【PaddlePaddle Hackathon 2】11新增 API index_fill 【错误分支请删除】

paddle/fluid/operators/index_fill_op.cc

@@ -0,0 +1,115 @@
+/*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 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"
+             "int32, int64, float16, float32, float64.");
+    AddInput("Index",
+             "(Tensor, default 1-d Tensor<int>), "
+             "the 1-D tensor containing the indices to index, "
+             "dtype should be int32, int64");
+    AddAttr<int>("axis",
+                 "(int, default 0), "
+                 "the dimension in which we index.")
+        .SetDefault(0);
+    AddAttr<float>("fill_value",
+                   "(float, default 0.0f) The value to be filled.")
+        .SetDefault(0.0f);
+    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");
+    op->SetInput("Index", this->Input("Index"));
+    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")});
+DECLARE_NO_NEED_BUFFER_VARS_INFERER(IndexFillGradNoNeedBufferVarsInferer, "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,
+                  ops::IndexFillGradNoNeedBufferVarsInferer,
+                  IndexFillGradInferShapeFunctor);

paddle/fluid/operators/nanmedian_op.cc

@@ -0,0 +1,120 @@
+/*Copyright (c) 2019 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 NanmedianOp : 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 NanmedianOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor, default Tensor<float>), "
+             "the input feature data of NanmedianOp, dtype should be"
+             "int32, int64, float16, float32, float64.");
+    AddAttr<bool>(
+        "ignore_nan",
+        "(bool, default true) Set to true if nan values should be ignored. "
+        "Set to false when no nan value in x were considered. ")
+        .SetDefault(true);
+    AddOutput("Medians",
+              "The calculation differs in the odd or even of the valid "
+              "elements amount."
+              "Along the axis, two elements contributed to the median value in "
+              "each row."
+              "If the amount of valid elements were even, both were the same.")
+        .AsIntermediate()
+        .AsExtra();
+    AddOutput("Out",
+              "(Tensor, default Tensor<float>),"
+              " the output of  NanmedianOp, whose dtype is the same as X");
+    AddComment(R"DOC(
+                Nanmedian operator
+
+                This operator is considered as an extention of median operation,
+                which supports specifically the case of nan values in the input.
+
+                If all the elements in input are NaN it will also return NaN.
+                If no elements in input are Nan, this op is identical to thie median op.
+
+                This operator can also supports multiple axis,
+                and could be switched to median operator when `ignore_nan` were set to False.
+        )DOC");
+  }
+};
+
+template <typename T>
+class NanmedianGradMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("nanmedian_grad");
+    op->SetInput("X", this->Input("X"));
+    op->SetInput("Medians", this->Output("Medians"));
+    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+  }
+};
+
+class NanmedianGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "nanmedian");
+    OP_INOUT_CHECK(ctx->HasInput("Medians"), "Input", "Medians", "nanmedian");
+    OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input",
+                   framework::GradVarName("Out"), "nanmedian");
+
+    auto x_dims = ctx->GetInputDim("X");
+    ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
+                                       ctx, framework::GradVarName("Out")),
+                                   ctx.GetPlace());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+DECLARE_INFER_SHAPE_FUNCTOR(nanmedian, NanmedianInferShapeFunctor,
+                            PD_INFER_META(phi::NanmedianInferMeta));
+
+REGISTER_OPERATOR(nanmedian, ops::NanmedianOp, ops::NanmedianOpMaker,
+                  ops::NanmedianGradMaker<paddle::framework::OpDesc>,
+                  ops::NanmedianGradMaker<paddle::imperative::OpBase>,
+                  NanmedianInferShapeFunctor);
+
+REGISTER_OPERATOR(nanmedian_grad, ops::NanmedianGradOp);

paddle/phi/infermeta/binary.cc

@@ -1181,6 +1181,57 @@ void HuberLossInferMeta(const MetaTensor& input,
   out->share_lod(input);
 }
 
+void IndexFillInferMeta(const MetaTensor& x,
+                        const MetaTensor& index,
+                        int axis,
+                        float fill_value,
+                        MetaTensor* output) {
+  auto input_dim = x.dims();
+  auto index_dim = index.dims();
+
+  PADDLE_ENFORCE_EQ(
+      axis < input_dim.size() && axis >= (0 - input_dim.size()),
+      true,
+      phi::errors::OutOfRange(
+          "Attr(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_dim.size() == 1 || (index_dim.size() == 2 && index_dim[1] == 1),
+      true,
+      phi::errors::InvalidArgument(
+          "The 'shape' of Input(Index) must be 1-D tensor. "
+          "But received: the 'shape' of Input(Index) is [%s], "
+          "the dimension of Input(Index) is [%d].",
+          index_dim,
+          index_dim.size()));
+
+  PADDLE_ENFORCE_EQ(
+      index_dim[0] != 0,
+      true,
+      phi::errors::InvalidArgument("The length of Input(Index) can't be 0."));
+
+  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 MetaTensor& index,
+                            int axis,
+                            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);
+}
+
 void IndexSampleInferMeta(const MetaTensor& x,
                           const MetaTensor& y,
                           MetaTensor* out,

paddle/phi/infermeta/binary.h

@@ -184,6 +184,18 @@ void HuberLossInferMeta(const MetaTensor& input_meta,
                         MetaTensor* residual,
                         MetaConfig config = MetaConfig());
 
+void IndexFillInferMeta(const MetaTensor& x,
+                        const MetaTensor& index,
+                        int axis,
+                        float fill_value,
+                        MetaTensor* output);
+
+void IndexFillGradInferMeta(const MetaTensor& out_grad,
+                            const MetaTensor& index,
+                            int axis,
+                            float fill_value,
+                            MetaTensor* x_grad);
+
 void IndexSampleInferMeta(const MetaTensor& x,
                           const MetaTensor& y,
                           MetaTensor* out,

paddle/phi/infermeta/unary.cc

@@ -1243,6 +1243,30 @@ void MultinomialInferMeta(const MetaTensor& x,
   out->set_dtype(DataType::INT64);
 }
 
+void NanmedianInferMeta(const MetaTensor& x,
+                        bool ignore_nan,
+                        MetaTensor* out,
+                        MetaTensor* medians) {
+  auto x_dim = x.dims();
+  int64_t x_rank = x_dim.size();
+
+  std::vector<int64_t> out_dims(x_rank);
+  std::vector<int64_t> median_dims(x_rank);
+  for (int64_t i = 0; i < x_rank - 1; i++) {
+    out_dims[i] = x_dim[i];
+    median_dims[i] = x_dim[i];
+  }
+
+  out_dims[x_rank - 1] = 1;
+  median_dims[x_rank - 1] = 2;
+
+  out->set_dims(make_ddim(out_dims));
+  out->set_dtype(x.dtype());
+
+  medians->set_dims(make_ddim(median_dims));
+  medians->set_dtype(x.dtype());
+}
+
 void NormInferMeta(const MetaTensor& x,
                    int axis,
                    float epsilon,

paddle/phi/infermeta/unary.h

@@ -177,6 +177,12 @@ void MultinomialInferMeta(const MetaTensor& x,
                           int num_samples,
                           bool replacement,
                           MetaTensor* out);
+
+void NanmedianInferMeta(const MetaTensor& x,
+                        bool ignore_nan,
+                        MetaTensor* out,
+                        MetaTensor* medians);
+
 void NormInferMeta(const MetaTensor& x,
                    int axis,
                    float epsilon,