modification

Author: sneaxiy

paddle/fluid/API.spec

@@ -130,7 +130,7 @@ paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name
 paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.edit_distance ArgSpec(args=['input', 'label', 'normalized', 'ignored_tokens'], varargs=None, keywords=None, defaults=(True, None))
 paddle.fluid.layers.l2_normalize ArgSpec(args=['x', 'axis', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(1e-12, None))
-paddle.fluid.layers.matmul ArgSpec(args=['x', 'y', 'transpose_x', 'transpose_y', 'scale', 'bias', 'name'], varargs=None, keywords=None, defaults=(False, False, 1.0, 0.0, None))
+paddle.fluid.layers.matmul ArgSpec(args=['x', 'y', 'transpose_x', 'transpose_y', 'alpha', 'name'], varargs=None, keywords=None, defaults=(False, False, 1.0, None))
 paddle.fluid.layers.topk ArgSpec(args=['input', 'k', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_times'], varargs=None, keywords=None, defaults=(0, False))
 paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None)

paddle/fluid/operators/elementwise_mul_op.h

@@ -93,7 +93,6 @@ class ElementwiseMulGradKernel : public ElemwiseGradKernel<T> {
 
     auto* x = ctx.Input<Tensor>("X");
     auto* y = ctx.Input<Tensor>("Y");
-    // auto* out = ctx.Input<Tensor>("Out");
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
     auto* out = dout;  // out is not necessary
     auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));

paddle/fluid/operators/matmul_op.cc

@@ -59,9 +59,8 @@ class MatMulKernel : public framework::OpKernel<T> {
         RowMatrixFromVector(x.dims()), 0, context.Attr<bool>("transpose_X"));
     auto mat_dim_b = math::CreateMatrixDescriptor(
         ColumnMatrixFromVector(y.dims()), 0, context.Attr<bool>("transpose_Y"));
-    auto scale = static_cast<T>(context.Attr<float>("scale"));
-    auto bias = static_cast<T>(context.Attr<float>("bias"));
-    blas.MatMul(x, mat_dim_a, y, mat_dim_b, scale, out, bias);
+    auto scale = static_cast<T>(context.Attr<float>("alpha"));
+    blas.MatMul(x, mat_dim_a, y, mat_dim_b, scale, out, T(0));
   }
 };
 
@@ -188,7 +187,7 @@ class MatMulGradKernel : public framework::OpKernel<T> {
     auto mat_dim_a = math::CreateMatrixDescriptor(a.dims(), 0, trans_a);
     auto mat_dim_b = math::CreateMatrixDescriptor(b.dims(), 0, trans_b);
     blas.MatMul(a, mat_dim_a, b, mat_dim_b,
-                static_cast<T>(context.Attr<float>("scale")), out, T(0));
+                static_cast<T>(context.Attr<float>("alpha")), out, T(0));
   }
 
   void CalcInputGrad(const framework::ExecutionContext &context,
@@ -337,8 +336,7 @@ class MatMulOpMaker : public framework::OpProtoAndCheckerMaker {
                   R"DOC(If true, use the transpose of `Y`.
         )DOC")
         .SetDefault(false);
-    AddAttr<float>("scale", "Scale").SetDefault(1.0f);
-    AddAttr<float>("bias", "Bias").SetDefault(0.0f);
+    AddAttr<float>("alpha", "The scale of Out").SetDefault(1.0f);
     AddComment(R"DOC(
 MatMul Operator.
 

paddle/fluid/operators/scale_op.cc

@@ -53,6 +53,11 @@ Multiply the input tensor with a float scalar to scale the input tensor.
     AddAttr<float>("scale", "The scaling factor of the scale operator.")
         .SetDefault(1.0);
     AddAttr<float>("bias", "The bias of the scale operator.").SetDefault(0.0);
+    AddAttr<bool>(
+        "bias_after_scale",
+        "Apply bias addition after or before scaling. It is useful for "
+        "numeric stability in some circumstances.")
+        .SetDefault(true);
   }
 };
 
@@ -82,6 +87,7 @@ class ScaleGradMaker : public framework::SingleGradOpDescMaker {
     grad_op->SetOutput("Out", InputGrad("X"));
     grad_op->SetAttr("scale", GetAttr("scale"));
     grad_op->SetAttr("bias", 0.0f);
+    grad_op->SetAttr("bias_after_scale", true);
     return std::unique_ptr<framework::OpDesc>(grad_op);
   }
 };

paddle/fluid/operators/scale_op.h

@@ -35,6 +35,7 @@ class ScaleKernel : public framework::OpKernel<T> {
 
     auto scale = static_cast<T>(ctx.Attr<float>("scale"));
     auto bias = static_cast<T>(ctx.Attr<float>("bias"));
+    auto bias_after_scale = ctx.Attr<bool>("bias_after_scale");
 
     if (in_var->IsType<framework::SelectedRows>() && in_var != out_var) {
       auto& in_slr = in_var->Get<framework::SelectedRows>();
@@ -46,8 +47,11 @@ class ScaleKernel : public framework::OpKernel<T> {
     auto eigen_out = framework::EigenVector<T>::Flatten(*out);
     auto eigen_in = framework::EigenVector<T>::Flatten(*in);
     auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
-    eigen_out.device(dev) =
-        static_cast<T>(scale) * eigen_in + static_cast<T>(bias);
+    if (bias_after_scale) {
+      eigen_out.device(dev) = scale * eigen_in + bias;
+    } else {
+      eigen_out.device(dev) = scale * (eigen_in + bias);
+    }
   }
 };
 

python/paddle/fluid/layers/nn.py

@@ -3388,13 +3388,7 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None):
     return out
 
 
-def matmul(x,
-           y,
-           transpose_x=False,
-           transpose_y=False,
-           scale=1.0,
-           bias=0.0,
-           name=None):
+def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
     """
     Applies matrix multiplication to two tensors.
 
@@ -3428,8 +3422,7 @@ def matmul(x,
         y (Variable): The input variable which is a Tensor or LoDTensor.
         transpose_x (bool): Whether to transpose :math:`x` before multiplication.
         transpose_y (bool): Whether to transpose :math:`y` before multiplication.
-        scale (float): The scale of output. Default 1.0.
-        bias (float): The bias added to output. Default 0.0.
+        alpha (float): The scale of output. Default 1.0.
         name(str|None): A name for this layer(optional). If set None, the layer
             will be named automatically.
 
@@ -3500,8 +3493,7 @@ def __check_input(x, y):
         attrs={
             'transpose_X': transpose_x,
             'transpose_Y': transpose_y,
-            'scale': scale,
-            'bias': bias
+            'alpha': alpha,
         })
     return out