Add element-wise multiplication operator. #3787

gongweibao · 2017-08-31T09:44:59Z

… addop

qingqing01 · 2017-08-31T13:08:39Z

paddle/operators/CMakeLists.txt

@@ -57,6 +57,7 @@ op_library(add_op SRCS add_op.cc add_op.cu)
 op_library(mean_op SRCS mean_op.cc mean_op.cu)

 op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
+op_library(element_wise_mul_op SRCS element_wise_mul_op.cc element_wise_mul_op.cu DEPS math_function)


remove DEPS math_function

Thanks.
Done.

qingqing01 · 2017-08-31T13:11:46Z

paddle/operators/element_wise_mul_op.h

+    auto Y_e = framework::EigenVector<T>::Flatten(*Y);
+    auto Z_e = framework::EigenVector<T>::Flatten(*Z);
+
+    Z_e.device(context.GetEigenDevice<Place>()) = X_e * Y_e;


The names of variables are all lowercas. https://google.github.io/styleguide/cppguide.html#Variable_Names

It's better to support DotMulProjection , where the second input is a row vector,
https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/layers/DotMulProjection.cpp#L57

The broadcast of Eigen can be used.

Both are needed.

emailweixu · 2017-08-31T23:14:17Z

paddle/operators/CMakeLists.txt

@@ -57,6 +57,7 @@ op_library(add_op SRCS add_op.cc add_op.cu)
 op_library(mean_op SRCS mean_op.cc mean_op.cu)

 op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
+op_library(element_wise_mul_op SRCS element_wise_mul_op.cc element_wise_mul_op.cu DEPS math_function)


Elementwise is one word.
elementwise_mul_op

Thanks.
Done.

emailweixu · 2017-08-31T23:17:17Z

paddle/operators/element_wise_mul_op.h

+    auto Y_e = framework::EigenVector<T>::Flatten(*Y);
+    auto Z_e = framework::EigenVector<T>::Flatten(*Z);
+
+    Z_e.device(context.GetEigenDevice<Place>()) = X_e * Y_e;


Both are needed.

emailweixu · 2017-08-31T23:18:38Z

paddle/operators/element_wise_mul_op.cc

+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(elemwisemul, ops::ElemWiseMulOp, ops::ElemWiseMulOpMaker,


elemwisemul => elementwise_mul

Thanks.
Done.

emailweixu · 2017-08-31T23:20:14Z

paddle/operators/element_wise_mul_op.h

+    auto dOut_e = framework::EigenVector<T>::Flatten(*dOut);
+
+    dX_e.device(ctx.GetEigenDevice<Place>()) = dOut_e * Y_e;
+    dY_e.device(ctx.GetEigenDevice<Place>()) = X_e * dOut_e;


Need to support the case where the gradient dX_e or dY_e is not needed.
@qingqing01 Most of the gradient operator need to handle the case where is one of more of the gradient of the input is not needed. Please change the unittest framework to explicitly test whether the op correctly handles this

@emailweixu Ok, I'll change the unit test framework to handle this.

… addop

qingqing01 · 2017-09-11T12:39:59Z

paddle/operators/CMakeLists.txt

    recurrent_op
    scale_op)
 op_library(identity_op DEPS scale_op)
 op_library(minus_op DEPS scale_op)
 op_library(mul_op DEPS math_function)
+op_library(elementwise_mul_op SRCS elementwise_mul_op.cc elementwise_mul_op.cu)


Remove line 57 and line 63. There is no DEPS for this op.

qingqing01 · 2017-09-11T12:42:01Z

paddle/operators/elementwise_mul_op.cc

+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto x_dim = ctx.Input<Tensor>("X")->dims();
+    auto y_dim = ctx.Input<Tensor>("Y")->dims();


The PADDLE_ENFORCE_NOT_NULL is needed for input X and Y, similar check in ElementWiseMulOpGrad.

qingqing01 · 2017-09-11T12:44:07Z

paddle/operators/elementwise_mul_op.cc

+                        framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The first input of elementwise mul op");
+    AddInput("Y", "The second input of elementwise mul op");


Add the supported shape for input X and Y.

#3885 (comment)

qingqing01 · 2017-09-11T12:46:02Z

paddle/operators/elementwise_mul_op.cc

+    AddInput("X", "The first input of elementwise mul op");
+    AddInput("Y", "The second input of elementwise mul op");
+    AddAttr<int>("axis", "Optional input parameter of elementwise mul op");
+    AddAttr<int>("broadcast", "Optional input parameter of elementwise mul op");


Need to describe the meaning and usage for axis and broadcast attr.

#3885 (comment)

从下面单测来看 axis像是y在x中的start_axis，需要更详细的说明。

qingqing01 · 2017-09-11T12:53:05Z

paddle/operators/elementwise_mul_op.h

+using Tensor = framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;


From googlge code style:

https://google.github.io/styleguide/cppguide.html#Namespaces

Do not use Namespace aliases at namespace scope in header files except in explicitly marked internal-only namespaces, because anything imported into a namespace in a header file becomes part of the public API exported by that file.

qingqing01 · 2017-09-11T12:56:38Z

paddle/operators/elementwise_mul_op.h

+
+    // TODO(gongweibao): if axis is optional?
+    bool broadcast = ctx.template Attr<int>("broadcast");
+    PADDLE_ENFORCE(broadcast, "Do you forget broadcast parameter?");


从这里实现来看 broadcast 要始终为True，为什么还要加这个参数？另外这个参数作用是啥？

broadcast和axis都是optional的参数。

qingqing01 · 2017-09-11T13:01:13Z

paddle/operators/elementwise_op.h

+
+inline void get_slice(const framework::DDim& x_dims,
+                      const framework::DDim& y_dims, const int axis, int& pre,
+                      int& n, int& post) {


从下面实现来看，不是给x_dims, y_dims做slice得，叫这个名字不是很理解。另外，单独放到ementwise_op.h里是否合适？

如果暂时只有这一个op用到，觉得先放到 elementwise_mul_op.h 中，后续增加新的op，再移出来都行。

qingqing01 · 2017-09-11T13:16:08Z

paddle/operators/elementwise_mul_op.h

+    PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
+                   "Axis should be in range [0, x_dims)");
+
+    int pre, n, post;


不是特别理解pre, n, post 的意思，能够加上comments吗？或者更明确的name吗？

Out = X ⊙ Y 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 pre=2, n=3*4,post=5 2.shape(X) = (2, 3, 4, 5), shape(Y) = (4,5) pre=2*3, n=4*5,post=1

已经加入到comment。

qingqing01 · 2017-09-11T13:19:34Z

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

+        """
+        self.compare_grad(self.op, self.inputs)
+        self.check_grad(
+            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5)


max_relative_error能调小一些吗？

qingqing01 · 2017-09-11T13:21:50Z

paddle/operators/elementwise_mul_op.h

+
+    if (x_dims == y_dims || product(y_dims) == 1) {
+      dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e;
+      dy_e.device(ctx.GetEigenDevice<Place>()) = x_e * dout_e;


if (dx) dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e; if (dy) dy_e.device(ctx.GetEigenDevice<Place>()) = x_e * dout_e;

需要处理某个输入可能不用计算grad的情况

QiJune · 2017-09-12T03:49:03Z

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

+from paddle.v2.framework.op import Operator
+
+
+class TestElementwiseMulOp_Matrix(unittest.TestCase):


需要先merge一下develop分支，因为目前的operator的测试框架有一些调整，也需要对应修改一下单元测试的写法

… addop

qingqing01 · 2017-09-12T05:20:10Z

paddle/operators/elementwise_mul_op.cc

+        "broadcast",
+        R"DOC(When broadcast is set, Y will be broadcast to match shape of X.
+        )DOC")
+        .SetDefault(0);


broadcast can be removed.

已经删除掉了。

qingqing01 · 2017-09-12T05:22:25Z

paddle/operators/elementwise_mul_op.h

+ * Out = X ⊙ Y
+ * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
+ *    pre=2, n=3*4, post=5
+ * 2.shape(X) = (2, 3, 4, 5), shape(Y) = (4,5)


shape(X) = (2, 3, 4, 5), shape(Y) = (4,5)

qingqing01 · 2017-09-12T05:23:35Z

paddle/operators/elementwise_mul_op.h

+      return;
+    }
+
+    int axis = ctx.template Attr<int>("axis");


ctx.Attr<int>

qingqing01 · 2017-09-12T05:31:38Z

paddle/operators/elementwise_mul_op.cc

+    AddOutput("Out", "The output of elementwise mul op");
+    AddComment(R"DOC(
+Limited elementwise multiple operator.The equation is: Out = X ⊙ Y.
+1. The shape of Y should be same of X or


The shape of Y should be same with X

qingqing01 · 2017-09-12T05:37:01Z

paddle/operators/elementwise_mul_op.h

+    }
+    if (dy) {
+      dy->mutable_data<T>(ctx.GetPlace());
+    }


short lines for line 119 - 124:

if (dx) dx->mutable_data<T>(ctx.GetPlace()); if (dy) dy->mutable_data<T>(ctx.GetPlace());

https://google.github.io/styleguide/cppguide.html#Conditionals
这个应该不是强制的。

主要是gdb调试的时候比较郁闷，不知道if后边的语句到底是不是执行了。

qingqing01 · 2017-09-12T05:39:10Z

paddle/operators/elementwise_mul_op.cc

+
+    if (y_grad) {
+      y_grad->Resize(y_dims);
+    }


for shorter lines:

if (x_grad) x_grad->Resize(x_dims); if (y_grad) y_grad->Resize(y_dims);

qingqing01 · 2017-09-12T05:39:57Z

paddle/operators/elementwise_mul_op.h

+      return;
+    }
+
+    int axis = ctx.template Attr<int>("axis");


ctx.Attr<int>("axis");

* Update bbox_utils.py

gongweibao added 6 commits August 30, 2017 09:16

add element-wise multiply

0bb7b0d

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

eae9acb

… addop

compile ok

0a3fd42

fix link bug

e2c48d1

test ok

2c0c858

add somae warning

bcbe92c

gongweibao changed the title ~~Add element-wise multiply operator.~~ Add element-wise multiple operator. Aug 31, 2017

gongweibao requested review from qingqing01, QiJune and reyoung August 31, 2017 09:45

qingqing01 requested a review from jacquesqiao August 31, 2017 13:07

qingqing01 reviewed Aug 31, 2017

View reviewed changes

qingqing01 requested a review from pkuyym August 31, 2017 13:21

emailweixu requested changes Aug 31, 2017

View reviewed changes

emailweixu changed the title ~~Add element-wise multiple operator.~~ Add element-wise multiplication operator. Aug 31, 2017

qingqing01 added the OpPorting label Sep 5, 2017

gongweibao added 5 commits September 7, 2017 11:34

fix and add some

157c327

modify test

2f8dfb1

modify to attr

e254981

compile ok

c5cb8f6

test mul forward ok

781f7cc

gongweibao changed the title ~~Add element-wise multiplication operator.~~ [WIP]Add element-wise multiplication operator. Sep 10, 2017

test ok

7580167

gongweibao changed the title ~~[WIP]Add element-wise multiplication operator.~~ Add element-wise multiplication operator. Sep 10, 2017

gongweibao added 2 commits September 10, 2017 10:18

rm not need

4bf8f9d

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

6219ce0

… addop

qingqing01 reviewed Sep 11, 2017

View reviewed changes

gongweibao added 2 commits September 12, 2017 03:40

fix by comments

611895c

rm not nedd

797422d

QiJune reviewed Sep 12, 2017

View reviewed changes

gongweibao added 3 commits September 12, 2017 03:52

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

cd9b492

… addop

fix by comments

4425e73

fix by comments

aa0042f

qingqing01 reviewed Sep 12, 2017

View reviewed changes

fix by comments

1c6b7e4

qingqing01 approved these changes Sep 12, 2017

View reviewed changes

emailweixu approved these changes Sep 13, 2017

View reviewed changes

gongweibao merged commit 8778957 into PaddlePaddle:develop Sep 13, 2017

gongweibao deleted the addop branch September 13, 2017 06:18

kuke mentioned this pull request Sep 14, 2017

Add scaling operator #3942

Closed

gongweibao restored the addop branch September 15, 2017 03:38

gongweibao deleted the addop branch January 17, 2021 07:41

heavengate pushed a commit to heavengate/Paddle that referenced this pull request Aug 16, 2021

Fix bug in nonempty_bbox for height (PaddlePaddle#3787)

7c50785

* Update bbox_utils.py

		from paddle.v2.framework.op import Operator


		class TestElementwiseMulOp_Matrix(unittest.TestCase):

Add element-wise multiplication operator. #3787

Add element-wise multiplication operator. #3787

Conversation

gongweibao commented Aug 31, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

emailweixu Aug 31, 2017 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gongweibao Sep 12, 2017 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

qingqing01 Sep 12, 2017 • edited by gongweibao

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

emailweixu Aug 31, 2017 •

edited

gongweibao Sep 12, 2017 •

edited

qingqing01 Sep 12, 2017 •

edited by gongweibao