tensorflow添加自定义的auc计算operator
2016-02-19 17:01
357 查看
tensorflow可以很方便的添加用户自定义的operator(如果不添加也可以采用sklearn的auc计算函数或者自己写一个
但是会在python执行,这里希望在graph中也就是c++端执行这个计算)
这里根据工作需要添加一个计算auc的operator,只给出最简单实现,后续高级功能还是参考官方wiki
https://www.tensorflow.org/versions/r0.7/how_tos/adding_an_op/index.html
注意tensorflow现在和最初的官方wiki有变化,原wiki貌似是需要重新bazel编译整个tensorflow,然后使用比如tf.user_op.auc这样。
目前wiki给出的方式>=0.6.0版本,采用plug-in的方式,更加灵活可以直接用g++编译一个so载入,解耦合,省去了编译tensorflow过程,即插即用。
首先auc的operator计算的文件
tensorflow/core/user_ops/auc.cc
/* Copyright 2015 Google Inc. All Rights Reserved.
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.
==============================================================================*/
// An auc Op.
#include
"tensorflow/core/framework/op.h"
#include
"tensorflow/core/framework/op_kernel.h"
using
namespace tensorflow;
using
std::vector;
//@TODO add weight as optional input
REGISTER_OP("Auc")
.Input("predicts: T1")
.Input("labels: T2")
.Output("z: float")
.Attr("T1: {float, double}")
.Attr("T2: {float, double}")
//.Attr("T1: {float, double}")
//.Attr("T2: {int32, int64}")
.SetIsCommutative()
.Doc(R"doc(
Given preidicts and labels output it's auc
)doc");
class
AucOp : public OpKernel {
public:
explicit
AucOp(OpKernelConstruction* context) : OpKernel(context) {}
template<typename
ValueVec>
void
index_sort(const
ValueVec& valueVec, vector<int>& indexVec)
{
indexVec.resize(valueVec.size());
for (size_t
i = 0; i < indexVec.size(); i++)
{
indexVec[i] = i;
}
std::sort(indexVec.begin(), indexVec.end(),
[&valueVec](const
int
l, const
int
r) { return
valueVec(l) > valueVec(r); });
}
void
Compute(OpKernelContext* context) override {
// Grab the input tensor
const Tensor& predicts_tensor = context->input(0);
const Tensor& labels_tensor = context->input(1);
auto
predicts = predicts_tensor.flat<float>(); //输入能接受float double那么这里如何都处理?
auto
labels = labels_tensor.flat<float>();
vector<int> indexes;
index_sort(predicts, indexes);
typedef
float
Float;
Float
oldFalsePos = 0;
Float
oldTruePos = 0;
Float
falsePos = 0;
Float
truePos = 0;
Float
oldOut = std::numeric_limits<Float>::infinity();
Float
result = 0;
for (size_t
i = 0; i < indexes.size(); i++)
{
int
index = indexes[i];
Float
label = labels(index);
Float
prediction = predicts(index);
Float
weight = 1.0;
//Pval3(label, output, weight);
if (prediction != oldOut) //存在相同值得情况是特殊处理的
{
result += 0.5 * (oldTruePos + truePos) * (falsePos - oldFalsePos);
oldOut = prediction;
oldFalsePos = falsePos;
oldTruePos = truePos;
}
if (label > 0)
truePos += weight;
else
falsePos += weight;
}
result += 0.5 * (oldTruePos + truePos) * (falsePos - oldFalsePos);
Float
AUC = result / (truePos * falsePos);
// Create an output tensor
Tensor* output_tensor = NULL;
TensorShape output_shape;
OP_REQUIRES_OK(context, context->allocate_output(0, output_shape, &output_tensor));
output_tensor->scalar<float>()() = AUC;
}
};
REGISTER_KERNEL_BUILDER(Name("Auc").Device(DEVICE_CPU), AucOp);
编译:
$cat gen-so.sh
TF_INC=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())')
TF_LIB=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_lib())')
i=$1
o=${i/.cc/.so}
g++ -std=c++11 -shared $i -o $o -I $TF_INC -l tensorflow_framework -L $TF_LIB -fPIC -Wl,-rpath $TF_LIB
$sh gen-so.sh auc.cc
会生成auc.so
使用的时候
auc_module = tf.load_op_library('auc.so')
#auc = tf.user_ops.auc #0.6.0之前的tensorflow 自定义op方式
auc = auc_module.auc
evaluate_op = auc(py_x, Y) #py_x is predicts, Y is labels
但是会在python执行,这里希望在graph中也就是c++端执行这个计算)
这里根据工作需要添加一个计算auc的operator,只给出最简单实现,后续高级功能还是参考官方wiki
https://www.tensorflow.org/versions/r0.7/how_tos/adding_an_op/index.html
注意tensorflow现在和最初的官方wiki有变化,原wiki貌似是需要重新bazel编译整个tensorflow,然后使用比如tf.user_op.auc这样。
目前wiki给出的方式>=0.6.0版本,采用plug-in的方式,更加灵活可以直接用g++编译一个so载入,解耦合,省去了编译tensorflow过程,即插即用。
首先auc的operator计算的文件
tensorflow/core/user_ops/auc.cc
/* Copyright 2015 Google Inc. All Rights Reserved.
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.
==============================================================================*/
// An auc Op.
#include
"tensorflow/core/framework/op.h"
#include
"tensorflow/core/framework/op_kernel.h"
using
namespace tensorflow;
using
std::vector;
//@TODO add weight as optional input
REGISTER_OP("Auc")
.Input("predicts: T1")
.Input("labels: T2")
.Output("z: float")
.Attr("T1: {float, double}")
.Attr("T2: {float, double}")
//.Attr("T1: {float, double}")
//.Attr("T2: {int32, int64}")
.SetIsCommutative()
.Doc(R"doc(
Given preidicts and labels output it's auc
)doc");
class
AucOp : public OpKernel {
public:
explicit
AucOp(OpKernelConstruction* context) : OpKernel(context) {}
template<typename
ValueVec>
void
index_sort(const
ValueVec& valueVec, vector<int>& indexVec)
{
indexVec.resize(valueVec.size());
for (size_t
i = 0; i < indexVec.size(); i++)
{
indexVec[i] = i;
}
std::sort(indexVec.begin(), indexVec.end(),
[&valueVec](const
int
l, const
int
r) { return
valueVec(l) > valueVec(r); });
}
void
Compute(OpKernelContext* context) override {
// Grab the input tensor
const Tensor& predicts_tensor = context->input(0);
const Tensor& labels_tensor = context->input(1);
auto
predicts = predicts_tensor.flat<float>(); //输入能接受float double那么这里如何都处理?
auto
labels = labels_tensor.flat<float>();
vector<int> indexes;
index_sort(predicts, indexes);
typedef
float
Float;
Float
oldFalsePos = 0;
Float
oldTruePos = 0;
Float
falsePos = 0;
Float
truePos = 0;
Float
oldOut = std::numeric_limits<Float>::infinity();
Float
result = 0;
for (size_t
i = 0; i < indexes.size(); i++)
{
int
index = indexes[i];
Float
label = labels(index);
Float
prediction = predicts(index);
Float
weight = 1.0;
//Pval3(label, output, weight);
if (prediction != oldOut) //存在相同值得情况是特殊处理的
{
result += 0.5 * (oldTruePos + truePos) * (falsePos - oldFalsePos);
oldOut = prediction;
oldFalsePos = falsePos;
oldTruePos = truePos;
}
if (label > 0)
truePos += weight;
else
falsePos += weight;
}
result += 0.5 * (oldTruePos + truePos) * (falsePos - oldFalsePos);
Float
AUC = result / (truePos * falsePos);
// Create an output tensor
Tensor* output_tensor = NULL;
TensorShape output_shape;
OP_REQUIRES_OK(context, context->allocate_output(0, output_shape, &output_tensor));
output_tensor->scalar<float>()() = AUC;
}
};
REGISTER_KERNEL_BUILDER(Name("Auc").Device(DEVICE_CPU), AucOp);
编译:
$cat gen-so.sh
TF_INC=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())')
TF_LIB=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_lib())')
i=$1
o=${i/.cc/.so}
g++ -std=c++11 -shared $i -o $o -I $TF_INC -l tensorflow_framework -L $TF_LIB -fPIC -Wl,-rpath $TF_LIB
$sh gen-so.sh auc.cc
会生成auc.so
使用的时候
auc_module = tf.load_op_library('auc.so')
#auc = tf.user_ops.auc #0.6.0之前的tensorflow 自定义op方式
auc = auc_module.auc
evaluate_op = auc(py_x, Y) #py_x is predicts, Y is labels
相关文章推荐
- 查看Linux进程CPU过高具体的线程堆栈(不中断程序)
- 查看Linux进程CPU过高具体的线程堆栈(不中断程序)
- linux上调试崩溃程序
- Apache Kafka:下一代分布式消息系统
- linux常用查看日志命令
- linux下使用fstat来计算文件的大小
- 去掉xcode编译warning:ld: warning: directory not found for option:xxxxxx 的方法
- centos 常用命令
- .net网站关于导入excel,vs可以成功发布网站失败的问题
- Linux文件权限详解
- 启动spark-shell提示scala版本不一致的解决
- Linux启动流程图解
- Linux下文件的切分与合并
- Docker安装
- linux 版本source insight
- 配置eclipse+tomcat注意点
- nginx的搭建-Linux上
- centos访问windows的共享文件(通过挂载)
- 常用技术网站
- shell脚本:批量修改文件名(删除文件名中字符)