【PaddlePaddle Hackathon 2】11新增 API index_fill

paddle/fluid/operators/index_fill_op.cc

@@ -0,0 +1,124 @@
+/* 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 <memory>
+#include "paddle/fluid/framework/infershape_utils.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/op_version_registry.h"
+#include "paddle/phi/core/infermeta_utils.h"
+#include "paddle/phi/infermeta/unary.h"
+
+namespace paddle {
+namespace operators {
+
+class IndexFillOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        OperatorWithKernel::IndicateVarDataType(ctx, "X"), ctx.GetPlace());
+  }
+};
+
+class IndexFillOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor, default input Tensor<float>), "
+             "the input feature data of IndexFillOp, dtype should be"
+             "bool, int32, int64, float16, float32, float64.");
+    AddInput("IndexTensor",
+             "(Tensor, optional) If provided, index fill will use this."
+             "It has the highest priority of IndexTensor and attr(index).")
+        .AsDispensable();
+    AddInput("AxisTensor",
+             "(Tensor) If provided, use this as "
+             "axis, this has a higher priority than "
+             "attr(axis), the shape of this tensor MUST BE 1.")
+        .AsDispensable();
+    AddAttr<std::vector<int>>(
+        "index",
+        "(list<int>) Starting indices of corresponding axis in `axes`");
+    AddAttr<int>("axis", "(int), the dimension in which we index.");
+    AddAttr<float>("fill_value", "(float) The value to be filled.");
+    AddOutput("Out",
+              "(Tensor, default Tensor<float>),"
+              " the output of  IndexFillOp, whose dtype is the same as X.");
+    AddComment(R"DOC(
+                IndexFill operator
+                Fills the elements of the input tensor with value
+                by selecting the indices in the order given in index.
+
+                This operator also supports inplace modification.
+        )DOC");
+  }
+};
+
+template <typename T>
+class IndexFillGradMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("index_fill_grad");
+    if (this->HasInput("AxisTensor")) {
+      op->SetInput("AxisTensor", this->Input("AxisTensor"));
+    }
+    if (this->HasInput("IndexTensor")) {
+      op->SetInput("IndexTensor", this->Input("IndexTensor"));
+    }
+    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+    op->SetAttrMap(this->Attrs());
+  }
+};
+
+class IndexFillGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
+                                       ctx, framework::GradVarName("Out")),
+                                   ctx.GetPlace());
+  }
+};
+
+DECLARE_INPLACE_OP_INFERER(IndexFillInplaceInferer, {"X", "Out"});
+DECLARE_INPLACE_OP_INFERER(IndexFillGradInplaceInferer,
+                           {framework::GradVarName("Out"),
+                            framework::GradVarName("X")});
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+DECLARE_INFER_SHAPE_FUNCTOR(index_fill, IndexFillInferShapeFunctor,
+                            PD_INFER_META(phi::IndexFillInferMeta));
+
+REGISTER_OPERATOR(index_fill, ops::IndexFillOp, ops::IndexFillOpMaker,
+                  ops::IndexFillGradMaker<paddle::framework::OpDesc>,
+                  ops::IndexFillGradMaker<paddle::imperative::OpBase>,
+                  ops::IndexFillInplaceInferer, IndexFillInferShapeFunctor);
+
+DECLARE_INFER_SHAPE_FUNCTOR(index_fill_grad, IndexFillGradInferShapeFunctor,
+                            PD_INFER_META(phi::IndexFillGradInferMeta));
+
+REGISTER_OPERATOR(index_fill_grad, ops::IndexFillGradOp,
+                  ops::IndexFillGradInplaceInferer,
+                  IndexFillGradInferShapeFunctor);

paddle/fluid/operators/index_fill_tensor_op.cc

@@ -0,0 +1,133 @@
+/* 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 <memory>
+#include "paddle/fluid/framework/infershape_utils.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/op_version_registry.h"
+#include "paddle/phi/core/infermeta_utils.h"
+#include "paddle/phi/infermeta/binary.h"
+
+namespace paddle {
+namespace operators {
+
+class IndexFillTensorOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        OperatorWithKernel::IndicateVarDataType(ctx, "X"), ctx.GetPlace());
+  }
+};
+
+class IndexFillTensorOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor), "
+             "the input feature data of IndexFillTensorOp, dtype should be"
+             "bool, int32, int64, float16, float32, float64.");
+    AddInput("FillValue",
+             "(Tensor>), "
+             "the input tensor of IndexFillTensorOp, dtype should be"
+             "the same as input tensor X");
+    AddInput("IndexTensor",
+             "(Tensor, optional) If provided, index fill will use this."
+             "It has the highest priority of IndexTensor and attr(index).")
+        .AsDispensable();
+    AddInput("AxisTensor",
+             "(Tensor) If provided, use this as "
+             "axis, this has a higher priority than "
+             "attr(axis), the shape of this tensor MUST BE 1.")
+        .AsDispensable();
+    AddAttr<std::vector<int>>(
+        "index",
+        "(list<int>) Starting indices of corresponding axis in `axes`");
+    AddAttr<int>("axis", "(int), the dimension in which we index.");
+    AddOutput(
+        "Out",
+        "(Tensor, default Tensor<float>),"
+        " the output of  IndexFillTensorOp, whose dtype is the same as X.");
+    AddComment(R"DOC(
+                IndexFillTensor operator
+                Fills the elements of the input tensor with value
+                by selecting the indices in the order given in index.
+
+                This operator also supports inplace modification.
+        )DOC");
+  }
+};
+
+template <typename T>
+class IndexFillTensorGradMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("index_fill_tensor_grad");
+    if (this->HasInput("AxisTensor")) {
+      op->SetInput("AxisTensor", this->Input("AxisTensor"));
+    }
+    if (this->HasInput("IndexTensor")) {
+      op->SetInput("IndexTensor", this->Input("IndexTensor"));
+    }
+    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+    op->SetOutput(framework::GradVarName("FillValue"),
+                  this->InputGrad("FillValue"));
+    op->SetAttrMap(this->Attrs());
+  }
+};
+
+class IndexFillTensorGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
+                                       ctx, framework::GradVarName("Out")),
+                                   ctx.GetPlace());
+  }
+};
+
+DECLARE_INPLACE_OP_INFERER(IndexFillTensorInplaceInferer, {"X", "Out"});
+DECLARE_INPLACE_OP_INFERER(IndexFillTensorGradInplaceInferer,
+                           {framework::GradVarName("Out"),
+                            framework::GradVarName("X")});
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+DECLARE_INFER_SHAPE_FUNCTOR(index_fill_tensor, IndexFillTensorInferShapeFunctor,
+                            PD_INFER_META(phi::IndexFillTensorInferMeta));
+
+REGISTER_OPERATOR(index_fill_tensor, ops::IndexFillTensorOp,
+                  ops::IndexFillTensorOpMaker,
+                  ops::IndexFillTensorGradMaker<paddle::framework::OpDesc>,
+                  ops::IndexFillTensorGradMaker<paddle::imperative::OpBase>,
+                  ops::IndexFillTensorInplaceInferer,
+                  IndexFillTensorInferShapeFunctor);
+
+DECLARE_INFER_SHAPE_FUNCTOR(index_fill_tensor_grad,
+                            IndexFillTensorGradInferShapeFunctor,
+                            PD_INFER_META(phi::IndexFillTensorGradInferMeta));
+
+REGISTER_OPERATOR(index_fill_tensor_grad, ops::IndexFillTensorGradOp,
+                  ops::IndexFillTensorGradInplaceInferer,
+                  IndexFillTensorGradInferShapeFunctor);

paddle/phi/infermeta/binary.cc

@@ -1181,6 +1181,61 @@ void HuberLossInferMeta(const MetaTensor& input,
   out->share_lod(input);
 }
 
+void IndexFillTensorInferMeta(const MetaTensor& x,
+                              const MetaTensor& fill_tensor,
+                              const IntArray& index_arr,
+                              const Scalar& axis_scalar,
+                              MetaTensor* output) {
+  auto input_dim = x.dims();
+  auto index = index_arr.GetData();
+  auto axis = axis_scalar.to<int>();
+
+  PADDLE_ENFORCE_EQ(
+      fill_tensor.dtype() == x.dtype(),
+      true,
+      phi::errors::InvalidArgument(
+          "The dtype of fill_tensor should be same as input tensor."));
+
+  PADDLE_ENFORCE_EQ(
+      axis < input_dim.size() && axis >= (0 - input_dim.size()),
+      true,
+      phi::errors::OutOfRange(
+          "Axis is out of range, It's expected "
+          "to be in range of [-%d, %d). But received Attr(dim) = %d.",
+          input_dim.size(),
+          input_dim.size() - 1,
+          axis));
+
+  PADDLE_ENFORCE_EQ(
+      index.size() > 0,
+      true,
+      phi::errors::InvalidArgument("The index array should not be empty."));
+
+  output->set_dims(x.dims());
+  output->set_dtype(x.dtype());
+  output->set_layout(x.layout());
+  output->share_lod(x);
+}
+
+void IndexFillTensorGradInferMeta(const MetaTensor& out_grad,
+                                  const IntArray& index_arr,
+                                  const Scalar& axis_scalar,
+                                  MetaTensor* x_grad,
+                                  MetaTensor* fill_tensor_grad) {
+  auto do_dims = out_grad.dims();
+  if (x_grad) {
+    x_grad->set_dims(do_dims);
+    x_grad->set_dtype(out_grad.dtype());
+    x_grad->set_layout(out_grad.layout());
+    x_grad->share_lod(out_grad);
+  }
+
+  if (fill_tensor_grad) {
+    fill_tensor_grad->set_dims(phi::make_ddim({1}));
+    fill_tensor_grad->set_dtype(out_grad.dtype());
+  }
+}
+
 void IndexSampleInferMeta(const MetaTensor& x,
                           const MetaTensor& y,
                           MetaTensor* out,

paddle/phi/infermeta/binary.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/phi/common/int_array.h"
 #include "paddle/phi/common/scalar.h"
 #include "paddle/phi/core/meta_tensor.h"
 
@@ -184,6 +185,18 @@ void HuberLossInferMeta(const MetaTensor& input_meta,
                         MetaTensor* residual,
                         MetaConfig config = MetaConfig());
 
+void IndexFillTensorInferMeta(const MetaTensor& x,
+                              const MetaTensor& fill_tensor,
+                              const IntArray& index_arr,
+                              const Scalar& axis_scalar,
+                              MetaTensor* output);
+
+void IndexFillTensorGradInferMeta(const MetaTensor& out_grad,
+                                  const IntArray& index_arr,
+                                  const Scalar& axis_scalar,
+                                  MetaTensor* x_grad,
+                                  MetaTensor* fill_tensor_grad);
+
 void IndexSampleInferMeta(const MetaTensor& x,
                           const MetaTensor& y,
                           MetaTensor* out,

paddle/phi/infermeta/unary.cc

@@ -702,6 +702,48 @@ void IncrementInferMeta(const MetaTensor& x, float value, MetaTensor* out) {
   out->set_dtype(x.dtype());
 }
 
+void IndexFillInferMeta(const MetaTensor& x,
+                        const IntArray& index_arr,
+                        const Scalar& axis_scalar,
+                        float fill_value,
+                        MetaTensor* output) {
+  auto input_dim = x.dims();
+  auto index = index_arr.GetData();
+  auto axis = axis_scalar.to<int>();
+
+  PADDLE_ENFORCE_EQ(
+      axis < input_dim.size() && axis >= (0 - input_dim.size()),
+      true,
+      phi::errors::OutOfRange(
+          "Axis is out of range, It's expected "
+          "to be in range of [-%d, %d). But received Attr(dim) = %d.",
+          input_dim.size(),
+          input_dim.size() - 1,
+          axis));
+
+  PADDLE_ENFORCE_EQ(
+      index.size() > 0,
+      true,
+      phi::errors::InvalidArgument("The index array should not be empty."));
+
+  output->set_dims(x.dims());
+  output->set_dtype(x.dtype());
+  output->set_layout(x.layout());
+  output->share_lod(x);
+}
+
+void IndexFillGradInferMeta(const MetaTensor& out_grad,
+                            const IntArray& index_arr,
+                            const Scalar& axis_scalar,
+                            float fill_value,
+                            MetaTensor* x_grad) {
+  auto do_dims = out_grad.dims();
+  x_grad->set_dims(do_dims);
+  x_grad->set_dtype(out_grad.dtype());
+  x_grad->set_layout(out_grad.layout());
+  x_grad->share_lod(out_grad);
+}
+
 static phi::DDim ValidateShape(const std::vector<int64_t> shape,
                                const phi::DDim& in_dims) {
   const int64_t in_size = phi::product(in_dims);