new prelu with functor

paddle/operators/prelu_op.cc

@@ -27,13 +27,14 @@ class PReluOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
     auto *in = ctx.Input<framework::Tensor>("X");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     out->Resize(in->dims());
   }
 };
 
-// template <typename AttrType>
+template <typename AttrType>
 class PReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   PReluOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
@@ -43,10 +44,12 @@ class PReluOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(PRelu operator
 
 The equation is:
-f(x) = alpha * x , for x < 0
+
-f(x) = x         , for x >= 0
+  f(x) = alpha * x , for x < 0
+  f(x) = x         , for x >= 0
+
 )DOC");
-    AddAttr<float>("alpha", "The scaling factor alpha of prelu.")
+    AddAttr<AttrType>("alpha", "The scaling factor alpha of prelu.")
         .SetDefault(0.0);
   }
 };
@@ -59,6 +62,8 @@ class PReluGradOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
     auto *X_grad =
         ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
     auto *X = ctx.Input<framework::Tensor>("X");
@@ -72,7 +77,7 @@ class PReluGradOp : public framework::OperatorWithKernel {
 
 namespace ops = paddle::operators;
 
-REGISTER_OP(prelu, ops::PReluOp, ops::PReluOpMaker, prelu_grad,
+REGISTER_OP(prelu, ops::PReluOp, ops::PReluOpMaker<float>, prelu_grad,
             ops::PReluGradOp);
 REGISTER_OP_CPU_KERNEL(prelu,
                        ops::PReluKernel<paddle::platform::CPUPlace, float>);

paddle/operators/prelu_op.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/platform/transform.h"
 
 namespace paddle {
 namespace operators {
@@ -23,28 +24,60 @@ using Tensor = framework::Tensor;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+using platform::Transform;
 
-template <typename Place, typename T>
+template <typename T>
+class Prelu_functor {
+ public:
+  explicit Prelu_functor(const T& alpha) : alpha_(alpha) {}
+
+  HOSTDEVICE T operator()(const T& X) const {
+    if (X > 0)
+      return X;
+    else
+      return X * alpha_;
+  }
+
+ private:
+  T alpha_;
+};
+
+template <typename Place, typename T, typename AttrType = T>
 class PReluKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto* X = context.Input<Tensor>("X");
     auto* Out = context.Output<Tensor>("Out");
 
-    Out->mutable_data<T>(context.GetPlace());
+    const T* X_ptr = X->data<T>();
+    T* O_ptr = Out->mutable_data<T>(context.GetPlace());
 
-    auto alpha = static_cast<T>(context.Attr<float>("alpha"));
+    auto alpha = static_cast<T>(context.Attr<AttrType>("alpha"));
 
-    auto X_vec = EigenVector<T>::Flatten(*X);
+    int numel = X->numel();
-    auto Out_vec = EigenVector<T>::Flatten(*Out);
 
-    // auto place = context.GetEigenDevice<Place>();
+    auto place = context.GetPlace();
-    // Out_vec.device(place)
+    Transform(place, X_ptr, X_ptr + numel, O_ptr, Prelu_functor<T>(alpha));
-    Out_vec = X_vec.cwiseMax(0.f) + X_vec.cwiseMin(0.f) * alpha;
   }
 };
 
-template <typename Place, typename T>
+template <typename T>
+class Prelu_Grad_functor {
+ public:
+  explicit Prelu_Grad_functor(const T& alpha) : alpha_(alpha) {}
+
+  HOSTDEVICE T operator()(const T& Out, const T& dOut) const {
+    if (Out > 0)
+      return dOut;
+    else
+      return dOut * alpha_;
+  }
+
+ private:
+  T alpha_;
+};
+
+template <typename Place, typename T, typename AttrType = T>
 class PReluGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -53,16 +86,16 @@ class PReluGradKernel : public framework::OpKernel {
 
     auto* Out = context.Input<Tensor>("Out");
 
-    auto alpha = static_cast<T>(context.Attr<float>("alpha"));
+    auto alpha = static_cast<T>(context.Attr<AttrType>("alpha"));
 
-    dX->mutable_data<T>(context.GetPlace());
+    T* dX_ptr = dX->mutable_data<T>(context.GetPlace());
-    for (int i = 0; i < dX->numel(); ++i) {
+    const T* dO_ptr = dO->data<T>();
-      if (Out->data<T>()[i] > 0) {
+    const T* O_ptr = Out->data<T>();
-        dX->data<T>()[i] = dO->data<T>()[i];
+    int numel = dX->numel();
-      } else {
+
-        dX->data<T>()[i] = dO->data<T>()[i] * alpha;
+    auto place = context.GetPlace();
-      }
+    Transform(place, O_ptr, O_ptr + numel, dO_ptr, dX_ptr,
-    }
+              Prelu_Grad_functor<T>(alpha));
   }
 };
 

python/paddle/v2/framework/tests/test_prelu_op.py

@@ -6,7 +6,7 @@ from op_test import OpTest
 class PreluTest(OpTest):
     def setUp(self):
         self.op_type = "prelu"
-        self.inputs = {'X': np.random.normal(size=(3, 5)).astype("float32")}
+        self.inputs = {'X': np.random.normal(size=(10, 10)).astype("float32")}
         self.attrs = {'alpha': 0.1}
         out_np = np.maximum(self.inputs['X'], 0.)
         out_np = out_np + np.minimum(self.inputs['X'], 0.) * self.attrs['alpha']