Merge pull request #3379 from dzhwinter/mul_op

"mul op gradient"

paddle/operators/math/math_function.cc

@@ -25,8 +25,8 @@ void gemm<platform::CPUPlace, float>(const CBLAS_TRANSPOSE transA,
                                      const float alpha, const float* A,
                                      const float* B, const float beta, float* C,
                                      platform::DeviceContext* context) {
-  int lda = K;
-  int ldb = N;
+  int lda = (transA == CblasNoTrans) ? K : M;
+  int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
   cblas_sgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
               beta, C, ldc);
@@ -40,8 +40,8 @@ void gemm<platform::CPUPlace, double>(const CBLAS_TRANSPOSE transA,
                                       const double* B, const double beta,
                                       double* C,
                                       platform::DeviceContext* context) {
-  int lda = K;
-  int ldb = N;
+  int lda = (transA == CblasNoTrans) ? K : M;
+  int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
   cblas_dgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
               beta, C, ldc);

paddle/operators/mul_op.cc

@@ -18,6 +18,8 @@
 namespace paddle {
 namespace operators {
 
+using framework::Tensor;
+
 class MulOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -59,10 +61,23 @@ class MulOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {}
-  std::string DebugString() const override {
-    LOG(INFO) << "MulGrad";
-    return "";
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    PADDLE_ENFORCE(x_dims[0] == out_dims[0],
+                   "Out@GRAD M X N must equal to X dims 0, M ");
+    PADDLE_ENFORCE(y_dims[1] == out_dims[1],
+                   "Out@GRAD M X N must equal to Y dims 1, N ");
+
+    x_grad->Resize(x_dims);
+    y_grad->Resize(y_dims);
   }
 };
 
@@ -72,3 +87,5 @@ class MulOpGrad : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
 REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(mul_grad,
+                       ops::MulGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/mul_op.cu

@@ -17,3 +17,5 @@
 
 namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(mul_grad,
+                       ops::MulGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/mul_op.h

@@ -31,18 +31,34 @@ template <typename Place, typename T>
 class MulKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair = {
-        {Eigen::IndexPair<Eigen::DenseIndex>(1, 0)}};
-    auto* input0 = context.Input<Tensor>("X");
-    auto* input1 = context.Input<Tensor>("Y");
-    auto* output = context.Output<Tensor>("Out");
-    output->mutable_data<T>(context.GetPlace());
-    auto X = EigenMatrix<T>::From(*input0);
-    auto Y = EigenMatrix<T>::From(*input1);
-    auto Z = EigenMatrix<T>::From(*output);
-    auto& place = context.GetEigenDevice<Place>();
-
-    Z.device(place) = X.contract(Y, dim_pair);
+    auto* X = context.Input<Tensor>("X");
+    auto* Y = context.Input<Tensor>("Y");
+    auto* Z = context.Output<Tensor>("Out");
+    Z->mutable_data<T>(context.GetPlace());
+    auto* device_context =
+        const_cast<platform::DeviceContext*>(context.device_context_);
+    math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
+  }
+};
+
+template <typename Place, typename T>
+class MulGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* X = ctx.Input<Tensor>("X");
+    auto* Y = ctx.Input<Tensor>("Y");
+    auto* dOut = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+    auto* dX = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dY = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    dX->mutable_data<T>(ctx.GetPlace());
+    dY->mutable_data<T>(ctx.GetPlace());
+    auto* device_context =
+        const_cast<platform::DeviceContext*>(ctx.device_context_);
+    // dX = dOut * Y'. dX: M x K, dOut : M x N, Y : K x N
+    math::matmul<Place, T>(*dOut, false, *Y, true, 1, dX, 0, device_context);
+    // dY = X' * dOut. dY: K x N, dOut : M x N, X : M x K
+    math::matmul<Place, T>(*X, true, *dOut, false, 1, dY, 0, device_context);
   }
 };
 

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

 import unittest
-from op_test_util import OpTestMeta
 import numpy as np
+from gradient_checker import GradientChecker, create_op
+from op_test_util import OpTestMeta
 
 
 class TestMulOp(unittest.TestCase):
@@ -15,5 +16,19 @@ class TestMulOp(unittest.TestCase):
         self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
 
 
+class MulGradOpTest(GradientChecker):
+    def test_mul(self):
+        op = create_op("mul")
+        inputs = {
+            'X': np.random.random((32, 84)).astype("float32"),
+            'Y': np.random.random((84, 100)).astype("float32")
+        }
+        # mul op will enlarge the relative error
+        self.check_grad(
+            op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5)
+
+
+# TODO(dzh,qijun) : mulgrad test case need transpose feature of blas library
+
 if __name__ == '__main__':
     unittest.main()