Merge pull request #3181 from qingqing01/sigmoid_bp

Sigmoid backward implementation.

paddle/operators/sigmoid_op.cc

@@ -37,10 +37,8 @@ class SigmoidOpMaker : public OpProtoAndCheckerMaker {
 
 class SigmoidOpGrad : public OperatorWithKernel {
  protected:
-  void InferShape(const InferShapeContext &ctx) const override {}
-  std::string DebugString() const override {
-    LOG(INFO) << "SigmoidGrad";
-    return "";
+  void InferShape(const InferShapeContext &ctx) const override {
+    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
   }
 };
 
@@ -51,3 +49,5 @@ REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker);
 REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, ops::SigmoidOpGrad);
 
 REGISTER_OP_CPU_KERNEL(sigmoid, ops::SigmoidKernel<ops::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(sigmoid_grad,
+                       ops::SigmoidGradKernel<ops::CPUPlace, float>);

paddle/operators/sigmoid_op.cu

@@ -16,3 +16,5 @@
 #include "paddle/operators/sigmoid_op.h"
 
 REGISTER_OP_GPU_KERNEL(sigmoid, ops::SigmoidKernel<ops::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(sigmoid_grad,
+                       ops::SigmoidGradKernel<ops::GPUPlace, float>);

paddle/operators/sigmoid_op.h

@@ -27,6 +27,7 @@ class SigmoidKernel : public OpKernel {
     auto output = context.Output<Tensor>(0);
     output->mutable_data<T>(context.GetPlace());
 
+    // The clipping is used in Paddle's raw implenmention
     auto X = EigenVector<T>::Flatten(*input);
     auto Y = EigenVector<T>::Flatten(*output);
     auto place = context.GetEigenDevice<Place>();
@@ -34,5 +35,23 @@ class SigmoidKernel : public OpKernel {
     Y.device(place) = 1.0 / (1.0 + (-1.0 * X).exp());
   }
 };
+
+template <typename Place, typename T>
+class SigmoidGradKernel : public OpKernel {
+ public:
+  void Compute(const ExecutionContext& context) const override {
+    auto Y_t = context.Input<Tensor>("Y");
+    auto dY_t = context.Input<Tensor>(framework::GradVarName("Y"));
+    auto dX_t = context.Output<Tensor>(framework::GradVarName("X"));
+
+    dX_t->mutable_data<T>(context.GetPlace());
+
+    auto dX = EigenVector<T>::Flatten(*dX_t);
+    auto Y = EigenVector<T>::Flatten(*Y_t);
+    auto dY = EigenVector<T>::Flatten(*dY_t);
+    dX.device(context.GetEigenDevice<Place>()) = dY * Y * (1. - Y);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle

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

@@ -12,5 +12,8 @@ class TestSigmoidOp(unittest.TestCase):
         self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))}
 
 
+#class TestSigmoidGradOp(unittest.TestCase):
+#TODO(qingqing) add unit test
+
 if __name__ == '__main__':
     unittest.main()