[Phi] Migrate lable_smooth_op into Phi

paddle/fluid/framework/operator.cc

@@ -2037,7 +2037,7 @@ void OperatorWithKernel::BuildPtenKernelContext(
         (i == 0 ? 0 : pt_kernel_context->InputRangeAt(i - 1).second);
 
     // deal with optional here
-    if ((it == ctx.inputs.end()) &&
+    if ((it == ctx.inputs.end() || it->second.size() == 0) &&
         (input_defs[i].type_index ==
          std::type_index(typeid(paddle::optional<const phi::DenseTensor&>)))) {
       pt_kernel_context->EmplaceBackInputWithoutSetRange(nullptr);

paddle/fluid/operators/label_smooth_op.cc

@@ -12,9 +12,8 @@ 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/label_smooth_op.h"
-
 #include <string>
+#include "paddle/fluid/framework/op_registry.h"
 
 namespace paddle {
 namespace framework {
@@ -152,11 +151,3 @@ REGISTER_OPERATOR(label_smooth, ops::LabelSmoothOp, ops::LabelSmoothOpMaker,
                   ops::LabelSmoothGradMaker<paddle::framework::OpDesc>,
                   ops::LabelSmoothGradMaker<paddle::imperative::OpBase>);
 REGISTER_OPERATOR(label_smooth_grad, ops::LabelSmoothGradOp);
-REGISTER_OP_CPU_KERNEL(
-    label_smooth,
-    ops::LabelSmoothKernel<paddle::platform::CPUDeviceContext, float>,
-    ops::LabelSmoothKernel<paddle::platform::CPUDeviceContext, double>);
-REGISTER_OP_CPU_KERNEL(
-    label_smooth_grad,
-    ops::LabelSmoothGradKernel<paddle::platform::CPUDeviceContext, float>,
-    ops::LabelSmoothGradKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/label_smooth_op_npu.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/operators/label_smooth_op.h"
+#include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/platform/device/npu/npu_op_runner.h"
 
 namespace paddle {

paddle/fluid/operators/label_smooth_op_xpu.cc

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #ifdef PADDLE_WITH_XPU
-#include "paddle/fluid/operators/label_smooth_op.h"
 #include "paddle/fluid/framework/op_registry.h"
 
 namespace paddle {

paddle/phi/kernels/cpu/label_smooth_grad_kernel.cc

@@ -0,0 +1,45 @@
+// Copyright (c) 2022 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/label_smooth_grad_kernel.h"
+#include "paddle/phi/backends/cpu/cpu_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/funcs/eigen/common.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void LabelSmoothGradKernel(const Context& ctx,
+                           const DenseTensor& out_grad,
+                           float epsilon,
+                           DenseTensor* label_grad) {
+  ctx.template Alloc<T>(label_grad);
+  auto d_out_dim = out_grad.dims()[out_grad.dims().size() - 1];
+  if (d_out_dim != 0) {
+    auto d_out = EigenVector<T>::Flatten(out_grad);
+    auto d_in = EigenVector<T>::Flatten(*label_grad);
+
+    auto& dev = *ctx.eigen_device();
+    d_in.device(dev) = static_cast<T>(1 - epsilon) * d_out;
+  }
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(label_smooth_grad,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::LabelSmoothGradKernel,
+                   float,
+                   double) {}

paddle/phi/kernels/cpu/label_smooth_kernel.cc

@@ -0,0 +1,50 @@
+// Copyright (c) 2022 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/label_smooth_kernel.h"
+#include "paddle/phi/backends/cpu/cpu_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/funcs/eigen/common.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void LabelSmoothKernel(const Context& ctx,
+                       const DenseTensor& label,
+                       paddle::optional<const DenseTensor&> prior_dist,
+                       float epsilon,
+                       DenseTensor* out) {
+  auto label_dim = label.dims()[label.dims().size() - 1];
+  ctx.template Alloc<T>(out);
+  auto& dev = *ctx.eigen_device();
+  if (label_dim != 0) {
+    auto eigen_out = EigenVector<T>::Flatten(*out);
+    auto eigen_in = EigenVector<T>::Flatten(label);
+    if (prior_dist.is_initialized()) {
+      auto dist = EigenVector<T>::Flatten(*prior_dist.get_ptr());
+      eigen_out.device(dev) =
+          static_cast<T>(1 - epsilon) * eigen_in +
+          static_cast<T>(epsilon) *
+              dist.broadcast(Eigen::DSizes<int, 1>(label.numel() / label_dim));
+    } else {
+      eigen_out.device(dev) = static_cast<T>(1 - epsilon) * eigen_in +
+                              static_cast<T>(epsilon / label_dim);
+    }
+  }
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(
+    label_smooth, CPU, ALL_LAYOUT, phi::LabelSmoothKernel, float, double) {}

paddle/phi/kernels/gpu/label_smooth_grad_kernel.cu

@@ -0,0 +1,55 @@
+// Copyright (c) 2022 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/fluid/operators/elementwise/elementwise_op_impl.cu.h"
+#include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/label_smooth_grad_kernel.h"
+
+namespace phi {
+template <typename T>
+struct LabelSmoothGradFunctor {
+  T epsilon;
+
+  __forceinline__ LabelSmoothGradFunctor(float epsilon_data) {
+    epsilon = static_cast<T>(epsilon_data);
+  }
+
+  __device__ __forceinline__ T operator()(const T x) const {
+    return static_cast<T>(1 - epsilon) * x;
+  }
+};
+
+template <typename T, typename Context>
+void LabelSmoothGradKernel(const Context& ctx,
+                           const DenseTensor& out_grad,
+                           float epsilon,
+                           DenseTensor* label_grad) {
+  ctx.template Alloc<T>(label_grad);
+
+  std::vector<const DenseTensor*> ins = {&out_grad};
+  std::vector<DenseTensor*> outs = {label_grad};
+  auto functor = LabelSmoothGradFunctor<T>(epsilon);
+  paddle::operators::LaunchSameDimsElementwiseCudaKernel<T>(
+      ctx, ins, &outs, functor);
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(label_smooth_grad,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::LabelSmoothGradKernel,
+                   float,
+                   double) {}