Add elementwise_pow_op backward implementation and the unit test codes of it. (#18848)

paddle/fluid/operators/elementwise/elementwise_pow_op.cc

 /* Copyright (c) 2018 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.
@@ -13,11 +10,30 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/elementwise/elementwise_pow_op.h"
+#include <memory>
 #include <string>
 #include "paddle/fluid/operators/elementwise/elementwise_op.h"
 
 namespace paddle {
 namespace operators {
+
+class ElementwisePowOpGradDescMaker : 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("elementwise_pow_grad");
+    op->SetInput("X", Input("X"));
+    op->SetInput("Y", Input("Y"));
+    op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
+    op->SetAttrMap(Attrs());
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    op->SetOutput(framework::GradVarName("Y"), InputGrad("Y"));
+    return op;
+  }
+};
 class ElementwisePowOpMaker : public ElementwiseOpMaker {
  protected:
   std::string GetName() const override { return "Pow"; }
@@ -27,9 +43,20 @@ class ElementwisePowOpMaker : public ElementwiseOpMaker {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(elementwise_pow, ops::ElementwiseOp,
-                             ops::ElementwisePowOpMaker);
+REGISTER_OPERATOR(elementwise_pow, ops::ElementwiseOp,
+                  ops::ElementwisePowOpMaker, ops::ElementwiseOpInferVarType,
+                  ops::ElementwisePowOpGradDescMaker);
+REGISTER_OPERATOR(elementwise_pow_grad, ops::ElementwiseOpGrad);
+
 REGISTER_OP_CPU_KERNEL(
     elementwise_pow,
     ops::ElementwisePowKernel<paddle::platform::CPUDeviceContext, float>,
-    ops::ElementwisePowKernel<paddle::platform::CPUDeviceContext, double>);
+    ops::ElementwisePowKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ElementwisePowKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ElementwisePowKernel<paddle::platform::CPUDeviceContext, int64_t>);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_pow_grad,
+    ops::ElementwisePowGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ElementwisePowGradKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ElementwisePowGradKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ElementwisePowGradKernel<paddle::platform::CPUDeviceContext, int64_t>);

paddle/fluid/operators/elementwise/elementwise_pow_op.cu

@@ -15,4 +15,13 @@ namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
     elementwise_pow,
     ops::ElementwisePowKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::ElementwisePowKernel<paddle::platform::CUDADeviceContext, double>);
+    ops::ElementwisePowKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwisePowKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwisePowKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    elementwise_pow_grad,
+    ops::ElementwisePowGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwisePowGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwisePowGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwisePowGradKernel<paddle::platform::CUDADeviceContext,
+                                  int64_t>);

paddle/fluid/operators/elementwise/elementwise_pow_op.h

 /* Copyright (c) 2018 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.
@@ -15,6 +12,7 @@ limitations under the License. */
 #pragma once
 
 #include <cmath>
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 
 namespace paddle {
@@ -29,9 +27,11 @@ template <typename DeviceContext, typename T>
 class ElementwisePowKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    using Tensor = framework::Tensor;
-
+    using Tensor = framework::LoDTensor;
     auto* x = ctx.Input<Tensor>("X");
+    PADDLE_ENFORCE(x != nullptr,
+                   "Cannot get input Variable X, variable name = %s",
+                   ctx.op().Input("X"));
     auto* y = ctx.Input<Tensor>("Y");
     auto* z = ctx.Output<Tensor>("Out");
     z->mutable_data<T>(ctx.GetPlace());
@@ -41,5 +41,36 @@ class ElementwisePowKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename T>
+struct PowGradDX {
+  HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
+    return dout * y * std::pow(x, y - 1);
+  }
+};
+
+template <typename T>
+struct PowGradDY {
+  HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
+    return dout * std::log(x) * std::pow(x, y);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class ElementwisePowGradKernel : public ElemwiseGradKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
+    using Tensor = framework::Tensor;
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* out = dout;
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    int axis = ctx.Attr<int>("axis");
+    ElemwiseGradCompute<DeviceContext, T, PowGradDX<T>, PowGradDY<T>>(
+        ctx, *x, *y, *out, *dout, axis, dx, dy, PowGradDX<T>(), PowGradDY<T>());
+  }
+};
 }  // namespace operators
 }  // namespace paddle

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

@@ -22,24 +22,87 @@ class TestElementwisePowOp(OpTest):
     def setUp(self):
         self.op_type = "elementwise_pow"
         self.inputs = {
-            'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
-            'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
+            'X': np.random.uniform(0.1, 1, [2, 3]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [2, 3]).astype("float32")
         }
         self.outputs = {'Out': np.power(self.inputs['X'], self.inputs['Y'])}
 
     def test_check_output(self):
         self.check_output()
 
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out')
+
 
 class TestElementwisePowOp_scalar(TestElementwisePowOp):
     def setUp(self):
         self.op_type = "elementwise_pow"
         self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype('float32'),
-            'Y': np.random.rand(1).astype('float32')
+            'X': np.random.uniform(0.1, 1, [3, 3, 4]).astype(np.float32),
+            'Y': np.random.uniform(0.1, 1, [1]).astype(np.float32)
+        }
+        self.outputs = {'Out': np.power(self.inputs['X'], self.inputs['Y'])}
+
+
+class TestElementwisePowOp_tensor(TestElementwisePowOp):
+    def setUp(self):
+        self.op_type = "elementwise_pow"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [32]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [32]).astype("float32")
+        }
+        self.outputs = {'Out': np.power(self.inputs['X'], self.inputs['Y'])}
+
+
+class TestElementwisePowOp_broadcast_0(TestElementwisePowOp):
+    def setUp(self):
+        self.op_type = "elementwise_pow"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 3, 4]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [4]).astype("float32")
         }
         self.outputs = {'Out': np.power(self.inputs['X'], self.inputs['Y'])}
 
 
+class TestElementwisePowOp_broadcast_1(TestElementwisePowOp):
+    def setUp(self):
+        self.op_type = "elementwise_pow"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 3, 4]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [3]).astype("float32")
+        }
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out': np.power(self.inputs['X'], self.inputs['Y'].reshape(3, 1))
+        }
+
+
+class TestElementwisePowOp_broadcast_2(TestElementwisePowOp):
+    def setUp(self):
+        self.op_type = "elementwise_pow"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 3, 4]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [2]).astype("float32")
+        }
+        self.attrs = {'axis': 0}
+        self.outputs = {
+            'Out': np.power(self.inputs['X'], self.inputs['Y'].reshape(2, 1, 1))
+        }
+
+
+class TestElementwisePowOp_broadcast_3(TestElementwisePowOp):
+    def setUp(self):
+        self.op_type = "elementwise_pow"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 3, 4, 5]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [3, 4]).astype("float32")
+        }
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out': np.power(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 4,
+                                                                       1))
+        }
+
+
 if __name__ == '__main__':
     unittest.main()