Implement slice grad operator. #8130 (#12330)

* Implement slice grad operator. #8130
* test slice grad operator and bug fix
* Fix pre commit style

paddle/fluid/operators/slice_op.cc

@@ -25,7 +25,7 @@ class SliceOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
-  void InferShape(framework::InferShapeContext *ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Input"),
                    "Input (Input) of slice op should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -58,7 +58,7 @@ class SliceOp : public framework::OperatorWithKernel {
 
  protected:
   framework::OpKernelType GetExpectedKernelType(
-      const framework::ExecutionContext &ctx) const override {
+      const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
         ctx.GetPlace());
@@ -119,15 +119,54 @@ Following examples will explain how slice works:
   }
 };
 
+class SliceOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"), "Input should not be null");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) should not be null");
+    auto x_dims = ctx->GetInputDim("Input");
+    auto x_grad_name = framework::GradVarName("Input");
+    if (ctx->HasOutput(x_grad_name)) {
+      ctx->SetOutputDim(x_grad_name, x_dims);
+    }
+  }
+};
+
+class SliceOpGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto* bind = new framework::OpDesc();
+    bind->SetInput("Input", Input("Input"));
+    bind->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
+    bind->SetOutput(framework::GradVarName("Input"), InputGrad("Input"));
+    bind->SetAttrMap(Attrs());
+    bind->SetType("slice_grad");
+    return std::unique_ptr<framework::OpDesc>(bind);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(slice, ops::SliceOp, ops::SliceOpMaker,
-                  paddle::framework::EmptyGradOpMaker);
+                  ops::SliceOpGradMaker);
+REGISTER_OPERATOR(slice_grad, ops::SliceOpGrad);
 
 REGISTER_OP_CPU_KERNEL(
     slice, ops::SliceKernel<paddle::platform::CPUDeviceContext, int>,
     ops::SliceKernel<paddle::platform::CPUDeviceContext, int64_t>,
     ops::SliceKernel<paddle::platform::CPUDeviceContext, float>,
     ops::SliceKernel<paddle::platform::CPUDeviceContext, double>);
+
+REGISTER_OP_CPU_KERNEL(
+    slice_grad, ops::SliceGradKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::SliceGradKernel<paddle::platform::CPUDeviceContext, int64_t>,
+    ops::SliceGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SliceGradKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/slice_op.cu

@@ -20,3 +20,10 @@ REGISTER_OP_CUDA_KERNEL(
     ops::SliceKernel<paddle::platform::CUDADeviceContext, double>,
     ops::SliceKernel<paddle::platform::CUDADeviceContext, int>,
     ops::SliceKernel<paddle::platform::CUDADeviceContext, int64_t>);
+
+REGISTER_OP_CUDA_KERNEL(
+    slice_grad,
+    ops::SliceGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SliceGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::SliceGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::SliceGradKernel<paddle::platform::CUDADeviceContext, int64_t>);

paddle/fluid/operators/slice_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 #include <algorithm>
+#include <utility>
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 
@@ -84,5 +85,79 @@ class SliceKernel : public framework::OpKernel<T> {
     out_t.device(place) = in_t.slice(offsets, extents);
   }
 };
+
+template <typename DeviceContext, typename T>
+class SliceGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    size_t rank = ctx.Input<framework::Tensor>(framework::GradVarName("Out"))
+                      ->dims()
+                      .size();
+    switch (rank) {
+      case 1:
+        SliceCompute<1>(ctx);
+        break;
+      case 2:
+        SliceCompute<2>(ctx);
+        break;
+      case 3:
+        SliceCompute<3>(ctx);
+        break;
+      case 4:
+        SliceCompute<4>(ctx);
+        break;
+      case 5:
+        SliceCompute<5>(ctx);
+        break;
+      case 6:
+        SliceCompute<6>(ctx);
+        break;
+    }
+  }
+
+ private:
+  template <size_t D>
+  void SliceCompute(const framework::ExecutionContext& context) const {
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
+    auto* d_out =
+        context.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* d_input =
+        context.Output<framework::Tensor>(framework::GradVarName("Input"));
+    d_input->mutable_data<T>(context.GetPlace());
+    auto out_dims = d_out->dims();
+    auto in_dims = d_input->dims();
+    auto axes = context.Attr<std::vector<int>>("axes");
+    auto starts = context.Attr<std::vector<int>>("starts");
+
+    auto offsets = Eigen::array<int, D>();
+    auto extents = Eigen::array<int, D>();
+    for (size_t i = 0; i < D; ++i) {
+      offsets[i] = 0;
+      extents[i] = out_dims[i];
+    }
+    int start;
+    for (size_t i = 0; i < axes.size(); ++i) {
+      start = starts[i];
+      if (start < 0) {
+        start = (start + in_dims[axes[i]]);
+      }
+      start = std::max(start, 0);
+      offsets[axes[i]] = start;
+    }
+    Eigen::array<std::pair<int, int>, D> paddings;
+    for (size_t i = 0; i < paddings.size(); ++i) {
+      paddings[i].first = offsets[i];
+      paddings[i].second = (in_dims[i] - out_dims[i]) - offsets[i];
+    }
+    auto d_in_t =
+        framework::EigenTensor<T, D, Eigen::RowMajor, Eigen::DenseIndex>::From(
+            *d_input);
+    auto d_out_t =
+        framework::EigenTensor<T, D, Eigen::RowMajor, Eigen::DenseIndex>::From(
+            *d_out);
+    d_in_t.device(place) = d_out_t.pad(paddings, 0);
+  }
+};
 }  // namespace operators
 }  // namespace paddle

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

@@ -41,6 +41,9 @@ class TestSliceOp(OpTest):
     def test_check_output(self):
         self.check_output()
 
+    def test_check_grad_normal(self):
+        self.check_grad(['Input'], 'Out', max_relative_error=0.006)
+
 
 class TestCase1(TestSliceOp):
     def config(self):