Refine smooth L1 loss. (PaddlePaddle#10713)

qingqing01 · web-flow · commit 1f8243b7f8dd · 2018-05-17T09:14:09.000+08:00
diff --git a/paddle/fluid/operators/smooth_l1_loss_op.cc b/paddle/fluid/operators/smooth_l1_loss_op.cc
@@ -105,7 +105,7 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext* ctx) const override {
-    auto in_dims = ctx->GetInputDim("X");
+    auto in_dims = ctx->GetInputDim("Diff");
     auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
 
     PADDLE_ENFORCE_GE(out_dims.size(), 2,
@@ -127,12 +127,33 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel {
   }
 };
 
+class SmoothL1LossGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto* op = new framework::OpDesc();
+    op->SetType("smooth_l1_loss_grad");
+    op->SetInput("InsideWeight", Input("InsideWeight"));
+    op->SetInput("OutsideWeight", Input("OutsideWeight"));
+    op->SetInput("Diff", Output("Diff"));
+    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 std::unique_ptr<framework::OpDesc>(op);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(smooth_l1_loss, ops::SmoothL1LossOp, ops::SmoothL1LossOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
+                  ops::SmoothL1LossGradMaker);
 REGISTER_OPERATOR(smooth_l1_loss_grad, ops::SmoothL1LossGradOp);
 REGISTER_OP_CPU_KERNEL(
     smooth_l1_loss,
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
@@ -3263,35 +3263,35 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
     """
     **Smooth L1 Loss Operator. **
 
-    This operator computes the smooth l1 loss for X and Y.
+    This operator computes the smooth L1 loss for X and Y.
     The operator takes the first dimension of X and Y as batch size.
-    For each instance, it computes the smooth l1 loss element by element first
+    For each instance, it computes the smooth L1 loss element by element first
     and then sums all the losses. So the shape of Out is [batch_size, 1].
 
     Args:
         x (Variable): A tensor with rank at least 2. The input value of smooth
-            l1 loss op with shape [batch_size, dim1, ..., dimN].
+            L1 loss op with shape [batch_size, dim1, ..., dimN].
         y (Variable): A tensor with rank at least 2. The target value of smooth
-            l1 loss op with same shape as x.
+            L1 loss op with same shape as x.
         inside_weight (Variable|None):  A tensor with rank at least 2. This
             input is optional and should have same shape with x. If provided,
             the result of (x - y) will be multiplied by this tensor element by
             element.
         outside_weight (Variable|None): A tensor with rank at least 2. This
             input is optional and should have same shape with x. If provided,
-            the out smooth l1 loss will be multiplied by this tensor element
+            the out smooth L1 loss will be multiplied by this tensor element
             by element.
-        sigma (float|None): Hyper parameter of smooth l1 loss op. A float scalar
+        sigma (float|None): Hyper parameter of smooth L1 loss op. A float scalar
             with default value 1.0.
     Returns:
-        Variable: A tensor with rank be 2. The output smooth l1 loss with
+        Variable: A tensor with rank be 2. The output smooth L1 loss with
             shape [batch_size, 1].
 
     Examples:
         .. code-block:: python
 
             data = fluid.layers.data(name='data', shape=[128], dtype='float32')
-            label = fluid.layers.data(name='label', shape=[100], dtype='int64')
+            label = fluid.layers.data(name='label', shape=[100], dtype='float32')
             fc = fluid.layers.fc(input=data, size=100)
             out = fluid.layers.smooth_l1(x=fc, y=label)
     """
