PaddingRNN model memory optimize

paddle/fluid/operators/cross_entropy_op.cc

@@ -13,18 +13,21 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/cross_entropy_op.h"
+#include <memory>
 #include <string>
+#include <unordered_map>
 
 namespace paddle {
 namespace operators {
 
-class CrossEntropyOp : public framework::OperatorWithKernel {
+class CrossEntropyOpBase : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
+
     PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null.");
 
     auto x_dims = ctx->GetInputDim("X");
@@ -43,7 +46,8 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
                         "Input(X) and Input(Label) shall have the same shape "
                         "except the last dimension.");
     }
-    if (ctx->Attrs().Get<bool>("soft_label")) {
+
+    if (IsSoftLabel(ctx)) {
       if (check) {
         PADDLE_ENFORCE_EQ(x_dims[rank - 1], label_dims[rank - 1],
                           "If Attr(soft_label) == true, the last dimension of "
@@ -69,21 +73,24 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
     return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
                                    ctx.device_context());
   }
+
+  virtual bool IsSoftLabel(framework::InferShapeContext* ctx) const {
+    return ctx->Attrs().Get<bool>("soft_label");
+  }
 };
 
-class CrossEntropyGradientOp : public framework::OperatorWithKernel {
+class CrossEntropyGradientOpBase : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
-  void InferShape(framework::InferShapeContext* ctx) const override {
-    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
+  void InferShape(framework::InferShapeContext* ctx) const {
     PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
                    "Input(Y@GRAD) shoudl be not null.");
     PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
                    "Output(X@GRAD) should be not null.");
 
-    auto x_dims = ctx->GetInputDim("X");
+    auto x_dims = GetXDim(ctx);
     auto label_dims = ctx->GetInputDim("Label");
     auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y"));
     int rank = x_dims.size();
@@ -108,9 +115,7 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
                         "The Input(X) and Input(Y@Grad) should have the same "
                         "shape except the last dimension.");
     }
-    PADDLE_ENFORCE_EQ(dy_dims[rank - 1], 1,
-                      "The last dimension of Input(Y@Grad) should be 1.");
-    if (ctx->Attrs().Get<bool>("soft_label")) {
+    if (IsSoftLabel(ctx)) {
       if (check) {
         PADDLE_ENFORCE_EQ(
             x_dims[rank - 1], label_dims[rank - 1],
@@ -123,16 +128,39 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
                         "Input(Label) should be 1.");
     }
     ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
-    ctx->ShareLoD("X", framework::GradVarName("X"));
+    PADDLE_ENFORCE_EQ(dy_dims[rank - 1], 1,
+                      "The last dimension of Input(Y@Grad) should be 1.");
+    ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
+    ctx->ShareLoD(VarNameWithXLoD(), framework::GradVarName("X"));
   }
 
  protected:
   // Explicitly set that the data type of computation kernel of cross_entropy
   // is determined by its input "X".
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
-    return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
-                                   ctx.device_context());
+    return framework::OpKernelType(
+        ctx.Input<Tensor>(framework::GradVarName("Y"))->type(),
+        ctx.device_context());
+  }
+
+  virtual framework::DDim GetXDim(framework::InferShapeContext* ctx) const {
+    return ctx->GetInputDim("X");
+  }
+
+  virtual const char* VarNameWithXLoD() const { return "X"; }
+
+  virtual bool IsSoftLabel(framework::InferShapeContext* ctx) const {
+    return ctx->Attrs().Get<bool>("soft_label");
+  }
+};
+
+class CrossEntropyOpInferVarType
+    : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Y"}};
   }
 };
 
@@ -200,26 +228,137 @@ or not. But the output only shares the LoD information with input X.
   }
 };
 
-class CrossEntropyOpInferVarType
-    : public framework::PassInDtypeAndVarTypeToOutput {
+class CrossEntropyGradientOp : public CrossEntropyGradientOpBase {
+ public:
+  using CrossEntropyGradientOpBase::CrossEntropyGradientOpBase;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
+    CrossEntropyGradientOpBase::InferShape(ctx);
+  }
+};
+
+class CrossEntropyOp2 : public CrossEntropyOpBase {
+ public:
+  using CrossEntropyOpBase::CrossEntropyOpBase;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    CrossEntropyOpBase::InferShape(ctx);
+
+    PADDLE_ENFORCE(ctx->HasOutput("XShape"),
+                   "Output(XShape) should be not null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto x_dims_vec = framework::vectorize(x_dims);
+    x_dims_vec.push_back(0);
+    ctx->SetOutputDim("XShape", framework::make_ddim(x_dims_vec));
+    ctx->ShareLoD("X", /*->*/ "XShape");
+  }
+
  protected:
-  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
-      const override {
-    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Y"}};
+  bool IsSoftLabel(framework::InferShapeContext* ctx) const override {
+    return false;
+  }
+};
+
+class CrossEntropyGradientOp2 : public CrossEntropyGradientOpBase {
+ public:
+  using CrossEntropyGradientOpBase::CrossEntropyGradientOpBase;
+
+ protected:
+  virtual framework::DDim GetXDim(framework::InferShapeContext* ctx) const {
+    auto x_shape = ctx->GetInputDim("XShape");
+    return framework::DDim(x_shape.Get(), x_shape.size() - 1);
+  }
+
+  virtual const char* VarNameWithXLoD() const { return "XShape"; }
+
+  virtual bool IsSoftLabel(framework::InferShapeContext* ctx) const {
+    return false;
+  }
+};
+
+class CrossEntropyOpMaker2 : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor, default Tensor<float>), a tensor whose last dimension "
+             "size is equal to the number of classes. This input is a "
+             "probability computed by the previous operator, which is almost "
+             "always the result of a softmax operator.");
+    AddInput(
+        "Label",
+        "(Tensor), the tensor which represents the ground truth. It has the "
+        "same shape with 'X' except the last dimension. One hot Tensor.");
+    AddOutput("Y",
+              "(Tensor, default Tensor<float>), a tensor whose shape is same "
+              "with 'X' except that the last dimension size is 1. It "
+              "represents the cross entropy loss.");
+    AddOutput("XShape", "Temporaily variable to save shape and LoD of X.");
+    AddAttr<int>("ignore_index",
+                 "(int, default -100), Specifies a target value that is"
+                 "ignored and does not contribute to the input gradient."
+                 "Only valid if soft_label is set to False")
+        .SetDefault(-100);
+    AddComment(R"DOC(
+Hard-label CrossEntropy Operator.
+
+The input 'X' and 'Label' will first be logically flattened to 2-D matrixs. 
+The matrix's second dimension(row length) is as same as the original last 
+dimension, and the first dimension(column length) is the product of all other 
+original dimensions. Then the softmax computation will take palce on each raw 
+of flattened matrixs.
+
+Only support hard label.
+
+Both the input X and Label can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD information with input X.
+
+)DOC");
+  }
+};
+
+class CrossEntropyGradOpDescMaker2 : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    std::unique_ptr<framework::OpDesc> op(new framework::OpDesc());
+    op->SetType("cross_entropy_grad2");
+    op->SetInput("Label", Input("Label"));
+    op->SetInput("Y", Output("Y"));
+    op->SetInput("XShape", Output("XShape"));
+    op->SetInput(framework::GradVarName("Y"), OutputGrad("Y"));
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    op->SetAttrMap(Attrs());
+    return op;
   }
 };
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 using CPUCtx = paddle::platform::CPUDeviceContext;
 
-REGISTER_OPERATOR(cross_entropy, ops::CrossEntropyOp, ops::CrossEntropyOpMaker,
-                  ops::CrossEntropyOpInferVarType,
+REGISTER_OPERATOR(cross_entropy, ops::CrossEntropyOpBase,
+                  ops::CrossEntropyOpMaker, ops::CrossEntropyOpInferVarType,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(cross_entropy_grad, ops::CrossEntropyGradientOp);
 REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<CPUCtx, float>,
                        ops::CrossEntropyOpKernel<CPUCtx, double>);
 REGISTER_OP_CPU_KERNEL(cross_entropy_grad,
                        ops::CrossEntropyGradientOpKernel<CPUCtx, float>,
                        ops::CrossEntropyGradientOpKernel<CPUCtx, double>);
+
+REGISTER_OPERATOR(cross_entropy2, ops::CrossEntropyOp2,
+                  ops::CrossEntropyOpMaker2, ops::CrossEntropyOpInferVarType,
+                  ops::CrossEntropyGradOpDescMaker2);
+REGISTER_OPERATOR(cross_entropy_grad2, ops::CrossEntropyGradientOp2);
+REGISTER_OP_CPU_KERNEL(cross_entropy2,
+                       ops::CrossEntropyOpKernel2<CPUCtx, float>,
+                       ops::CrossEntropyOpKernel2<CPUCtx, double>);
+REGISTER_OP_CPU_KERNEL(cross_entropy_grad2,
+                       ops::CrossEntropyGradientOpKernel2<CPUCtx, float>,
+                       ops::CrossEntropyGradientOpKernel2<CPUCtx, double>);

paddle/fluid/operators/cross_entropy_op.cu

@@ -27,3 +27,13 @@ REGISTER_OP_CUDA_KERNEL(
     cross_entropy_grad, ops::CrossEntropyGradientOpKernel<CUDACtx, float>,
     ops::CrossEntropyGradientOpKernel<CUDACtx, double>,
     ops::CrossEntropyGradientOpKernel<CUDACtx, plat::float16>);
+
+REGISTER_OP_CUDA_KERNEL(cross_entropy2,
+                        ops::CrossEntropyOpKernel2<CUDACtx, float>,
+                        ops::CrossEntropyOpKernel2<CUDACtx, double>,
+                        ops::CrossEntropyOpKernel2<CUDACtx, plat::float16>);
+
+REGISTER_OP_CUDA_KERNEL(
+    cross_entropy_grad2, ops::CrossEntropyGradientOpKernel2<CUDACtx, float>,
+    ops::CrossEntropyGradientOpKernel2<CUDACtx, double>,
+    ops::CrossEntropyGradientOpKernel2<CUDACtx, plat::float16>);

paddle/fluid/operators/cross_entropy_op.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math.h"
 #include "paddle/fluid/operators/math/cross_entropy.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/platform/for_range.h"
@@ -137,5 +138,85 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename T>
+struct HardLabelCrossEntropyBackwardFunctor {
+  HardLabelCrossEntropyBackwardFunctor(T* dx, const T* y, const T* dy,
+                                       const int64_t* label,
+                                       int64_t ignore_index,
+                                       int64_t feature_size)
+      : dx_(dx),
+        y_(y),
+        dy_(dy),
+        label_(label),
+        ignore_index_(ignore_index),
+        feature_size_(feature_size) {}
+
+  HOSTDEVICE void operator()(int64_t idx) const {
+    auto row_idx = idx / feature_size_;
+    auto col_idx = idx % feature_size_;
+    auto label = label_[row_idx];
+    if (label == col_idx && label != ignore_index_) {
+      dx_[idx] = -dy_[row_idx] * real_exp(y_[row_idx]);
+    } else {
+      dx_[idx] = 0;
+    }
+  }
+
+  T* dx_;
+  const T* y_;
+  const T* dy_;
+  const int64_t* label_;
+  int64_t ignore_index_;
+  int64_t feature_size_;
+};
+
+template <typename DeviceContext, typename T>
+class CrossEntropyOpKernel2 : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x_original = ctx.Input<Tensor>("X");
+    int rank = x_original->dims().size();
+
+    auto x = framework::ReshapeToMatrix(*x_original, rank - 1);
+    auto label =
+        framework::ReshapeToMatrix(*ctx.Input<Tensor>("Label"), rank - 1);
+    auto* y = ctx.Output<Tensor>("Y");
+    y->mutable_data<T>(ctx.GetPlace());
+
+    auto ignore_index = ctx.Attr<int>("ignore_index");
+
+    math::CrossEntropyFunctor<DeviceContext, T>()(
+        ctx.template device_context<DeviceContext>(), y, &x, &label, false,
+        ignore_index);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class CrossEntropyGradientOpKernel2 : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    auto* label = ctx.Input<Tensor>("Label");
+
+    auto* p_dx = dx->mutable_data<T>(ctx.GetPlace());
+    auto* p_y = y->data<T>();
+    auto* p_dy = dy->data<T>();
+    auto* p_label = label->data<int64_t>();
+
+    int64_t ignore_index = ctx.Attr<int>("ignore_index");
+    int rank = dx->dims().size();
+    int64_t feature_size = dx->dims()[rank - 1];
+    int64_t batch_size = framework::product(dx->dims()) / feature_size;
+
+    platform::ForRange<DeviceContext> for_range(
+        ctx.template device_context<DeviceContext>(),
+        batch_size * feature_size);
+    for_range(HardLabelCrossEntropyBackwardFunctor<T>(
+        p_dx, p_y, p_dy, p_label, ignore_index, feature_size));
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle