errors::InvalidArgument(
-                          "The Var(%s)'s type should be phi::DenseTensor, "
-                          "but the received is %s",
-                          ctx.InputNames("Grad").front(),
-                          framework::ToTypeName(grad_var->Type())));
-
-    auto param_out_tensor = ctx.Output<phi::DenseTensor>("ParamOut");
-    auto moment_out_tensor = ctx.Output<phi::DenseTensor>("MomentOut");
-
-    param_out_tensor->mutable_data<T>(ctx.GetPlace());
-    moment_out_tensor->mutable_data<T>(ctx.GetPlace());
-
-    float decay = ctx.Attr<float>("decay");
-    float epsilon = ctx.Attr<float>("epsilon");
-
-    auto param = framework::EigenVector<T>::Flatten(
-        *ctx.Input<phi::DenseTensor>("Param"));
-    auto grad = framework::EigenVector<T>::Flatten(
-        *ctx.Input<phi::DenseTensor>("Grad"));
-    auto moment = framework::EigenVector<T>::Flatten(
-        *ctx.Input<phi::DenseTensor>("Moment"));
-    auto lr = framework::EigenVector<T>::Flatten(
-        *ctx.Input<phi::DenseTensor>("LearningRate"));
-
-    auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
-    auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
-    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
-
-    moment_out.device(place) = decay * moment + (1 - decay) * grad * grad;
-    Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
-    param_out.device(place) =
-        param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
-  }
-};
-
-}  // namespace operators
-}  // namespace paddle

paddle/phi/infermeta/multiary.cc

@@ -1057,6 +1057,45 @@ void CudnnLSTMInferMeta(
   state_out->set_dtype(phi::DataType::UINT8);
 }
 
+void DecayedAdagradInferMeta(const MetaTensor& param,
+                             const MetaTensor& grad,
+                             const MetaTensor& moment,
+                             const MetaTensor& learning_rate,
+                             float decay,
+                             float epsilon,
+                             MetaTensor* param_out,
+                             MetaTensor* moment_out) {
+  auto lr_dims = learning_rate.dims();
+  PADDLE_ENFORCE_NE(phi::product(lr_dims),
+                    0,
+                    phi::errors::InvalidArgument(
+                        "Maybe the Input variable LearningRate has not "
+                        "been initialized. You may need to confirm "
+                        "if you put exe.run(startup_program) "
+                        "after optimizer.minimize function."));
+  PADDLE_ENFORCE_EQ(
+      phi::product(lr_dims),
+      1,
+      phi::errors::InvalidArgument("LearningRate should have one element"));
+  auto param_dims = param.dims();
+  PADDLE_ENFORCE_EQ(param_dims,
+                    grad.dims(),
+                    phi::errors::InvalidArgument(
+                        "Param and Grad input of DecayedAdagradOp should have "
+                        "the same dimension."));
+  PADDLE_ENFORCE_EQ(
+      param_dims,
+      moment.dims(),
+      phi::errors::InvalidArgument(
+          "Param and Moment input of DecayedAdagradOp should have "
+          "the same dimension."));
+
+  param_out->set_dims(param_dims);
+  param_out->set_dtype(param.dtype());
+  moment_out->set_dims(param_dims);
+  moment_out->set_dtype(param.dtype());
+}
+
 inline int ConvOutputSize(
     int input_size, int filter_size, int dilation, int padding, int stride) {
   const int dkernel = dilation * (filter_size - 1) + 1;

paddle/phi/infermeta/multiary.h

@@ -258,6 +258,15 @@ void CudnnLSTMInferMeta(
     MetaTensor* reserve,
     MetaTensor* state_out);
 
+void DecayedAdagradInferMeta(const MetaTensor& param,
+                             const MetaTensor& grad,
+                             const MetaTensor& moment,
+                             const MetaTensor& learning_rate,
+                             float decay,
+                             float epsilon,
+                             MetaTensor* param_out,
+                             MetaTensor* moment_out);
+
 void DeformableConvInferMeta(const MetaTensor& x,
                              const MetaTensor& offset,
                              const MetaTensor& filter,

paddle/phi/kernels/cpu/decayed_adagrad_kernel.cc

+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/phi/kernels/decayed_adagrad_kernel.h"
+#include "paddle/phi/backends/cpu/cpu_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/impl/decayed_adagrad_kernel_impl.h"
+
+PD_REGISTER_KERNEL(
+    decayed_adagrad, CPU, ALL_LAYOUT, phi::DecayedAdagradDenseKernel, float) {}

paddle/phi/kernels/decayed_adagrad_kernel.h

+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/phi/core/dense_tensor.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void DecayedAdagradDenseKernel(const Context& dev_ctx,
+                               const DenseTensor& param,
+                               const DenseTensor& grad,
+                               const DenseTensor& moment,
+                               const DenseTensor& learning_rate,
+                               float decay,
+                               float epsilon,
+                               DenseTensor* param_out,
+                               DenseTensor* moment_out);
+}  // namespace phi

paddle/phi/kernels/gpu/decayed_adagrad_kernel.cu

+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/phi/kernels/decayed_adagrad_kernel.h"
+#include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/impl/decayed_adagrad_kernel_impl.h"
+
+PD_REGISTER_KERNEL(
+    decayed_adagrad, GPU, ALL_LAYOUT, phi::DecayedAdagradDenseKernel, float) {}

paddle/phi/kernels/impl/decayed_adagrad_kernel_impl.h

+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/phi/kernels/decayed_adagrad_kernel.h"
+#include "paddle/phi/kernels/funcs/eigen/common.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void DecayedAdagradDenseKernel(const Context& dev_ctx,
+                               const DenseTensor& param_t,
+                               const DenseTensor& grad_t,
+                               const DenseTensor& moment_t,
+                               const DenseTensor& learning_rate,
+                               float decay,
+                               float epsilon,
+                               DenseTensor* param_out_t,
+                               DenseTensor* moment_out_t) {
+  dev_ctx.template Alloc<T>(param_out_t);
+  dev_ctx.template Alloc<T>(moment_out_t);
+
+  auto param = EigenVector<T>::Flatten(param_t);
+  auto grad = EigenVector<T>::Flatten(grad_t);
+  auto moment = EigenVector<T>::Flatten(moment_t);
+  auto lr = EigenVector<T>::Flatten(learning_rate);
+
+  auto param_out = EigenVector<T>::Flatten(*param_out_t);
+  auto moment_out = EigenVector<T>::Flatten(*moment_out_t);
+  auto& place = *dev_ctx.eigen_device();
+
+  moment_out.device(place) = decay * moment + (1 - decay) * grad * grad;
+  Eigen::DSizes<int, 1> m_dsize(moment_out_t->numel());
+  param_out.device(place) =
+      param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
+}
+}  // namespace phi

paddle/fluid/operators/optimizers/decayed_adagrad_op.cu → paddle/phi/ops/compat/decayed_adagrad_sig.cc

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
@@ -11,9 +11,20 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
-#include "paddle/fluid/operators/optimizers/decayed_adagrad_op.h"
 
-namespace ops = paddle::operators;
+#include "paddle/phi/core/compat/op_utils.h"
 
-PD_REGISTER_STRUCT_KERNEL(
-    decayed_adagrad, GPU, ALL_LAYOUT, ops::DecayedAdagradOpKernel, float) {}
+namespace phi {
+
+KernelSignature DecayedAdagradOpArgumentMapping(
+    const ArgumentMappingContext& ctx UNUSED) {
+  return KernelSignature("decayed_adagrad",
+                         {"Param", "Grad", "Moment", "LearningRate"},
+                         {"decay", "epsilon"},
+                         {"ParamOut", "MomentOut"});
+}
+
+}  // namespace phi
+
+PD_REGISTER_ARG_MAPPING_FN(decayed_adagrad,
+                           phi::DecayedAdagradOpArgumentMapping);

test/legacy_test/test_decayed_adagrad_op.py

@@ -46,7 +46,7 @@ class TestDecayedAdagradOp1(OpTest):
         self.outputs = {'ParamOut': param_out, 'MomentOut': moment_out}
 
     def test_check_output(self):
-        self.check_output()
+        self.check_output(check_dygraph=False)
 
 
 class TestDecayedAdagradOp2(OpTest):
@@ -77,7 +77,7 @@ class TestDecayedAdagradOp2(OpTest):
         self.outputs = {'ParamOut': param_out, 'MomentOut': moment_out}
 
     def test_check_output(self):
-        self.check_output()
+        self.check_output(check_dygraph=False)
 
 
 if __name__ == "__main__":