提交 8d88c52d 编写于 作者: C caoying03

Merge branch 'develop' into softmax_with_cross_entropy_op

...@@ -262,7 +262,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, ...@@ -262,7 +262,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
- 生成库 - 生成库
无需修改 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件,`paddle/operators` 目录下新增的 `*_op.cc` 文件会被自动添加链接到生成的lib库中。 `paddle/operators` 目录下新增的 `*_op.cc` 文件会被自动添加链接到生成的lib库中。
## 实现单元测试 ## 实现单元测试
...@@ -354,11 +354,7 @@ class TestMulGradOp(GradientChecker): ...@@ -354,11 +354,7 @@ class TestMulGradOp(GradientChecker):
### 编译和执行单元测试 ### 编译和执行单元测试
单元测试编写完成之后,在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)中添加以下内容,将单元测试加入工程: `python/paddle/v2/framework/tests` 目录下新增的 `test_*.py` 单元测试会被自动加入工程进行编译。
```
py_test(test_mul_op SRCS test_mul_op.py)
```
请注意,**不同于Op的编译测试,运行单元测试测时需要编译整个工程**,并且编译时需要打开`WITH_TESTING`, 即`cmake paddle_dir -DWITH_TESTING=ON`。编译成功后,执行下面的命令来运行单元测试: 请注意,**不同于Op的编译测试,运行单元测试测时需要编译整个工程**,并且编译时需要打开`WITH_TESTING`, 即`cmake paddle_dir -DWITH_TESTING=ON`。编译成功后,执行下面的命令来运行单元测试:
......
...@@ -2,11 +2,22 @@ ...@@ -2,11 +2,22 @@
## Motivation ## Motivation
In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the gradient operators/expressions together with the chain rule. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass. In Neural Network, many model is solved by the the backpropagation algorithm(known as BP) at present. Technically it caculates the gradient of the loss function, then distributed back through the networks. Follows the chain rule, so we need a module chains the gradient operators/expressions together with to construct the backward pass. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
## Backward Operator Registry ## Implementation
A backward network is built up with several backward operators. Backward operators take forward operators' inputs outputs, and output gradients and then calculate its input gradients. In this design doc, we exported only one API for generating the backward pass.
```c++
std::unique_ptr<OperatorBase> Backward(const OperatorBase& forwardOp,
const std::unordered_set<std::string>& no_grad_vars);
```
The implementation behind it can be divided into two parts, **Backward Operator Creating** and **Backward Operator Building**.
### Backward Operator Registry
A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs, and output gradients and then calculate its input gradients.
| | forward operator | backward operator | | forward operator | backward operator
| ---------------------- | ---------------- |------------------------- | | ---------------------- | ---------------- |------------------------- |
...@@ -25,7 +36,7 @@ REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad); ...@@ -25,7 +36,7 @@ REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
`mul_grad` is the type of backward operator, and `MulOpGrad` is its class name. `mul_grad` is the type of backward operator, and `MulOpGrad` is its class name.
## Backward Opeartor Creating ### Backward Opeartor Creating
Given a certain forward operator, we can get its corresponding backward operator by calling: Given a certain forward operator, we can get its corresponding backward operator by calling:
...@@ -43,40 +54,47 @@ The function `BuildGradOp` will sequentially execute following processes: ...@@ -43,40 +54,47 @@ The function `BuildGradOp` will sequentially execute following processes:
4. Building backward operator with `inputs`, `outputs` and forward operator's attributes. 4. Building backward operator with `inputs`, `outputs` and forward operator's attributes.
## Backward Network Building ### Backward Network Building
A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and put them together.
In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and append them together one by one. There is some corner case need to process specially.
given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`, `InputGradients`.
1. Op 1. Op
when the input forward network is an Op, return its gradient Operator Immediately. When the input forward network is an Op, return its gradient Operator Immediately. If all of its outputs are in no gradient set, then return a special `NOP`.
2. NetOp 2. NetOp
when the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp. In our design, the network itself is also a kind of operator(**NetOp**). So the operators contained by a big network may be some small network. When the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
3. RnnOp
RnnOp is a nested stepnet operator. Backward module need to recusively call `Backward` for every stepnet.
4. Sharing Variables
**sharing variables**. As illustrated in the pictures, two operator's share the same variable name of W@GRAD, which will overwrite their sharing input variable.
<p align="center">
<img src="./images/duplicate_op.png" width="50%" ><br/>
**shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwrite their shared input variable. ​ pic 1. Sharing variables in operators.
<p align="center"> </p>
<img src="./images/duplicate_op.png" width="50%" ><br/>
1. Shared variable in operators. ​ Sharing variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator to replace the overwrite links.
</p> <p align="center">
<img src="images/duplicate_op2.png" width="40%" ><br/>
Share variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator replace the overwrite links. ​ pic 2. Replace sharing variable's gradient with `Add` operator.
<p align="center"> </p>
<img src="images/duplicate_op2.png" width="50%" ><br/>
2. Replace shared variable's gradient with `Add` operator. ​ Because our framework finds variables accord to their names, we need to rename the output links. We add a suffix of number to represent its position in clockwise.
</p> 5. Part of Gradient is Zero.
In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator, we need to fill a same shape gradient matrix in the position. In our implement, we insert a special `fillZeroLike` operator.
​ Then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it. Follow these rules above, then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.
...@@ -83,8 +83,7 @@ void SwitchOrderLayer::forward(PassType passType) { ...@@ -83,8 +83,7 @@ void SwitchOrderLayer::forward(PassType passType) {
setOutDims(); setOutDims();
resetOutput(outDims_[0], outDims_[1] * outDims_[2] * outDims_[3]); resetOutput(outDims_[0], outDims_[1] * outDims_[2] * outDims_[3]);
if (heightAxis_.size() > 0) { if (heightAxis_.size() > 0) {
getOutputValue()->reshape(reshapeHeight_, reshapeWidth_); resetOutput(reshapeHeight_, reshapeWidth_);
getOutputGrad()->reshape(reshapeHeight_, reshapeWidth_);
} }
// switch NCHW to NHWC // switch NCHW to NHWC
......
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/elementwise_mul_op.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
class ElementWiseMulOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
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");
auto x_dim = ctx.Input<Tensor>("X")->dims();
auto y_dim = ctx.Input<Tensor>("Y")->dims();
PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
"Rank of first input must >= rank of second input.")
ctx.Output<Tensor>("Out")->Resize(x_dim);
}
};
class ElementWiseMulOpMaker : public framework::OpProtoAndCheckerMaker {
public:
ElementWiseMulOpMaker(framework::OpProto *proto,
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");
AddAttr<int>("axis",
R"DOC(
When shape(Y) does not equal shape(X),Y will be broadcasted
to match the shape of X and axis should be dimension index Y in X
)DOC")
.SetDefault(-1)
.EqualGreaterThan(-1);
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 with X or
2. Y's shape is a subset of X.
Y will be broadcasted to match the shape of X and axis should be dimension index Y in X.
example:
shape(X) = (2, 3, 4, 5), shape(Y) = (,)
shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
)DOC");
}
};
class ElementWiseMulOpGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
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_GE(x_dims.size(), y_dims.size(),
"Rank of first input must >= rank of second input.")
if (x_grad) {
x_grad->Resize(x_dims);
}
if (y_grad) {
y_grad->Resize(y_dims);
}
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(elementwise_mul, ops::ElementWiseMulOp, ops::ElementWiseMulOpMaker,
elementwise_mul_grad, ops::ElementWiseMulOpGrad);
REGISTER_OP_CPU_KERNEL(
elementwise_mul,
ops::ElementWiseMulKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
elementwise_mul_grad,
ops::ElementWiseMulGradKernel<paddle::platform::CPUPlace, float>);
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#define EIGEN_USE_GPU
#include "paddle/operators/elementwise_mul_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
elementwise_mul,
ops::ElementWiseMulKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
elementwise_mul_grad,
ops::ElementWiseMulGradKernel<paddle::platform::GPUPlace, float>);
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <iostream>
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/math_function.h"
namespace paddle {
namespace operators {
/*
* 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
*/
inline void get_mid_dims(const framework::DDim& x_dims,
const framework::DDim& y_dims, const int axis,
int& pre, int& n, int& post) {
pre = 1;
n = 1;
post = 1;
for (int i = 0; i < axis; ++i) {
pre *= x_dims[i];
}
for (int i = 0; i < y_dims.size(); ++i) {
PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i],
"Broadcast dimension mismatch.");
n *= y_dims[i];
}
for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
post *= x_dims[i];
}
}
template <typename Place, typename T>
class ElementWiseMulKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
using Tensor = framework::Tensor;
auto* x = ctx.Input<Tensor>("X");
auto* y = ctx.Input<Tensor>("Y");
auto* z = ctx.Output<Tensor>("Out");
z->mutable_data<T>(ctx.GetPlace());
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto z_e = framework::EigenVector<T>::Flatten(*z);
auto x_dims = x->dims();
auto y_dims = y->dims();
PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
"Rank of first input must >= rank of second input.")
if (x_dims == y_dims || product(y_dims) == 1) {
z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_e;
return;
}
int axis = ctx.Attr<int>("axis");
axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
"Axis should be in range [0, x_dims)");
int pre, n, post;
get_mid_dims(x_dims, y_dims, axis, pre, n, post);
if (post == 1) {
auto y_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
return;
} else {
auto y_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
return;
}
}
};
template <typename Place, typename T>
class ElementWiseMulGradKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
using Tensor = framework::Tensor;
auto* x = ctx.Input<Tensor>("X");
auto* y = ctx.Input<Tensor>("Y");
auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dout_e = framework::EigenVector<T>::Flatten(*dout);
auto x_dims = x->dims();
auto y_dims = y->dims();
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
if (dx) {
dx->mutable_data<T>(ctx.GetPlace());
}
if (dy) {
dy->mutable_data<T>(ctx.GetPlace());
}
if (x_dims == y_dims || product(y_dims) == 1) {
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(ctx.GetEigenDevice<Place>()) = x_e * dout_e;
}
return;
}
int axis = ctx.Attr<int>("axis");
axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
int pre, n, post;
get_mid_dims(x_dims, y_dims, axis, pre, n, post);
// TODO(gongweibao): wrap reshape to a function.
if (post == 1) {
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(ctx.GetEigenDevice<Place>()) =
(x_e * dout_e)
.reshape(Eigen::DSizes<int, 2>(pre, n))
.sum(Eigen::array<int, 1>{{0}});
}
return;
} else {
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(ctx.GetEigenDevice<Place>()) =
(x_e * dout_e)
.reshape(Eigen::DSizes<int, 3>(pre, n, post))
.sum(Eigen::array<int, 2>{{0, 2}});
}
return;
}
}
};
} // namespace operators
} // namespace paddle
## Operator's Parameter Name Convention
To make the operator document itself more clear, we recommend operator names obey the listing conventions.
### OpProtoMaker names
When defining an operator in Paddle, a corresponding [OpProtoMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L170) (TODO: OpProtoMaker Doc)need to be defined. All the Input/Output and Attributes will write into the [OpProto](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L61) , and will be used in client language to create operator.
- Input/Output.
- Input/Output names follow the **CamelCase**. e.g. `X`, `Y`, `Matrix`, `LastAxisInMatrix`. Input/Output much more like Variables, we prefer to meaningful English words.
- If an operator's Input/Output are tensors in math, not match to any meaningful words, input name should starts from `X`. e.g. `X`, `Y`, and output name should starts from `Out`. e.g. `Out`. This rule intends making operators which have few inputs/outputs unified.
- Attribute.
- Attribute name follows the **camelCase**. e.g. `x`, `y`, `axis`, `rowwiseMatrix`. Also, attribute name prefers to meaningful English words.
- Comments.
- Input/Output/Attr comment follow the format of **(type,default value) usage**, corresponding to which type it can be and how it will be used in the operator. e.g. Attribute in Accumulator`"gamma" `,`(float, default 1.0) Accumulation multiplier`.
- Operator comment format of` R"DOC(your comment here)DOC"`. You should explain the input/output of the operator first. If there is math calculation in this operator, you should write the equation in the comment. e.g. `Out = X + Y`.
- Order.
- Follow the order of Input/Output, then Attribute, then Comments. See the example in best practice.
### Best Practice
Here we give some examples to show how these rules will be used.
- The operator has one input, one output. e.g.`relu`, inputs: `X`, outputs: `Out`.
- The operator has two input, one output. e.g. `rowwise_add`, inputs : `X`, `Y`, outputs : `Out`.
- The operator contains attribute. e.g. `cosine`, inputs : `X`, `axis`, outputs : `Out`.
We give a full example of Accumulator Operator.
```c++
class AccumulateOpMaker : public framework::OpProtoAndCheckerMaker {
public:
AccumulateOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "(Tensor) The input tensor that has to be accumulated to the output tensor. If the output size is not the same as input size, the output tensor is first reshaped and initialized to zero, and only then, accumulation is done.");
AddOutput("Out", "(Tensor) Accumulated output tensor");
AddAttr<float>("gamma", "(float, default 1.0) Accumulation multiplier");
AddComment(R"DOC(
Accumulate operator accumulates the input tensor to the output tensor. If the
output tensor already has the right size, we add to it; otherwise, we first
initialize the output tensor to all zeros, and then do accumulation. Any
further calls to the operator, given that no one else fiddles with the output
in the interim, will do simple accumulations.
Accumulation is done as shown:
Out = 1*X + gamma*Out
where X is the input tensor, Y is the output tensor and gamma is the multiplier
argument.
)DOC");
}
};
```
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License"); Licensed under the Apache License, Version 2.0 (the "License");
...@@ -52,5 +51,5 @@ class ReshapeGradKernel : public framework::OpKernel { ...@@ -52,5 +51,5 @@ class ReshapeGradKernel : public framework::OpKernel {
d_x->Resize(in_dims); d_x->Resize(in_dims);
} }
}; };
} } // namespace operators
} } // namespace paddle
...@@ -35,6 +35,7 @@ USE_OP(add); ...@@ -35,6 +35,7 @@ USE_OP(add);
USE_OP(onehot_cross_entropy); USE_OP(onehot_cross_entropy);
USE_OP(sgd); USE_OP(sgd);
USE_OP(mul); USE_OP(mul);
USE_OP(elementwise_mul);
USE_OP(mean); USE_OP(mean);
USE_OP(sigmoid); USE_OP(sigmoid);
USE_OP(softmax); USE_OP(softmax);
......
...@@ -2034,6 +2034,7 @@ class ParameterReluLayer(LayerBase): ...@@ -2034,6 +2034,7 @@ class ParameterReluLayer(LayerBase):
config_assert(input_layer.size % partial_sum == 0, config_assert(input_layer.size % partial_sum == 0,
"a wrong setting for partial_sum") "a wrong setting for partial_sum")
self.set_layer_size(input_layer.size) self.set_layer_size(input_layer.size)
self.config.partial_sum = partial_sum
self.create_input_parameter(0, input_layer.size / partial_sum) self.create_input_parameter(0, input_layer.size / partial_sum)
......
...@@ -14,6 +14,29 @@ layers { ...@@ -14,6 +14,29 @@ layers {
input_layer_name: "input" input_layer_name: "input"
input_parameter_name: "___prelu_layer_0__.w0" input_parameter_name: "___prelu_layer_0__.w0"
} }
partial_sum: 1
}
layers {
name: "__prelu_layer_1__"
type: "prelu"
size: 300
active_type: ""
inputs {
input_layer_name: "input"
input_parameter_name: "___prelu_layer_1__.w0"
}
partial_sum: 1
}
layers {
name: "__prelu_layer_2__"
type: "prelu"
size: 300
active_type: ""
inputs {
input_layer_name: "input"
input_parameter_name: "___prelu_layer_2__.w0"
}
partial_sum: 5
} }
parameters { parameters {
name: "___prelu_layer_0__.w0" name: "___prelu_layer_0__.w0"
...@@ -23,14 +46,32 @@ parameters { ...@@ -23,14 +46,32 @@ parameters {
initial_strategy: 0 initial_strategy: 0
initial_smart: true initial_smart: true
} }
parameters {
name: "___prelu_layer_1__.w0"
size: 300
initial_mean: 0.0
initial_std: 0.057735026919
initial_strategy: 0
initial_smart: true
}
parameters {
name: "___prelu_layer_2__.w0"
size: 60
initial_mean: 0.0
initial_std: 0.129099444874
initial_strategy: 0
initial_smart: true
}
input_layer_names: "input" input_layer_names: "input"
output_layer_names: "__prelu_layer_0__" output_layer_names: "__prelu_layer_2__"
sub_models { sub_models {
name: "root" name: "root"
layer_names: "input" layer_names: "input"
layer_names: "__prelu_layer_0__" layer_names: "__prelu_layer_0__"
layer_names: "__prelu_layer_1__"
layer_names: "__prelu_layer_2__"
input_layer_names: "input" input_layer_names: "input"
output_layer_names: "__prelu_layer_0__" output_layer_names: "__prelu_layer_2__"
is_recurrent_layer_group: false is_recurrent_layer_group: false
} }
...@@ -2,5 +2,7 @@ from paddle.trainer_config_helpers import * ...@@ -2,5 +2,7 @@ from paddle.trainer_config_helpers import *
data = data_layer(name='input', size=300) data = data_layer(name='input', size=300)
prelu = prelu_layer(input=data) prelu = prelu_layer(input=data)
prelu = prelu_layer(input=data, partial_sum=1)
prelu = prelu_layer(input=data, partial_sum=5)
outputs(prelu) outputs(prelu)
py_test(test_net SRCS test_net.py) file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
py_test(test_scope SRCS test_scope.py) foreach(src ${TEST_OPS})
py_test(${src} SRCS ${src}.py)
py_test(test_tensor SRCS test_tensor.py) endforeach()
py_test(test_mul_op SRCS test_mul_op.py)
py_test(test_cos_sim_op SRCS test_cos_sim_op.py)
py_test(test_mean_op SRCS test_mean_op.py)
py_test(test_protobuf SRCS test_protobuf.py)
py_test(test_add_two_op SRCS test_add_two_op.py)
py_test(test_sigmoid_op SRCS test_sigmoid_op.py)
py_test(test_softmax_op SRCS test_softmax_op.py)
py_test(test_cross_entropy_op SRCS test_cross_entropy_op.py)
py_test(test_gather_op SRCS test_gather_op.py)
py_test(test_scatter_op SRCS test_scatter_op.py)
py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py)
py_test(test_top_k_op SRCS test_top_k_op.py)
py_test(test_rowwise_add_op SRCS test_rowwise_add_op.py)
py_test(test_default_scope_funcs SRCS test_default_scope_funcs.py)
py_test(test_operator SRCS test_operator.py)
py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
py_test(test_recurrent_op SRCS test_recurrent_op.py)
py_test(test_sgd_op SRCS test_sgd_op.py)
py_test(test_gradient_checker SRCS test_gradient_checker.py)
py_test(test_lookup_table SRCS test_lookup_table.py)
py_test(test_scale_and_identity_op SRCS test_scale_and_identity_op.py)
py_test(test_sum_op SRCS test_sum_op.py)
py_test(mnist SRCS mnist.py)
py_test(test_concat_op SRCS test_concat_op.py)
py_test(test_squared_l2_distance_op SRCS test_squared_l2_distance_op.py)
py_test(test_reshape_op SRCS test_reshape_op.py)
import unittest
import numpy as np
from op_test import OpTest
class TestElementwiseMulOp_Matrix(OpTest):
def setUp(self):
self.op_type = "elementwise_mul"
""" Warning
CPU gradient check error!
'X': np.random.random((32,84)).astype("float32"),
'Y': np.random.random((32,84)).astype("float32")
"""
self.inputs = {
'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
}
self.outputs = {'Out': np.multiply(self.inputs['X'], self.inputs['Y'])}
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
class TestElementwiseMulOp_Vector(OpTest):
def setUp(self):
self.op_type = "elementwise_mul"
self.inputs = {
'X': np.random.random((32, )).astype("float32"),
'Y': np.random.random((32, )).astype("float32")
}
self.outputs = {'Out': np.multiply(self.inputs['X'], self.inputs['Y'])}
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
class TestElementwiseMulOp_broadcast_0(OpTest):
def setUp(self):
self.op_type = "elementwise_mul"
self.inputs = {
'X': np.random.rand(2, 3, 4).astype(np.float32),
'Y': np.random.rand(2).astype(np.float32)
}
self.attrs = {'axis': 0}
self.outputs = {
'Out': self.inputs['X'] * self.inputs['Y'].reshape(2, 1, 1)
}
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
class TestElementwiseMulOp_broadcast_1(OpTest):
def setUp(self):
self.op_type = "elementwise_mul"
self.inputs = {
'X': np.random.rand(2, 3, 4).astype(np.float32),
'Y': np.random.rand(3).astype(np.float32)
}
self.attrs = {'axis': 1}
self.outputs = {
'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 3, 1)
}
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
class TestElementwiseMulOp_broadcast_2(OpTest):
def setUp(self):
self.op_type = "elementwise_mul"
self.inputs = {
'X': np.random.rand(2, 3, 4).astype(np.float32),
'Y': np.random.rand(4).astype(np.float32)
}
self.outputs = {
'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 1, 4)
}
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
class TestElementwiseMulOp_broadcast_3(OpTest):
def setUp(self):
self.op_type = "elementwise_mul"
self.inputs = {
'X': np.random.rand(2, 3, 4, 5).astype(np.float32),
'Y': np.random.rand(3, 4).astype(np.float32)
}
self.attrs = {'axis': 1}
self.outputs = {
'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 3, 4, 1)
}
if __name__ == '__main__':
unittest.main()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册