Merge pull request #7593 from JiayiFeng/dev_elementwise_scalar

Make elementwise_op supporting scalar input `Y`

paddle/operators/elementwise_div_op.h

@@ -19,11 +19,16 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+template <typename T>
+struct DivFunctor {
+  inline HOSTDEVICE T operator()(T a, T b) const { return a / b; }
+};
+
 template <typename DeviceContext, typename T>
 class ElementwiseDivKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseCompute<EigenDivFunctor, DeviceContext, T>(ctx);
+    ElementwiseComputeEx<DivFunctor<T>, DeviceContext, T>(ctx);
   }
 };
 

paddle/operators/elementwise_mul_op.h

@@ -18,11 +18,16 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+template <typename T>
+struct MulFunctor {
+  inline HOSTDEVICE T operator()(T a, T b) const { return a * b; }
+};
+
 template <typename DeviceContext, typename T>
 class ElementwiseMulKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseCompute<EigenMulFunctor, DeviceContext, T>(ctx);
+    ElementwiseComputeEx<MulFunctor<T>, DeviceContext, T>(ctx);
   }
 };
 

paddle/operators/elementwise_op_function.h

@@ -340,6 +340,13 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
     return;
   }
 
+  if (y_dims.size() == 1 && y_dims[0] == 1) {
+    // y is a scalar
+    auto extended_dims = framework::vectorize(x_dims);
+    extended_dims.push_back(1);
+    x_dims = framework::make_ddim(extended_dims);
+  }
+
   int axis = ctx.Attr<int>("axis");
   axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
 
@@ -378,6 +385,13 @@ void ElementwiseComputeEx(const framework::ExecutionContext& ctx) {
     return;
   }
 
+  if (y_dims.size() == 1 && y_dims[0] == 1) {
+    // y is a scalar
+    auto extended_dims = framework::vectorize(x_dims);
+    extended_dims.push_back(1);
+    x_dims = framework::make_ddim(extended_dims);
+  }
+
   int axis = ctx.Attr<int>("axis");
   axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
   PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),

paddle/operators/elementwise_sub_op.h

@@ -18,11 +18,16 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+template <typename T>
+struct SubFunctor {
+  inline HOSTDEVICE T operator()(T a, T b) const { return a - b; }
+};
+
 template <typename DeviceContext, typename T>
 class ElementwiseSubKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseCompute<EigenSubFunctor, DeviceContext, T>(ctx);
+    ElementwiseComputeEx<SubFunctor<T>, DeviceContext, T>(ctx);
   }
 };
 

python/paddle/v2/fluid/tests/test_elementwise_add_op.py

@@ -40,6 +40,16 @@ class TestElementwiseOp(OpTest):
             ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
 
 
+class TestElementwiseAddOp_scalar(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_add"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(1).astype(np.float32)
+        }
+        self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
+
+
 class TestElementwiseAddOp_Vector(TestElementwiseOp):
     def setUp(self):
         self.op_type = "elementwise_add"

python/paddle/v2/fluid/tests/test_elementwise_div_op.py

@@ -45,6 +45,16 @@ class ElementwiseDivOp(OpTest):
             ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y'))
 
 
+class TestElementwiseDivOp_scalar(ElementwiseDivOp):
+    def setUp(self):
+        self.op_type = "elementwise_div"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 3, 4]).astype(np.float32),
+            'Y': np.random.uniform(0.1, 1, [1]).astype(np.float32)
+        }
+        self.outputs = {'Out': self.inputs['X'] / self.inputs['Y']}
+
+
 class TestElementwiseDivOp_Vector(ElementwiseDivOp):
     def setUp(self):
         self.op_type = "elementwise_div"

python/paddle/v2/fluid/tests/test_elementwise_max_op.py

@@ -43,6 +43,15 @@ class TestElementwiseOp(OpTest):
             ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
 
 
+class TestElementwiseMaxOp_scalar(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_max"
+        x = np.random.random_integers(-5, 5, [2, 3, 4]).astype("float32")
+        y = np.array([0.5]).astype("float32")
+        self.inputs = {'X': x, 'Y': y}
+        self.outputs = {'Out': np.maximum(self.inputs['X'], self.inputs['Y'])}
+
+
 class TestElementwiseMaxOp_Vector(TestElementwiseOp):
     def setUp(self):
         self.op_type = "elementwise_max"

python/paddle/v2/fluid/tests/test_elementwise_min_op.py

@@ -43,6 +43,15 @@ class TestElementwiseOp(OpTest):
             ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
 
 
+class TestElementwiseMinOp_scalar(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_min"
+        x = np.random.random_integers(-5, 5, [2, 3, 4]).astype("float32")
+        y = np.array([0.5]).astype("float32")
+        self.inputs = {'X': x, 'Y': y}
+        self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
+
+
 class TestElementwiseMaxOp_Vector(TestElementwiseOp):
     def setUp(self):
         self.op_type = "elementwise_min"

python/paddle/v2/fluid/tests/test_elementwise_mul_op.py

@@ -38,6 +38,16 @@ class ElementwiseMulOp(OpTest):
         self.check_grad(['X'], 'Out', no_grad_set=set('Y'))
 
 
+class TestElementwiseMulOp_scalar(ElementwiseMulOp):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(1).astype(np.float32)
+        }
+        self.outputs = {'Out': self.inputs['X'] * self.inputs['Y']}
+
+
 class TestElementwiseMulOp_Vector(ElementwiseMulOp):
     def setUp(self):
         self.op_type = "elementwise_mul"

python/paddle/v2/fluid/tests/test_elementwise_sub_op.py

@@ -40,6 +40,16 @@ class TestElementwiseOp(OpTest):
             ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
 
 
+class TestElementwiseSubOp_scalar(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_sub"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(1).astype(np.float32)
+        }
+        self.outputs = {'Out': self.inputs['X'] - self.inputs['Y']}
+
+
 class TestElementwiseSubOp_Vector(TestElementwiseOp):
     def setUp(self):
         self.op_type = "elementwise_sub"