Merge pull request #3183 from dzhwinter/add_op_gradient

"add rowwise add backward op"

paddle/operators/rowwise_add_op.cc

@@ -17,7 +17,9 @@
 namespace paddle {
 namespace operators {
 
-class RowWiseAddOp : public framework::OperatorWithKernel {
+using framework::Tensor;
+
+class RowwiseAddOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
@@ -34,9 +36,9 @@ class RowWiseAddOp : public framework::OperatorWithKernel {
   }
 };
 
-class RowWiseAddOpMaker : public framework::OpProtoAndCheckerMaker {
+class RowwiseAddOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  RowWiseAddOpMaker(framework::OpProto *proto,
+  RowwiseAddOpMaker(framework::OpProto *proto,
                     framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The left input of row-wise add op, must be matrix");
@@ -49,12 +51,32 @@ for i in xrange(X.shape[0]):
 )DOC");
   }
 };
+class RowwiseAddGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "X should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"), "b should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+    auto dims0 = ctx.Input<Tensor>("X")->dims();
+    auto dims1 = ctx.Input<Tensor>("b")->dims();
+    PADDLE_ENFORCE_EQ(1, dims1.size(), "b dims should be 1")
+    ctx.Output<Tensor>(framework::GradVarName("X"))->Resize(dims0);
+    ctx.Output<Tensor>(framework::GradVarName("b"))->Resize(dims1);
+  }
+};
 
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(rowwise_add, ops::RowWiseAddOp,
-                             ops::RowWiseAddOpMaker);
+REGISTER_OP(rowwise_add, ops::RowwiseAddOp, ops::RowwiseAddOpMaker,
+            rowwise_add_grad, ops::RowwiseAddGradOp);
+REGISTER_OP_CPU_KERNEL(
+    rowwise_add, ops::RowwiseAddKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
-    rowwise_add, ops::RowWiseAddKernel<paddle::platform::CPUPlace, float>);
+    rowwise_add_grad,
+    ops::RowwiseAddGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/rowwise_add_op.cu

@@ -17,4 +17,4 @@
 
 namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(
-    rowwise_add, ops::RowWiseAddKernel<paddle::platform::GPUPlace, float>);
+    rowwise_add, ops::RowwiseAddKernel<paddle::platform::GPUPlace, float>);

paddle/operators/rowwise_add_op.h

@@ -28,7 +28,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
 template <typename Place, typename T>
-class RowWiseAddKernel : public framework::OpKernel {
+class RowwiseAddKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto out = context.Output<Tensor>("Out");
@@ -47,5 +47,25 @@ class RowWiseAddKernel : public framework::OpKernel {
   }
 };
 
+template <typename Place, typename T>
+class RowwiseAddGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* dOut = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dX = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* db = context.Output<Tensor>(framework::GradVarName("b"));
+    dX->mutable_data<T>(context.GetPlace());
+    db->mutable_data<T>(context.GetPlace());
+
+    auto OutGrad = EigenMatrix<T>::From(*dOut);
+    auto place = context.GetEigenDevice<Place>();
+    EigenMatrix<T>::From(*dX).device(place) = OutGrad;
+
+    // https://eigen.tuxfamily.org/dox/unsupported/TensorBase_8h_source.html
+    // colwise add
+    Eigen::array<int, 1> dims{{1}}; /* dimension to reduce */
+    EigenVector<T>::Flatten(*db).device(place) = OutGrad.sum(dims);
+  }
+};
 }  // namespace operators
 }  // namespace paddle

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

@@ -1,6 +1,7 @@
 import unittest
-from op_test_util import OpTestMeta
 import numpy as np
+from op_test_util import OpTestMeta
+from gradient_checker import GradientChecker, create_op
 
 
 class TestRowwiseAddOp(unittest.TestCase):
@@ -15,5 +16,15 @@ class TestRowwiseAddOp(unittest.TestCase):
         self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])}
 
 
+class RowwiseAddGradOpTest(GradientChecker):
+    def test_rowwise_add(self):
+        op = create_op("rowwise_add")
+        inputs = {
+            "X": np.random.uniform(0.1, 1, [10, 10]).astype("float32"),
+            "b": np.random.uniform(0.1, 1, [10]).astype("float32")
+        }
+        self.check_grad(op, inputs, set(["X", "b"]), "Out")
+
+
 if __name__ == '__main__':
     unittest.main()