提交 1ccd4d31 编写于 作者: C chenjiaoAngel 提交者: Shan Yi

Add Anakin Doc (#83)

* add anakin docker

* add

* add doc

*  delete docker

* Update index_anakin_ch.rst

* Update index_anakin_ch.rst

* Update index_anakin_ch.rst

* Update index_anakin_ch.rst

* Update index_anakin.rst

* Update anakin_parser_design_ch.md

* Update and rename anakin_parser_design_ch.md to anakin_parser_design.md

* Delete index_anakin_ch.rst

* Update anakin_parser_design.md

* update

* update

* Update anakin_parser_design.md

*  fix format

*  fix format

* fix format

* Delete .DS_Store

* Delete .DS_Store

* Delete menu.json

* Update anakin_parser_design.md

* change fluid to paddle
上级 9a8ce09c
......@@ -25,15 +25,15 @@
### <span id = '11'> mobilenetv1 </span>
|platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)|
|platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)|
|:---: | :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:|
|麒麟960|107.7ms|61.1ms|38.2ms|152.8ms|85.2ms|51.9ms|152.6ms|nan|nan|
|高通835|105.7ms|63.1ms|~~46.8ms~~|152.7ms|87.0ms|~~92.7ms~~|146.9ms|nan|nan|
|高通653|120.3ms|64.2ms|46.6ms|202.5ms|117.6ms|84.8ms|158.6ms|nan|nan|
|高通653|120.3ms|64.2ms|46.6ms|202.5ms|117.6ms|84.8ms|158.6ms|nan|nan|
### <span id = '22'> mobilenetv2 </span>
|platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)|
|platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)|
|:---: | :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:|
|麒麟960|93.1ms|53.9ms|34.8ms|144.4ms|84.3ms|55.3ms|100.6ms|nan|nan|
|高通835|93.0ms|55.6ms|41.1ms|139.1ms|88.4ms|58.1ms|95.2ms|nan|nan|
......@@ -41,7 +41,7 @@
### <span id = '33'> mobilenet-ssd </span>
|platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)|
|platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)|
|:---: | :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:|
|麒麟960|213.9ms|120.5ms|74.5ms|307.9ms|166.5ms|104.2ms|nan|nan|nan|
|高通835|213.0ms|125.7ms|~~98.4ms~~|292.9ms|177.9ms|~~167.8ms~~|nan|nan|nan|
......@@ -49,8 +49,8 @@
## How to run those Benchmark models?
1. 首先, 使用[External Converter](../docs/Manual/Converter_en.md)对caffe model 进行转换
2. 然后将转换后的Anakin model和编译好的benchmark_arm 二进制文件通过'adb push'命令上传至测试机
3. 接着在测试机含有Anakin model的目录中运行'./benchmark_arm ./ anakin_model.anakin.bin 1 10 10 1' 命令
4. 最后,终端显示器上将会打印该模型的运行时间
5. 其中运行命令的参数个数和含义可以通过运行'./benchmark_arm'看到
1. 首先, 使用[External Converter](./convert_paddle_to_anakin.md)对caffe model 进行转换
2. 然后将转换后的Anakin model和编译好的benchmark_arm 二进制文件通过'adb push'命令上传至测试机
3. 接着在测试机含有Anakin model的目录中运行'./benchmark_arm ./ anakin_model.anakin.bin 1 10 10 1' 命令
4. 最后,终端显示器上将会打印该模型的运行时间
5. 其中运行命令的参数个数和含义可以通过运行'./benchmark_arm'看到
# Example
# Anakin 运行模型示例
Anakin目前只支持NCHW的格式
示例文件在test/framework/net下
## 在NV的GPU上运行CNN模型
示例文件为打开example_nv_cnn_net.cpp,整体流程如下:
- 将模型的的path设置为anakin模型的路径,初始化NV平台的图对象。 anakin模型可以通过转换器转化caffe或fluid的模型得到
- 将模型的的path设置为anakin模型的路径,初始化NV平台的图对象。 anakin模型可以通过转换器转化caffe或Paddle的模型得到
- 根据模型设置网络图的输入尺寸,进行图优化
- 根据优化后的网络图初始化网络执行器
- 取出网络的输入tensor,将数据拷贝到输入tensor
......@@ -14,15 +18,21 @@
以NV平台为例演示Anakin框架的使用方法,注意编译时需要打开GPU编译开关
## 在X86上运行RNN模型
示例文件为example_x86_rnn_net.cpp
整体流程与在NV的GPU上运行CNN模型相似,不同之处如下:
- 使用X86标识初始化图对象和网络执行器对象
- rnn模型的输入尺寸是可变的,初始化图时的输入维度是维度的最大值,输入维度N代表总的词的个数。还需要设置输入tensor的seq_offset来标示这些词是如何划分为句子的,如{0,5,12}表示共有12个词,其中第0到第4个词是第一句话,第5到第11个词是第二句话
以X86平台为例演示Anakin框架的使用方法,注意编译时需要打开X86编译开关
## 在NV的GPU上使用Anakin的线程池运行CNN模型
示例文件为example_nv_cnn_net_multi_thread.cpp ,示例使用worker的同步预测接口
整体流程与在NV的GPU上运行CNN模型相似,不同之处如下:
- 用模型地址和线程池大小初始化worker对象
- 将输入tensor注入任务队列,获得输出tensor
# Anakin GPU Benchmark
# Anakin GPU 性能测试
## Machine:
## 环境:
> CPU: `12-core Intel(R) Xeon(R) CPU E5-2620 v2 @2.10GHz`
> GPU: `Tesla P4`
> cuDNN: `v7`
## Counterpart of anakin :
## anakin 对比对象:
The counterpart of **`Anakin`** is the acknowledged high performance inference engine **`NVIDIA TensorRT 3`** , The models which TensorRT 3 doesn't support we use the custom plugins to support.
**`Anakin`** 将与高性能的推理引擎 **`NVIDIA TensorRT 3`** 进行比较
## Benchmark Model
The following convolutional neural networks are tested with both `Anakin` and `TenorRT3`.
You can use pretrained caffe model or the model trained by youself.
> 注意在性能测试之前,请先将测试model通过 `External Converter` 工具转换为Anakin model
> 对这些model,本文在GPU上进行单线程单GPU卡的性能测试。
> Please note that you should transform caffe model or others into anakin model with the help of [`external converter ->`](../docs/Manual/Converter_en.md)
- [Vgg16](#1) *caffe model can be found [here->](https://gist.github.com/jimmie33/27c1c0a7736ba66c2395)*
- [Yolo](#2) *caffe model can be found [here->](https://github.com/hojel/caffe-yolo-model)*
- [Resnet50](#3) *caffe model can be found [here->](https://github.com/KaimingHe/deep-residual-networks#models)*
- [Resnet101](#4) *caffe model can be found [here->](https://github.com/KaimingHe/deep-residual-networks#models)*
- [Mobilenet v1](#5) *caffe model can be found [here->](https://github.com/shicai/MobileNet-Caffe)*
- [Mobilenet v2](#6) *caffe model can be found [here->](https://github.com/shicai/MobileNet-Caffe)*
- [RNN](#7) *not support yet*
We tested them on single-GPU with single-thread.
- [Vgg16](#1) *caffe model 可以在[这儿](https://gist.github.com/jimmie33/27c1c0a7736ba66c2395)下载*
- [Yolo](#2) *caffe model 可以在[这儿](https://github.com/hojel/caffe-yolo-model)下载*
- [Resnet50](#3) *caffe model 可以在[这儿](https://github.com/KaimingHe/deep-residual-networks#models)下载*
- [Resnet101](#4) *caffe model 可以在[这儿](https://github.com/KaimingHe/deep-residual-networks#models)下载*
- [Mobilenet v1](#5) *caffe model 可以在[这儿](https://github.com/shicai/MobileNet-Caffe)下载*
- [Mobilenet v2](#6) *caffe model 可以在[这儿](https://github.com/shicai/MobileNet-Caffe)下载*
- [RNN](#7) *暂不支持*
### <span id = '1'>VGG16 </span>
......@@ -162,9 +157,9 @@ We tested them on single-GPU with single-thread.
| 8 | 421 | 351 |
| 32 | 637 | 551 |
## How to run those Benchmark models?
## How to run those Benchmark models
> 1. At first, you should parse the caffe model with [`external converter`](https://github.com/PaddlePaddle/Anakin/blob/b95f31e19993a192e7428b4fcf852b9fe9860e5f/docs/Manual/Converter_en.md).
> 2. Switch to *source_root/benchmark/CNN* directory. Use 'mkdir ./models' to create ./models and put anakin models into this file.
> 3. Use command 'sh run.sh', we will create files in logs to save model log with different batch size. Finally, model latency summary will be displayed on the screen.
> 4. If you want to get more detailed information with op time, you can modify CMakeLists.txt with setting `ENABLE_OP_TIMER` to `YES`, then recompile and run. You will find detailed information in model log file.
1. 首先, 使用[External Converter](./convert_paddle_to_anakin.md)对caffe model 进行转换
2. 然后跳转至 *source_root/benchmark/CNN* 目录下,使用 'mkdir ./models'创建存放模型的目录,并将转换好的Anakin模型放在该目录下
3. 运行脚本 `sh run.sh`,运行结束后,该模型的运行时间将会显示到终端上
4. 如果你想获取每层OP的运行时间,你只用将 CMakeLists.txt 中的`ENABLE_OP_TIMER` 设置为 `YES` 即可
# Parser的编写指南
Parser是一种网络框架转换工具,将其他框架如Caffe、TensorFlow的网络结构转换为Anakin网络结构图,然后对转换后的Anakin图进行预测处理
本文主要介绍Parser功能的框架结构和根据已有的网络框架改写Parser,以解析得到Anakin框架图,进行Anakin预测
下文称Anakin为AK,运算操作为OP,本文参考TensorFlow的Parser编写,参考代码目录为tools/external_converter_v2/parser/tensorflow
## Parser的功能和执行流程
Parser功能是将其他深度学习框架(如Caffe,TensorFlow,ONNX)的模型转换为AK的模型
对AK的作用是屏蔽不同框架间的差异,这种差异包括模型存储、OP的定义、图差异
因此Parser的执行流程是:
- 将源框架的模型载入Parser
- 将原框架的图解析为AK中的OP节点和OP节点的连接关系
- 进行OP定义的转换和图优化
- 将符合AK标准的图写入protobuf
## Parser的目录结构
Parser工具在tools/external_converter_v2/parser目录下
Parser的目录主要包含3部分:
- Parser的运行配置文件包括 config.py, config.yaml, converter.py, 用户只用执行converter.py,Parser就会按照config.yaml中的声明去解析模型
- Parser的公共定义,包括operations,pbs,proto三个目录。Parser的公共工具函数 graph*.py logger.py utils.py
- 各个框架对应的Parser,其目录的命名方式为框架名,如Caffe, TensorFlow
## Parser的编写流程
### 1、声明你的Parser
- 在config.yaml中填写你的Parser运行的必要信息,包括ProtoPath和SavePath等。OPTIONS/Framework改为你的Parser的类型,TARGET下填写对应的参数列表
- 添加你的Parser目录,如TensorFlow,导出你的Parser符号。注意,Parser的框架默认调用你的Parser类中的__call__方法来执行解析,这个方法需要返回填写完毕的GraphProtoIO对象
- 在config.py中Configuration下__init__函数中增加对你的Parser的调用,将yaml中读取的配置信息传给你的Parser,此处调用你的Parser中的__init__方法
### 2、添加你的Parser主体
可以参考parser_tf.py
- 你需要在Parser主体构造时获取模型路径,input,ouput名字等解析必须的信息
- 在__call__中返回填写好的GraphProtoIO对象,该对象为填写protobuf的辅助工具
- 建议Parser的解析过程分成三部分,先将原框架的模型载入并转换为一种便于修改的中间的图形式;对中间图修改使得图满足AK的要求;将满足要求的中间图利用NodeProtoIO和GraphProtoIO这两个辅助类填入protobuf,具体细节可以参考parser_tf
### 3、读取原始模型,并将模型转换为中间类型
可以参考parse_tf_2_med.py
- 这一步与原始框架结合紧密,你可能需要import原始框架的工具函数来完成模型的裁剪、固定、加载等操作
- 大部分的框架都是使用tensor来连接OP的,但AK中是OP直接相连,这点需要注意
- AK的shape默认是4维的,有的参数的shape不足4维,需要Parser补全
### 4、对中间类型的图进行优化
可以参考med_graph.py
- 由于AK不支持普通OP多输出的情况,需要在多输出的OP后面补上Splite类型的OP节点
- 对于Convlution后接Batchnorm这种可以合并又不会导致OP定义改变的情况,需要Parser在这一步做掉
- AK规定所有的输入类型OP的名字必须是input_x这种命名方式,其中x为从0开始的数字
### 5、将中间类型的图以GraphProtoIO的方式保存
可以参考parse_med_2_ak.py 和 parser_tf.py
- 你首先需要构造Node节点,Node节点的名字是OP的名字(如conv2d_1_a_0),Node节点中OP成员变量的名字是Node节点的类型(如Convlution)
- Node节点需要按照输入的顺序用Node的add_in方法填写输入Node的名字,add_out方法按顺序填写输出Node的名字
- 通过调用GraphProtoIO的add_node方法将构造好的Node的__call__方法的返回值作为参数,将Node节点加入AK的graph中
- 调用GraphProtoIO的add_in_edge和add_out_edge完成AK图中OP间关系的构建。如果Node中的in和out填写正确,你也可以通过调用GraphProtoIO的format_edge_from_nodes方法完成这个工作
- AK的模型需要Parser给出输出Node的名字,使用GraphProtoIO的add_out方法填写输出Node的名字
### 6、检查模型解析的正确性
- 默认的config.yaml配置会在解析结束后启动一个web服务器展示解析后的AK模型图,你需要对比原框架的模型图进行验证。这里最容易出现的错误是边关系的错误,表现为图非常乱,你需要逐条边地检查错误;第二个容易出错的地方是参数漏填,需要你检查OP中的属性
- 将解析后的模型放入AK中执行,使用相同的输入,原框架与AK有相同的输出。若果输出不一致可以开启AK的DEBUG模式,在net.cpp中将没层的输出打印;如果AK在解析阶段陷入死循环,大概率是边的关系出错
## 如何添加新OP
- 需要在AK代码中加入该OP的实现,包括对应设备Saber的OP,Saber单测和Framework中的OP
- 根据Framework的OP在ops.py中添加Parser公共的OP定义
- 从原框架的模型中解析出该OP的节点,并在AK的graph中填入该OP节点
## AK模型与其他框架模型的不同之处
+ AK模型与caffe的模型相似,因此与其他模型有很多不同的地方,需要Parser在解析过程中处理掉
+ 最大的不同是与PaddlePaddle或TensorFlow的模型中OP粒度很细,而AK的模型中OP的粒度很粗(目的是为了节省访存开销)。这会导致解析这些框架的模型时存在大量的合并操作
+ 其次是OP的行为不同,如TensorFlow中Pooling默认都是exclusive的,而AK中是inclusive的。TensorFlow的Padding,如果是奇数pad,则在右方和下方多pad,而AK是在左方和上方多Pad
+ AK默认的布局是NCHW,如果其他框架的OP是其他形式的,需要在Parser中做weights的布局转换,并处理reshape的问题
+ AK中有的weights是需要预先做布局转换的(如GRU,LSTM),AK中也支持同一OP的不同算法,如(GRU,Pooling)
## ARM 源码编译 Anakin ##
目前Anakin支持ARM Android平台,采用Android NDK交叉编译工具链,已在mac os和centos上编译和测试通过。
### 安装概览 ###
* [系统需求](#0001)
* [安装第三方依赖](#0002)
* [Anakin源码编译](#0003)
* [验证安装](#0004)
### <span id = '0001'> 1. 系统需求 </span> ###
* 宿主机: linux, mac
* cmake 3.8.2+
* Android NDK r14, Linux 版本[从这里下载](https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip)
### <span id = '0002'> 2. 安装第三方依赖 </span> ###
- 2.1 protobuf3.4.0
源码从这里[下载](https://github.com/google/protobuf/releases/tag/v3.4.0)
- 2.1.1 为宿主机编译protobuf
```bash
$ tar -xzf protobuf-3.4.0.tar.gz
$ cd protobuf-3.4.0
$ ./autogen.sh
$ ./configure
$ make
$ make check
$ make install
```
上述 $make install 执行后,可在 `/usr/local/include/google` 找到 libprotobuf 所需的头文件,将整个google文件夹拷贝至Anakin/third-party/arm-android/protobuf/下, 然后将已经生成文件清除。
如有问题,请点[这里](https://github.com/google/protobuf/blob/v3.4.0/src/README.md)
```bash
$ make distclean
```
- 2.1.1 交叉编译Android`armeabi-v7a`的protobuf,注意设置ANDROID_NDK的路径,以及ARCH_ABI、HOSTOSN的值
```bash
$ export ANDROID_NDK=your_ndk_path
$ ARCH_ABI="arm-linux-androideabi-4.9"
$ HOSTOSN="darwin-x86_64"
$ export SYSROOT=$ANDROID_NDK/platforms/android-9/arch-arm
$ export PREBUILT=$ANDROID_NDK/toolchains/$ARCH_ABI
$ export LDFLAGS="--sysroot=$SYSROOT"
$ export LD="$ANDROID_NDK/toolchains/$ARCH_ABI/prebuilt/$HOSTOSN/arm-linux-androideabi/bin/ld $LDFLAGS"
$ export LIBS="-llog $ANDROID_NDK/sources/cxx-stl/gnu-libstdc++/4.9/libs/armeabi-v7a/libgnustl_static.a"
$ export CPPFLAGS=""
$ export INCLUDES="-I$ANDROID_NDK/sources/cxx-stl/gnu-libstdc++/4.9/include/ -I$ANDROID_NDK/platforms/android-9/arch-arm/usr/include/ -I$ANDROID_NDK/sources/cxx-stl/gnu-libstdc++/4.9/libs/armeabi-v7a/include/"
$ export CXXFLAGS="-march=armv7-a -mfloat-abi=softfp -DGOOGLE_PROTOBUF_NO_RTTI --sysroot=$SYSROOT"
$ export CCFLAGS="$CXXFLAGS"
$ export CXX="$PREBUILT/prebuilt/$HOSTOSN/bin/arm-linux-androideabi-g++ $CXXFLAGS"
$ export CC="$CXX"
$ export RANLIB="$ANDROID_NDK/toolchains/$ARCH_ABI/prebuilt/$HOSTOSN/bin/arm-linux-androideabi-ranlib"
$ ./autogen.sh
$ ./configure --host=arm-linux-androideabi --with-sysroot=$SYSROOT --enable-cross-compile --with-protoc=protoc --disable-shared CXX="$CXX" CC="$CC" LD="$LD"
$ make
```
编译生成 *.a 静态库,若希望编译*.so 动态链接库 ,请在./configure参数中改--disable-shared为--disable-static --enable-shared
生成文件在`src/.libs/`下,将生成的文件拷贝至`Anakin/third-party/arm-android/protobuf/lib`
[cmake](../../cmake/find_modules.cmake)中更新`ARM_RPOTO_ROOT`的路径。
```cmake
set(ARM_RPOTO_ROOT "${CMAKE_SOURCE_DIR}/third-party/arm-android/protobuf")
```
- 2.2 opencv 2.4.3+(optional)
Anakin只在examples示例中使用opencv
Android系统的opencv从[这里下载](https://opencv.org/releases.html)
解压后将 `3rdparty/libs/armeabi-v7a`中的库文件拷贝到`libs/armeabi-v7a`
在[cmake](../../cmake/find_modules.cmake)中搜索`anakin_find_opencv`
并设置 `include_directories` 和 `LINK_DIRECTORIES`为自己安装的库的路径
```cmake
include_directories(${CMAKE_SOURCE_DIR}/third-party/arm-android/opencv/sdk/native/jni/include/)
LINK_DIRECTORIES(${CMAKE_SOURCE_DIR}/third-party/arm-android/opencv/sdk/native/libs/armeabi-v7a/)
```
### <span id = '0003'> 3. Anakin源码编译 </span> ###
#### 编译Android版本
克隆[源码](https://github.com/PaddlePaddle/Anakin/tree/arm)
```bash
cd your_dir
git clone https://github.com/PaddlePaddle/Anakin.git
cd Anakin
git fetch origin arm
git checkout arm
```
修改`android_build.sh`
- 修改NDK路径
```bash
#modify "your_ndk_path" to your NDK path
export ANDROID_NDK=your_ndk_path
```
- 修改ARM 处理器架构
对于32位ARM处理器, 将ANDROID_ABI 设置为 `armeabi-v7a with NEON`
对于64位ARM处理器, 可以将ANDROID_ABI 设置为 `armeabi-v7a with NEON`或者`arm64-v8a`
目前我们只支持 `armeabi-v7a with NEON`;`arm64-v8a` 还在开发中
```bash
-DANDROID_ABI="armeabi-v7a with NEON"
```
- 设置Android API
根据Android系统的版本设置API level, 例如API Level 21 -> Android 5.0.1
```bash
-DANDROID_NATIVE_API_LEVEL=21
```
- 选择编译静态库或动态库
设置`BUILD_SHARED=NO`编译静态库
设置`BUILD_SHARED=YES`编译动态库
```bash
-DBUILD_SHARED=NO
```
- OpenMP多线程支持
设置`USE_OPENMP=YES`开启OpenMP多线程
```bash
-DUSE_OPENMP=YES
```
- 编译单测文件
设置`BUILD_WITH_UNIT_TEST=YES`将会编译单测文件
```bash
-DBUILD_WITH_UNIT_TEST=YES
```
- 编译示例文件
设置`BUILD_EXAMPLES=YES`将会编译示例文件
```bash
-DBUILD_EXAMPLES=YES
```
- 开启opencv
如果使用opencv,设置`USE_OPENCV=YES`
```bash
-DUSE_OPENCV=YES
```
- 开始编译
运行脚本 `android_build.sh` 将自动编译Anakin
```bash
./android_build.sh
```
### <span id = '0004'> 4. 验证安装 </span> ###
编译好的库会放在目录`${Anakin_root}/output`下;
编译好的单测文件会放在`${Anakin_root}/output/unit_test`目录下;
编译好的示例文件会放在`${Anakin_root}/output/examples`目录下。
对于Android系统,打开设备的调试模式,通过ADB可以访问的目录是`data/local/tmp`,通过ADB push将测试文件、模型和数据发送到设备目录, 运行测试文件。
# Anakin 使用教程 ##
本教程将会简略的介绍Anakin的工作原理,一些基本的Anakin API,以及如何调用这些API。
## 内容 ###
- [Anakin的工作原理](#principle)
......@@ -14,31 +14,38 @@
用Anakin来进行前向计算主要分为三个步骤:
- 将外部模型通过[Anakin Parser](Converter_ch.md)解析为Anakin模型
在使用Anakin之前,用户必须将所有其他模型转换成Anakin模型,我们提供了转换脚本,用户可通过[Anakin Parser](Converter_ch.md)进行模型转换。
- 生成Anakin计算图
加载Anakin模型生成原始计算图,然后需要对原始计算图进行优化。你只需要调用相应的API优化即可。
- 执行计算图
Anakin会选择不同硬件平台执行计算图。
- 将外部模型通过[Anakin Parser](Converter_ch.md)解析为Anakin模型
在使用Anakin之前,用户必须将所有其他模型转换成Anakin模型,我们提供了转换脚本,用户可通过[Anakin Parser](Converter_ch.md)进行模型转换。
- 生成Anakin计算图
加载Anakin模型生成原始计算图,然后需要对原始计算图进行优化。你只需要调用相应的API优化即可。
- 执行计算图
Anakin会选择不同硬件平台执行计算图。
## <span id ='api'>Anakin APIs </span> ###
### Tensor ####
`Tensor`提供基础的数据操作和管理,为ops提供统一的数据接口。`Tensor`包含以下几个属性:
`Tensor`提供基础的数据操作和管理,为ops提供统一的数据接口。`Tensor`包含以下几个属性:
- Buffer
数据存储区
- Shape
数据的维度信息
- Event
用于异步计算的同步
- Buffer
数据存储区
- Shape
数据的维度信息
- Event
用于异步计算的同步
`Tensor`类包含三个`Shape`对象, 分别是`_shape`, `_valid_shape``offset`
`Tensor` 类包含三个`Shape`对象, 分别是`_shape`, `_valid_shape``offset``_shape``tensor`真正空间信息,`_valid_shape`表示当前`tensor`使用的空间信息, `_offset`表示当前`tensor`数据指针相对于真正数据空间的信息。 `Tensor`不同维度与分别与数学中的向量、矩阵等相对应如下表所示。
- `_shape``tensor`真正空间信息
- `_valid_shape`表示当前`tensor`使用的空间信息
- `tensor`使用的空间信息
- `_offset`表示当前`tensor`数据指针相对于真正数据空间的信息
`Tensor`不同维度与分别与数学中的向量、矩阵等相对应如下表所示
Dimentions | Math entity |
:----: | :----:
:----: | :----:
1 | vector
2 | matrix
3 | 3-tensor
......@@ -57,195 +64,202 @@ n | n-tensor
};
```
TargetType是平台类型,如X86,GPU等等,在Anakin内部有相应的标识与之对应;datatype是普通的数据类型,在Anakin内部也有相应的标志与之对应[LayOutType](#layout)是数据分布类型,如batch x channel x height x width [NxCxHxW], 在Anakin内部用一个struct来标识。 Anakin中数据类型与基本数据类型的对应如下:
TargetType是平台类型,如X86,GPU等等,在Anakin内部有相应的标识与之对应;datatype是普通的数据类型,在Anakin内部也有相应的标志与之对应
1. <span id='target'>TargetType</sapn>
[LayOutType](#layout)是数据分布类型,如batch x channel x height x width [NxCxHxW], 在Anakin内部用一个struct来标识
Anakin TargetType | platform
:----: | :----:|
NV | NVIDIA GPU
ARM | ARM
AMD | AMD GPU
X86 | X86
NVHX86 | NVIDIA GPU with Pinned Memory
Anakin中数据类型与基本数据类型的对应如下:
2. <sapn id='datatype'>DataType</span>
1. <span id = 'target'> TargetType </span>
Anakin DataType | C++ | Description
:---: | :---: | :---: |
AK_HALF | short | fp16
AK_FLOAT | float | fp32
AK_DOUBLE | double | fp64
AK_INT8 | char | int8
AK_INT16 | short | int16
AK_INT32 | int | int32
AK_INT64 | long | int64
AK_UINT8 | unsigned char | uint8
AK_UINT16 | unsigned short | uint8
AK_UINT32 | unsigned int | uint32
AK_STRING | std::string | /
AK_BOOL | bool | /
AK_SHAPE | / | Anakin Shape
AK_TENSOR | / | Anakin Tensor
Anakin TargetType | platform
:----: | :----:
NV | NVIDIA GPU
ARM | ARM
AMD | AMD GPU
X86 | X86
NVHX86 | NVIDIA GPU with Pinned Memory
2. <sapn id='datatype'> DataType </span>
3. <span id = 'layout'>LayOutType </span>
Anakin DataType | C++ | Description
:---: | :---: | :---:
AK_HALF | short | fp16
AK_FLOAT | float | fp32
AK_DOUBLE | double | fp64
AK_INT8 | char | int8
AK_INT16 | short | int16
AK_INT32 | int | int32
AK_INT64 | long | int64
AK_UINT8 | unsigned char | uint8
AK_UINT16 | unsigned short | uint8
AK_UINT32 | unsigned int | uint32
AK_STRING | std::string | /
AK_BOOL | bool | /
AK_SHAPE | / | Anakin Shape
AK_TENSOR | / | Anakin Tensor
Anakin LayOutType ( Tensor LayOut ) | Tensor Dimention | Tensor Support | Op Support
:---: | :---: | :---: | :---: |
W | 1-D | YES | NO
HW | 2-D | YES | NO
WH | 2-D | YES | NO
NW | 2-D | YES | YES
NHW | 3-D | YES |YES
NCHW ( default ) | 4-D | YES | YES
NHWC | 4-D | YES | NO
NCHW_C4 | 5-D | YES | YES
3. <span id = 'layout'> LayOutType </span>
Anakin LayOutType ( Tensor LayOut ) | Tensor Dimention | Tensor Support | Op Support
:---: | :---: | :---: | :---:
W | 1-D | YES | NO
HW | 2-D | YES | NO
WH | 2-D | YES | NO
NW | 2-D | YES | YES
NHW | 3-D | YES |YES
NCHW ( default ) | 4-D | YES | YES
NHWC | 4-D | YES | NO
NCHW_C4 | 5-D | YES | YES
理论上,Anakin支持申明1维以上的tensor,但是对于Anakin中的Op来说,只支持NW、NHW、NCHW、NCHW_C4这四种LayOut,其中NCHW是默认的LayOutType,NCHW_C4是专门针对于int8这种数据类型的。
理论上,Anakin支持申明1维以上的tensor,但是对于Anakin中的Op来说,只支持NW、NHW、NCHW、NCHW_C4这四种LayOut,其中NCHW是默认的LayOuteType,NCHW_C4是专门针对于int8这种数据类型的。
例子
例子
下面的代码将展示如何使用tensor, 我们建议先看看这些示例。
> 下面的代码将展示如何使用tensor, 我们建议先看看这些示例。
要想获得更多关于tensor的信息, 请参考 *soure_path/core/tensor.h*
> 要想获得更多关于tensor的信息, 请参考 *soure_path/core/tensor.h*
> 1. 使用shape对象初始化tensor
> 1. 使用shape对象初始化tensor
``` c++
//create a null tensor. A null tensor holds for nothing.
//tensor's buffer is resident at CPU and its datatype is AK_FLOAT.
//tensor's Layout is NCHW(default)
Tensor<X86, AK_FLOAT> mytensor;
```c++
//create a null tensor. A null tensor holds for nothing.
//tensor's buffer is resident at CPU and its datatype is AK_FLOAT.
//tensor's Layout is NCHW(default)
Tensor<X86, AK_FLOAT> mytensor;
//1. using shape object to create a tensor.
Shape shape1(NUM); //1-D shape. NUM is the number of dimention.
Tensor<X86, AK_FLOAT, W> mytensor1(shape1); //1-D tensor.
//1. using shape object to create a tensor.
Shape shape1(NUM); //1-D shape. NUM is the number of dimention.
Tensor<X86, AK_FLOAT, W> mytensor1(shape1); //1-D tensor.
// A 4-D shape
Shape shape2(N, C, H, W); // batch x channel x height x width
```
// A 4-D shape
Shape shape2(N, C, H, W); // batch x channel x height x width
```
>`注意:Shape的维度必须和tensor的`[LayoutType](#layout)`相同,比如Shape(N,C,H,W), 那么Tensor的 LayoutType必须是NCHW,否则会出错。如下列代码所示`
>`注意:Shape的维度必须和tensor的`[LayoutType](#layout)`相同,比如Shape(N,C,H,W), 那么Tensor的 LayoutType必须是NCHW,否则会出错。如下列代码所示`
```c++
// A 4-D tensor.
Tensor<X86, AK_FLOAT> mytensor2(shape2); //right
```c++
// A 4-D tensor.
Tensor<X86, AK_FLOAT> mytensor2(shape2); //right
//A 4-D tensor which is resident at GPU and its datatype is AK_INT8
Tensor<NV, AK_INT8> mytensor3(shape2); //right
//A 4-D tensor which is resident at GPU and its datatype is AK_INT8
Tensor<NV, AK_INT8> mytensor3(shape2); //right
Tensor<X86, AK_FLOAT, NHW> mytensor4(shape2); //wrong!! shape's dimetion must be equal to tensor's Layout.
Tensor<NV, AK_FLOAT, NCHW_C4> mytensor5(shape2); //wrong!!!!
Tensor<X86, AK_FLOAT, NHW> mytensor4(shape2); //wrong!! shape's dimetion must be equal to tensor's Layout.
Tensor<NV, AK_FLOAT, NCHW_C4> mytensor5(shape2); //wrong!!!!
```
```
> 2. 使用现有的数据和shape初始化tensor
> 2. 使用现有的数据和shape初始化tensor
```c++
```c++
/**
* A construtor of Tensor.
* data_ptr is a pointer to any data type of data
* TargetType is type of a platform [Anakin TargetType]
* id : device id
* shape: a Anakin shape
*/
Tensor(Dtype* data_ptr, TargetType_t target, int id, Shape shape);
/**
* A construtor of Tensor.
* data_ptr is a pointer to any data type of data
* TargetType is type of a platform [Anakin TargetType]
* id : device id
* shape: a Anakin shape
*/
Tensor(Dtype* data_ptr, TargetType_t target, int id, Shape shape);
//using existing data feed to a tensor
Tensor<X86, AK_FLOAT> mytensor(data_ptr, TargetType, device_id, shape); //shape must has dimention (N, C, H, W).
//using existing data feed to a tensor
Tensor<X86, AK_FLOAT> mytensor(data_ptr, TargetType, device_id, shape); //shape must has dimention (N, C, H, W).
```
```
> 3. 使用tensor初始化tensor
> 3. 使用tensor初始化tensor
```c++
Tensor<NV, AK_FLOAT> tensor(exist_tensor);
```
> 提示: 你可以用` typedef Tensor<X86, AK_FLOAT> Tensor4d_X86 `方便定义tensor
```c++
Tensor<NV, AK_FLOAT> tensor(exist_tensor);
```
> 提示: 你可以用` typedef Tensor<X86, AK_FLOAT> Tensor4d_X86 `方便定义tensor
#### 填充tensor数据区
填充数据区得看你申明tensor的方式, 下面展示了如何填充tensor的数据区。
```c++
首先来看看tensor的四种声明方式:
1. Tensor<X86, AK_FLOAT> mytensor;
2. Tensor<X86, AK_FLOAT, W> mytensor1(shape1);
3. Tensor<X86, AK_FLOAT> mytensor(data_ptr, TargetType, device_id, shape);
4. Tensor<NV, AK_FLOAT> tensor(exist_tensor);
```c++
1. Tensor<X86, AK_FLOAT> mytensor;
2. Tensor<X86, AK_FLOAT, W> mytensor1(shape1);
3. Tensor<X86, AK_FLOAT> mytensor(data_ptr, TargetType, device_id, shape);
4. Tensor<NV, AK_FLOAT> tensor(exist_tensor);
```
相关的声明方式的数据填充方法如下:
1:声明一个空的tensor,此时没有为其分配内存,所以,我们需要手动的为其分配内存。
//parama shape
mytensor.re_alloc(Shape shape);
//Get writable pointer to mytensor.
//parama index (int): where you start to write.
//Dtype is your data type such int, float or double.
Dtype *p = mytensor.mutable_data(index/*=0*/);
//write data to mytensor
for(int i = 0; i < mytensor.size(); i++){
p[i] = 1.0f;
}
//do something ...
2: 这种声明方式会自动分配内存
//Get writable pointer to mytensor.
//parama index (int): where you start to write.
//Dtype is your data type such int, float or double.
Dtype *p = mytensor1.mutable_data(index/*=0*/);
//write data to mytensor
for(int i = 0; i < mytensor.size(); i++){
p[i] = 1.0f;
}
//do something ...
3:在该种声明方式中,我们仍不需要手动为其分配内存。但在构造函数内部是否为其分配内存,得依情况而定。如果data_ptr和申明的
tensor都在都一个目标平台上,那么该tensor就会与data_ptr共享内存空间,相反,如果他们不在同一个平台上(如data_ptrX86上,而
tensorGPU上),那么此时tensor就会开辟一个新的内存空间,并将data_ptr所指向的数据拷贝到tensorbuffer中。
//Get writable pointer to mytensor.
//parama index (int): where you start to write.
//Dtype is your data type such int, float or double.
Dtype *p = mytensor.mutable_data(index/*=0*/);
//write data to mytensor
for(int i = 0; i < mytensor.size(); i++){
p[i] = 1.0f;
}
//do something ...
- 声明一个空的tensor,此时没有为其分配内存,所以,我们需要手动的为其分配内存。
4:该种方式仍不需要手动分配内存
```c++
//Get writable pointer to mytensor.
//parama index (int): where you start to write.
//Dtype is your data type such int, float or double.
Dtype *p = mytensor.mutable_data(index/*=0*/);
//write data to mytensor
for(int i = 0; i < mytensor.size(); i++){
//parama shape
mytensor.re_alloc(Shape shape);
//Get writable pointer to mytensor.
//parama index (int): where you start to write.
//Dtype is your data type such int, float or double.
Dtype *p = mytensor.mutable_data(index/*=0*/);
//write data to mytensor
for(int i = 0; i < mytensor.size(); i++){
p[i] = 1.0f;
}
//do something ...
}
//do something ...
```
- 这种声明方式会自动分配内存
```c++
//Get writable pointer to mytensor.
//parama index (int): where you start to write.
//Dtype is your data type such int, float or double.
Dtype *p = mytensor1.mutable_data(index/*=0*/);
//write data to mytensor
for(int i = 0; i < mytensor.size(); i++){
p[i] = 1.0f;
}
//do something ...
```
- 在该种声明方式中,我们仍不需要手动为其分配内存。但在构造函数内部是否为其分配内存,得依情况而定。如果data_ptr和申明的
tensor都在都一个目标平台上,那么该tensor就会与data_ptr共享内存空间,相反,如果他们不在同一个平台上(如data_ptr在X86上,而
tensor在GPU上),那么此时tensor就会开辟一个新的内存空间,并将data_ptr所指向的数据拷贝到tensor的buffer中。
```c++
//Get writable pointer to mytensor.
//parama index (int): where you start to write.
//Dtype is your data type such int, float or double.
Dtype *p = mytensor.mutable_data(index/*=0*/);
//write data to mytensor
for(int i = 0; i < mytensor.size(); i++){
p[i] = 1.0f;
}
//do something ...
```
- 该种方式仍不需要手动分配内存
```c++
//Get writable pointer to mytensor.
//parama index (int): where you start to write.
//Dtype is your data type such int, float or double.
Dtype *p = mytensor.mutable_data(index/*=0*/);
//write data to mytensor
for(int i = 0; i < mytensor.size(); i++){
p[i] = 1.0f;
}
//do something ...
```
- 另外,你还可以获取一个tensor的可读指针,示例如下:
另外,你还可以获取一个tensor的可读指针,示例如下:
```c++
//Get read-only pointer to mytensor.
//parama index (int): where you start to read.
//Dtype is your data type such int, float or double.
Dtype *p = mytensor.data(index/*=0*/);
Dtype *p = mytensor.data(index/*=0*/);
//do something ...
```
......@@ -254,77 +268,75 @@ tensor在GPU上),那么此时tensor就会开辟一个新的内存空间,
#### 获取tensor的shape
```c++
//some declarations
// ...
Shape shape = mytensor.shape();
//some declarations
// ...
Shape shape = mytensor.shape();
//Get a first dimetion size of tesor, if it has.
int d1 = shape[0];
//Get a first dimetion size of tesor, if it has.
int d1 = shape[0];
//Get a second dimention size of tensor, if it has.
int d2 = shape[1];
//Get a second dimention size of tensor, if it has.
int d2 = shape[1];
...
...
//Get a n-th dimention size of tensor, if it has.
int dn = shape[n-1];
//Get a n-th dimention size of tensor, if it has.
int dn = shape[n-1];
//Get a tensor's dimention
int dims = mytensor.dims();
//Get a tensor's dimention
int dims = mytensor.dims();
//Get the size of tensor.
//size = d1 x d2 x ... x dn.
int size = mytensor.size();
//Get the size of tensor.
//size = d1 x d2 x ... x dn.
int size = mytensor.size();
//Get the size of tensor at interval [Di, Dj)
// form i-th dimention to j-th dimention, but not including the j-th dimention.
// which means di x (di+1) x ... x (dj -1)
int size = mytensor.count(start, end);
//Get the size of tensor at interval [Di, Dj)
// form i-th dimention to j-th dimention, but not including the j-th dimention.
// which means di x (di+1) x ... x (dj -1)
int size = mytensor.count(start, end);
```
#### 设置tensor的shape
我们可以用tensor的成员函数set_shape来设置tensor的shape。 下面是set_shape的定义
```c++
/**
* \brief set a tensor's shape
* \param valid_shape [a Shape object]
* \param shape [a Shape object]
* \param offset [a Shape object]
* \return the status of this operation, that means whether it success * or not.
*/
SaberStatus set_shape(Shape valid_shape, Shape shape = Shape::zero(TensorAPI::layout_dims::value), Shape offset = Shape::minusone(TensorAPI::layout_dims::value));
/**
* \brief set a tensor's shape
* \param valid_shape [a Shape object]
* \param shape [a Shape object]
* \param offset [a Shape object]
* \return the status of this operation, that means whether it success * or not.
*/
SaberStatus set_shape(Shape valid_shape, Shape shape = Shape::zero(TensorAPI::layout_dims::value), Shape offset = Shape::minusone(TensorAPI::layout_dims::value));
```
这个成员函数只设置tensor的shape。这些shape对象(valid_shape, shape, offset)的[LayOutType](#layout)必须和当前的tensor的相应三个shape对象的LayOutType相同,如果不同就会出错,返回SaberInvalidValue。 如果相同,那么将成功设置tensor的shape。
```c++
// some declarations
// ...
//valid_shape, shape , offset are Shape object;
//All these Shape object's LayOutType must be equal to mytensor's.
mytensor.set_shape(valid_shape, shape, offset);
// some declarations
// ...
//valid_shape, shape , offset are Shape object;
//All these Shape object's LayOutType must be equal to mytensor's.
mytensor.set_shape(valid_shape, shape, offset);
```
#### 重置 tensor的shape
```c++
//some declarations
Shape shape, valid_shape, offset;
//some declarations
Shape shape, valid_shape, offset;
//do some initializations
...
mytensor.reshape(valid_shape, shape, offset);
//do some initializations
...
mytensor.reshape(valid_shape, shape, offset);
```
注意: Reshape操作仍然需要shape的[LayOutType](#layout) 与tensor的相同
### Graph ###
`Graph`类负责加载Anakin模型生成计算图、对图进行优化、存储模型等操作。
......@@ -335,62 +347,61 @@ mytensor.reshape(valid_shape, shape, offset);
```c++
template<typename TargetType, DataType Dtype, Precision Ptype>
class Graph ... /* inherit other class*/{
//some implements
...
template<typename TargetType, DataType Dtype, Precision Ptype>
class Graph ... /* inherit other class*/{
};
//some implements
...
};
```
前面已经介绍过[TargetType](#target)[DataType](#datatype)是Anakin内部自定义数据类型。[TargetType](#target)表示平台类型 (如NV、X86), [DataType](#datatype)是Anakin基本数据类型与C++/C中的基本数据类型相对应。 [Precision](#precision)为op所支持的精度类型, 稍后我们在介绍它。
```c++
//Create a empty graph object.
Graph graph = Graph<NV, AK_FLOAT, Precision::FP32> tmp();
//Create a empty graph object.
Graph graph = Graph<NV, AK_FLOAT, Precision::FP32> tmp();
//Create a pointer to a empty graph.
Graph *graph = new Graph<NV, AK_FLOAT, Precision::FP32>();
//Create a pointer to a empty graph.
Graph *graph = new Graph<NV, AK_FLOAT, Precision::FP32>();
//Create a pointer to a empty graph.
auto graph = new Graph<NV, AK_FLOAT, Precision::FP32>();
//Create a pointer to a empty graph.
auto graph = new Graph<NV, AK_FLOAT, Precision::FP32>();
```
#### 加载 Anakin 模型
```c++
//some declarations
...
auto graph = new Graph<NV, AK_FLOAT, Precision::FP32>();
std::string model_path = "the/path/to/where/your/models/are";
const char *model_path1 = "the/path/to/where/your/models/are";
//Loading Anakin model to generate a compute graph.
auto status = graph->load(model_path);
//Or this way.
auto status = graph->load(model_path1);
//Check whether load operation success.
if(!status){
std::cout << "error" << endl;
//do something...
}
//some declarations
...
auto graph = new Graph<NV, AK_FLOAT, Precision::FP32>();
std::string model_path = "the/path/to/where/your/models/are";
const char *model_path1 = "the/path/to/where/your/models/are";
//Loading Anakin model to generate a compute graph.
auto status = graph->load(model_path);
//Or this way.
auto status = graph->load(model_path1);
//Check whether load operation success.
if(!status){
std::cout << "error" << endl;
//do something...
}
```
#### 优化计算图
```c++
//some declarations
...
//Load graph.
...
//According to the ops of loaded graph, optimize compute graph.
graph->Optimize();
//some declarations
...
//Load graph.
...
//According to the ops of loaded graph, optimize compute graph.
graph->Optimize();
```
......@@ -400,34 +411,33 @@ graph->Optimize();
你可以在任何时候保存模型, 特别的, 你可以保存一个优化的模型,这样,下次再加载模型时,就不必进行优化操作。
```c++
//some declarations
...
//Load graph.
...
// save a model
//save_model_path: the path to where your model is.
auto status = graph->save(save_model_path);
//Checking
if(!status){
cout << "error" << endl;
//do somethin...
}
//some declarations
...
//Load graph.
...
// save a model
//save_model_path: the path to where your model is.
auto status = graph->save(save_model_path);
//Checking
if(!status){
cout << "error" << endl;
//do somethin...
}
```
#### 重新设置计算图里的tensor的shape
```c++
//some declarations
...
//Load graph.
...
vector<int> shape{10, 256, 256, 10};
//input_name : std::string.
//Reshape a tensor named input_name.
graph->Reshape(input_name, shape);//Note: shape is a vector, not a Shape object.
//some declarations
...
//Load graph.
...
vector<int> shape{10, 256, 256, 10};
//input_name : std::string.
//Reshape a tensor named input_name.
graph->Reshape(input_name, shape);//Note: shape is a vector, not a Shape object.
```
#### 设置 batch size
......@@ -435,14 +445,14 @@ graph->Reshape(input_name, shape);//Note: shape is a vector, not a Shape object.
`Graph` 支持重新设置batch size的大小。
```c++
//some declarations
...
//Load graph.
...
//input_name : std::string.
//Reset a tensor named input_name.
int new_batch_size = 4;
graph->ResetBatchSize(input_name, new_batch_size);
//some declarations
...
//Load graph.
...
//input_name : std::string.
//Reset a tensor named input_name.
int new_batch_size = 4;
graph->ResetBatchSize(input_name, new_batch_size);
```
### Net ###
......@@ -451,126 +461,122 @@ graph->ResetBatchSize(input_name, new_batch_size);
`Net` 是计算图的执行器。你可以通过Net对象获得输入和输出
#### Creating a graph executor
`Net`接受四个模板参数。
`Net`接受四个模板参数。
```c++
template<typename TargetType, DataType Dtype, Precision PType OpRunType RunType = OpRunType::ASYNC>
class Net{
//some implements
...
template<typename TargetType, DataType Dtype, Precision PType OpRunType RunType = OpRunType::ASYNC>
class Net{
//some implements
...
};
};
```
由于有些Op可能支持多种精度,我们可以通过Precision来指定。OpRunType表示同步或异步类型,异步是默认类型。OpRunType::SYNC表示同步,在GPU上只有单个流;OpRunType::ASYNC表示异步,在GPU上有多个流并以异步方式执行。实际上,Precision和OpRunType都是enum class, 详细设计请参考*source_root/framework/core/types.h*.
1. <span id = 'precision'> Precision </span>
Precision | Op support
:---: | :---:
Precision::INT4 | NO
Precision::INT8 | NO
Precision::FP16 | NO
Precision::FP32 | YES
Precision::FP64 | NO
Precision | Op support
:---: | :---:
Precision::INT4 | NO
Precision::INT8 | NO
Precision::FP16 | NO
Precision::FP32 | YES
Precision::FP64 | NO
现在Op的精度只支持FP32, 但在将来我们会支持剩下的Precision.
2. <span id = '1'> OpRunType </span>
OpRunType | Sync/Aync |Description
:---: | :---: | :---:
OpRunType::SYNC | Synchronization | single-stream on GPU
OpRunType::ASYNC | Asynchronization | multi-stream on GPU
2. OpRunType
用graph对象创建一个执行器
OpRunType | Sync/Aync |Description
:---: | :---: | :---:
OpRunType::SYNC | Synchronization | single-stream on GPU
OpRunType::ASYNC | Asynchronization | multi-stream on GPU
用graph对象创建一个执行器。
```c++
//some declarations
...
//Create a pointer to a graph.
auto graph = new Graph<NV, AK_FLOAT, Precision::FP32>();
//do something...
...
//some declarations
...
//Create a pointer to a graph.
auto graph = new Graph<NV, AK_FLOAT, Precision::FP32>();
//do something...
...
//create a executor
Net<NV, AK_FLOAT, Precision::FP32> executor(*graph);
//create a executor
Net<NV, AK_FLOAT, Precision::FP32> executor(*graph);
```
#### 获取输入输出tensor
获取输入输出tensor,并填充输入tensor的buffer。如果想要获取输入和输出tensor,那么必须指定输入的名字,如"input_0", "input_1", "input_2", ..., 必须传入如上字符串才能够获得输入tensor。另外,如果想知道input_i对应哪个输入,你需要去dash board查看,如何使用dash board请看[Anakin Parser](Converter_ch.md)。请看如下示例代码
```c++
//some declaratinos
...
//create a executor
//TargetType is NV [NVIDIA GPU]
Net<NV, AK_FLOAT, Precision::FP32> executor(*graph);
//Get the first input tensor.
//The following tensors(tensor_in0, tensor_in2 ...) are resident at GPU.
//Note: Member function get_in returns an pointer to tensor.
Tensor<NV, AK_FLOAT>* tensor_in0 = executor.get_in("input_0");
//If you have multiple input tensors
//You just type this code below.
Tensor<NV, AK_FLOAT>* tensor_in1 = executor.get_in("input_1");
...
auto tensor_inn = executor.get_in("input_n");
//some declaratinos
...
//create a executor
//TargetType is NV [NVIDIA GPU]
Net<NV, AK_FLOAT, Precision::FP32> executor(*graph);
//Get the first input tensor.
//The following tensors(tensor_in0, tensor_in2 ...) are resident at GPU.
//Note: Member function get_in returns an pointer to tensor.
Tensor<NV, AK_FLOAT>* tensor_in0 = executor.get_in("input_0");
//If you have multiple input tensors
//You just type this code below.
Tensor<NV, AK_FLOAT>* tensor_in1 = executor.get_in("input_1");
...
auto tensor_inn = executor.get_in("input_n");
```
当得到输入tensor之后,就可以填充它的数据区了。
```c++
//This tensor is resident at GPU.
auto tensor_d_in = executor.get_in("input_0");
//If we want to feed above tensor, we must feed the tensor which is resident at host. And then copy the host tensor to the device's one.
//using Tensor4d = Tensor<Ttype, Dtype>;
Tensor4d<X86, AK_FLOAT> tensor_h_in; //host tensor;
//Tensor<X86, AK_FLOAT> tensor_h_in;
//Allocate memory for host tensor.
tensor_h_in.re_alloc(tensor_d_in->valid_shape());
//Get a writable pointer to tensor.
float *h_data = tensor_h_in.mutable_data();
//Feed your tensor.
/** example
for(int i = 0; i < tensor_h_in.size(); i++){
h_data[i] = 1.0f;
}
*/
//Copy host tensor's data to device tensor.
tensor_d_in->copy_from(tensor_h_in);
// And then
//This tensor is resident at GPU.
auto tensor_d_in = executor.get_in("input_0");
//If we want to feed above tensor, we must feed the tensor which is resident at host. And then copy the host tensor to the device's one.
//using Tensor4d = Tensor<Ttype, Dtype>;
Tensor4d<X86, AK_FLOAT> tensor_h_in; //host tensor;
//Tensor<X86, AK_FLOAT> tensor_h_in;
//Allocate memory for host tensor.
tensor_h_in.re_alloc(tensor_d_in->valid_shape());
//Get a writable pointer to tensor.
float *h_data = tensor_h_in.mutable_data();
//Feed your tensor.
/** example
for(int i = 0; i < tensor_h_in.size(); i++){
h_data[i] = 1.0f;
}
*/
//Copy host tensor's data to device tensor.
tensor_d_in->copy_from(tensor_h_in);
// And then
```
类似的,我们可以利用成员函数get_out来获得输出tensor。但与获得输入tensor不同的是, 我们需要指定输入tensor结点的名字,这个可以从dash board中看到,请从[Anakin Parser](Converter_ch.md)中查看dash board的使用方法。假如有个输出结点叫pred_out, 那么我们可以通过如下代码获得相应的输出tensor:
```c++
//Note: this tensor are resident at GPU.
Tensor<NV, AK_FLOAT>* tensor_out_d = executor.get_out("pred_out");
//Note: this tensor are resident at GPU.
Tensor<NV, AK_FLOAT>* tensor_out_d = executor.get_out("pred_out");
```
#### Executing graph
当一切准备就绪后,我们就可以执行真正的计算了!
```c++
executor.prediction();
executor.prediction();
```
## <span id='example'> 示例代码 </span> ##
下面的例子展示了如何调用Anakin。
......@@ -579,61 +585,61 @@ executor.prediction();
### Single-thread
单线程例子在 *source_root/test/framework/net/net_exec_test.cpp`*
单线程例子在 *`source_root/test/framework/net/net_exec_test.cpp`*
```c++
std::string model_path = "your_Anakin_models/xxxxx.anakin.bin";
// Create an empty graph object.
auto graph = new Graph<NV, AK_FLOAT, Precision::FP32>();
// Load Anakin model.
auto status = graph->load(model_path);
if(!status ) {
LOG(FATAL) << " [ERROR] " << status.info();
}
// Reshape
graph->Reshape("input_0", {10, 384, 960, 10});
// You must optimize graph for the first time.
graph->Optimize();
// Create a executer.
Net<NV, AK_FLOAT, Precision::FP32> net_executer(*graph);
//Get your input tensors through some specific string such as "input_0", "input_1", and
//so on.
//And then, feed the input tensor.
//If you don't know Which input do these specific string ("input_0", "input_1") correspond with, you can launch dash board to find out.
auto d_tensor_in_p = net_executer.get_in("input_0");
Tensor4d<X86, AK_FLOAT> h_tensor_in;
auto valid_shape_in = d_tensor_in_p->valid_shape();
for (int i=0; i<valid_shape_in.size(); i++) {
LOG(INFO) << "detect input dims[" << i << "]" << valid_shape_in[i]; //see tensor's dimentions
}
h_tensor_in.re_alloc(valid_shape_in);
float* h_data = h_tensor_in.mutable_data();
for (int i=0; i<h_tensor_in.size(); i++) {
h_data[i] = 1.0f;
}
d_tensor_in_p->copy_from(h_tensor_in);
//Do inference.
net_executer.prediction();
//Get result tensor through the name of output node.
//And also, you need to see the dash board again to find out how many output nodes are and remember their name.
//For example, you've got a output node named obj_pre_out
//Then, you can get an output tensor.
auto d_tensor_out_0_p = net_executer.get_out("obj_pred_out"); //get_out returns a pointer to output tensor.
auto d_tensor_out_1_p = net_executer.get_out("lc_pred_out"); //get_out returns a pointer to output tensor.
//......
// do something else ...
//...
//save model.
//You might not optimize the graph when you load the saved model again.
std::string save_model_path = model_path + std::string(".saved");
auto status = graph->save(save_model_path);
if (!status ) {
LOG(FATAL) << " [ERROR] " << status.info();
}
std::string model_path = "your_Anakin_models/xxxxx.anakin.bin";
// Create an empty graph object.
auto graph = new Graph<NV, AK_FLOAT, Precision::FP32>();
// Load Anakin model.
auto status = graph->load(model_path);
if(!status ) {
LOG(FATAL) << " [ERROR] " << status.info();
}
// Reshape
graph->Reshape("input_0", {10, 384, 960, 10});
// You must optimize graph for the first time.
graph->Optimize();
// Create a executer.
Net<NV, AK_FLOAT, Precision::FP32> net_executer(*graph);
//Get your input tensors through some specific string such as "input_0", "input_1", and
//so on.
//And then, feed the input tensor.
//If you don't know Which input do these specific string ("input_0", "input_1") correspond with, you can launch dash board to find out.
auto d_tensor_in_p = net_executer.get_in("input_0");
Tensor4d<X86, AK_FLOAT> h_tensor_in;
auto valid_shape_in = d_tensor_in_p->valid_shape();
for (int i=0; i<valid_shape_in.size(); i++) {
LOG(INFO) << "detect input dims[" << i << "]" << valid_shape_in[i]; //see tensor's dimentions
}
h_tensor_in.re_alloc(valid_shape_in);
float* h_data = h_tensor_in.mutable_data();
for (int i=0; i<h_tensor_in.size(); i++) {
h_data[i] = 1.0f;
}
d_tensor_in_p->copy_from(h_tensor_in);
//Do inference.
net_executer.prediction();
//Get result tensor through the name of output node.
//And also, you need to see the dash board again to find out how many output nodes are and remember their name.
//For example, you've got a output node named obj_pre_out
//Then, you can get an output tensor.
auto d_tensor_out_0_p = net_executer.get_out("obj_pred_out"); //get_out returns a pointer to output tensor.
auto d_tensor_out_1_p = net_executer.get_out("lc_pred_out"); //get_out returns a pointer to output tensor.
//......
// do something else ...
//...
//save model.
//You might not optimize the graph when you load the saved model again.
std::string save_model_path = model_path + std::string(".saved");
auto status = graph->save(save_model_path);
if (!status ) {
LOG(FATAL) << " [ERROR] " << status.info();
}
```
# 模型转换指南
Anakin 支持不同框架的模型预测。但由于格式的差别,Anakin 需要您预先转换模型本文档介绍如何转换模型。
Anakin 支持不同框架的模型预测。但由于格式的差别,Anakin 需要您预先转换模型, 本文档介绍如何转换模型。
## 简介
Anakin 模型转换器输入支持 Caffe 和 Fluid 两种格式的预测模型,模型包含网络结构(model 或 prototxt)和权重参数(param 或 caffemodel)。
Anakin 模型转换器输入支持 Caffe 和 Paddle 两种格式的预测模型,模型包含网络结构(model 或 prototxt)和权重参数(param 或 caffemodel)。
模型转换的输出是一个 bin 文件,它作为 Anakin 框架的 graph 参数导入。
模型转换的输出是一个 bin 文件,它作为 Anakin 框架的 graph 参数导入。
您还可以使用模型转换器的 launch board 功能生成网络结构的 HTML 预览。
您还可以使用模型转换器的 launch board 功能生成网络结构的 HTML 预览。
## 系统要求
......@@ -22,7 +22,7 @@ Anakin 模型转换器输入支持 Caffe 和 Fluid 两种格式的预测模型
## 用法
### 1、环境
转换器所需的依赖标注于 *系统要求* 一节。
转换器所需的依赖标注于*系统要求*一节。
### 2、配置
您需要对 *config.yaml* 文件进行修改以告知您的需求。工程中给出了 *config.yaml* 示例,下面作进一步说明。
......@@ -30,7 +30,7 @@ Anakin 模型转换器输入支持 Caffe 和 Fluid 两种格式的预测模型
#### config.yaml
```bash
OPTIONS:
Framework: CAFFE # 依框架类型填写 CAFFE 或 FLUID
Framework: CAFFE # 依框架类型填写 CAFFE 或 Paddle
SavePath: ./output # 转换结束后模型的保存位置
ResultName: googlenet # 输出模型的名字
Config:
......@@ -53,13 +53,13 @@ TARGET:
PrototxtPath: /path/to/your/googlenet.prototxt
ModelPath: /path/to/your/googlenet.caffemodel
FLUID:
# 当 Framework 为 FLUID 时需填写
Paddle:
# 当 Framework 为 Paddle 时需填写
Debug: NULL
ProtoPaths:
- /
PrototxtPath: /path/to/fluid/inference_model
ModelPath: /path/to/fluid/inference_model
PrototxtPath: /path/to/paddle/inference_model
ModelPath: /path/to/paddle/inference_model
# ...
```
......@@ -68,6 +68,6 @@ TARGET:
### 4、预览
最后一步,就是在浏览器中查看令人振奋的转换结果!网址是在 *config.yaml* 中配置的,例如 http://0.0.0.0:8888 。
最后一步,就是在浏览器中查看转换结果!网址是在 *config.yaml* 中配置的,例如 http://0.0.0.0:8888 。
> 注意:若您使用了默认的 IP 地址 0.0.0.0,请在预览时使用真实的服务器地址 real_ip:port 替代它。
......@@ -52,7 +52,7 @@ endif()
#cmakedefine USE_TNEW_PLACE
```
* 其他依赖和编译选项
* 其他依赖和编译选项
修改`cmake`目录下的`compiler_options.cmake``find_modules.cmake`
......@@ -231,7 +231,7 @@ struct TargetWrapper<TNEW, __xxx_target> { //根据TNEW的具体类型修改__xx
4.`impl/`目录下添加设备目录和实现
`saber/core/impl`目录下添加设备目录`tnew`
* 实现`TargetWrapper<TNEW, __xxx_target>`结构体中各函数的定义。
* 实现`TargetWrapper<TNEW, __xxx_target>`结构体中各函数的定义。
如果`TargetWrapper<TNEW, __xxx_target>`的实现与默认的模板类一致,则不用特化出该类。
```c++
......@@ -243,11 +243,11 @@ void TNEW_API::get_device_count(int &count) {
void TNEW_API::set_device(int id){
// add implementation
}
void TNEW_API::mem_alloc(void** ptr, size_t n){
// add implementation
}
void TNEW_API::mem_free(void* ptr){
if(ptr != nullptr){
// add implementation
......@@ -275,7 +275,7 @@ void Device<TNEW>::get_info() {
### 在`saber/funcs`中实现设备相关的op
参考[如何增加新的Operator](addCustomOp.md)
参考[如何增加新的Operator](./how_to_add_anakin_op.md)
## <span id = '0003'> 在`framework`中添加设备的具体化或实例化 </span> ##
......@@ -329,7 +329,7 @@ public:
typedef Tensor4d<TNEW, DataTypeRecover<Dtype>::type> type;
PBlock() {
_inner_tensor = std::make_shared<type>();
_inner_tensor = std::make_shared<type>();
}
...
}
......@@ -348,7 +348,7 @@ struct target_host<saber::TNEW> {
### `framework/graph`
* `graph.cpp`中添加实例化
```c++
#ifdef USE_TNEW_PLACE
template class Graph<TNEW, AK_FLOAT, Precision::FP32>;
......@@ -360,7 +360,7 @@ struct target_host<saber::TNEW> {
### `framework/model_parser`
* `parser.cpp`中添加实例化
```c++
#ifdef USE_TNEW_PLACE
template
......@@ -372,7 +372,7 @@ struct target_host<saber::TNEW> {
template
Status load<TNEW, AK_FLOAT, Precision::INT8>(graph::Graph<TNEW, AK_FLOAT, Precision::INT8>* graph,
const char* model_path);
template
Status save<TNEW, AK_FLOAT, Precision::FP32>(graph::Graph<TNEW, AK_FLOAT, Precision::FP32>* graph,
std::string& model_path);
......@@ -382,7 +382,7 @@ struct target_host<saber::TNEW> {
template
Status save<TNEW, AK_FLOAT, Precision::INT8>(graph::Graph<TNEW, AK_FLOAT, Precision::INT8>* graph,
std::string& model_path);
template
Status load<TNEW, AK_FLOAT, Precision::FP32>(graph::Graph<TNEW, AK_FLOAT, Precision::FP32>* graph,
std::string& model_path);
......@@ -392,7 +392,7 @@ struct target_host<saber::TNEW> {
template
Status load<TNEW, AK_FLOAT, Precision::INT8>(graph::Graph<TNEW, AK_FLOAT, Precision::INT8>* graph,
std::string& model_path);
template
Status save<TNEW, AK_FLOAT, Precision::FP32>(graph::Graph<TNEW, AK_FLOAT, Precision::FP32>* graph,
const char* model_path);
......
......@@ -10,12 +10,13 @@ Anakin 预测引擎
install_anakin.md
convert_paddle_to_anakin.md
run_anakin_on_arm.md
anakin_tutorial.md
anakin_run_on_arm.md
anakin_example.md
anakin_gpu_benchmark.md
anakin_arm_benchmark.md
开发文档
~~~~~~~
......@@ -24,3 +25,4 @@ Anakin 预测引擎
how_to_add_anakin_op.md
how_to_support_new_device_in_anakin.md
anakin_parser_design.md
## 源码编译安装Anakin ##
## 源码编译安装Anakin ##
我们已经在CentOS 7.3上成功的安装和测试了Anakin,对于其他操作系统,我们将很快支持。
......@@ -6,7 +6,7 @@
* [在CentOS上安装 Anakin]()
* [在Ubuntu上安装 Anakin]()
* [在ARM上安装 Anakin](run_on_arm_ch.md)
* [在ARM上安装 Anakin](./anakin_run_on_arm.md)
* [验证安装]()
......@@ -26,30 +26,37 @@
#### 3. 编译支持NVIDIA GPU的Anakin ####
- 3.1. 安装依赖
- 3.1.1 protobuf
>$ git clone https://github.com/google/protobuf
>$ cd protobuf
>$ git submodule update --init --recursive
>$ ./autogen.sh
>$ ./configure --prefix=/path/to/your/insall_dir
>$ make
>$ make check
>$ make install
>$ sudo ldconfig
- 3.1.1 protobuf
如安装protobuf遇到任何问题,请访问[这里](https://github.com/google/protobuf/blob/master/src/README.md)
```
> git clone https://github.com/google/protobuf
> cd protobuf
> git submodule update --init --recursive
> ./autogen.sh
> ./configure --prefix=/path/to/your/insall_dir
> make
> make check
> make install
> sudo ldconfig
```
如安装protobuf遇到任何问题,请访问[这里](https://github.com/google/protobuf/blob/master/src/README.md)
- 3.2 CUDA Toolkit
- [CUDA 8.0](https://developer.nvidia.com/cuda-zone) or higher. 具体信息参见[NVIDIA's documentation](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/).
- [cuDNN v7](https://developer.nvidia.com/cudnn). 具体信息参见[NVIDIA's documentation](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/).
- [CUDA 8.0](https://developer.nvidia.com/cuda-zone) or higher, 具体信息参见[NVIDIA's documentation](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/).
- [cuDNN v7](https://developer.nvidia.com/cudnn), 具体信息参见[NVIDIA's documentation](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/).
- 3.3 编译Anakin
>$ git clone https:/xxxxx
>$ cd anakin
>$ mkdir build
>$ camke ..
>$ make
```
> git clone https:/xxxxx
> cd anakin
> mkdir build
> camke ..
> make
```
#### 4. 编译支持AMD GPU的Anakin ####
......@@ -63,7 +70,8 @@
### 在ARM上安装 Anakin ###
暂时还不支持
请参考[ARM安装文档](./anakin_run_on_arm.md)
### 验证安装 ###
we are coming soon...
安装完成后,如果没有报错信息,你可以通过运行 `output/unit_test`路径下的单测示例验证是否编译成功。
## 源码编译 Anakin ##
## ARM 源码编译 Anakin ##
目前Anakin支持ARM Android平台,采用Android NDK交叉编译工具链,已在mac os和centos上编译和测试通过。
......@@ -12,37 +12,44 @@
### <span id = '0001'> 1. 系统需求 </span> ###
* 宿主机: linux, mac
* cmake 3.8.2+
* 宿主机: linux, mac
* cmake 3.8.2+
* Android NDK r14, Linux 版本[从这里下载](https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip)
### <span id = '0002'> 2. 安装第三方依赖 </span> ###
- 2.1 protobuf3.4.0
源码从这里[下载](https://github.com/google/protobuf/releases/tag/v3.4.0)
- 2.1.1 为宿主机编译protobuf
```bash
$ tar -xzf protobuf-3.4.0.tar.gz
$ cd protobuf-3.4.0
$ ./autogen.sh
$ ./configure
$ make
$ make check
- 2.1 protobuf3.4.0
源码从这里[下载](https://github.com/google/protobuf/releases/tag/v3.4.0)
- 2.1.1 为宿主机编译protobuf
```bash
$ tar -xzf protobuf-3.4.0.tar.gz
$ cd protobuf-3.4.0
$ ./autogen.sh
$ ./configure
$ make
$ make check
$ make install
```
上述 $make install 执行后,可在 /usr/local/include/google 找到 libprotobuf 所需的头文件,将整个google文件夹拷贝至Anakin/third-party/arm-android/protobuf/下,
如有问题,请点[这里](https://github.com/google/protobuf/blob/v3.4.0/src/README.md)
然后将已经生成文件清除。
```bash
```
上述 $make install 执行后,可在 /usr/local/include/google 找到 libprotobuf 所需的头文件,将整个google文件夹拷贝至Anakin/third-party/arm-android/protobuf/下
如有问题,请点[这里](https://github.com/google/protobuf/blob/v3.4.0/src/README.md),然后将已经生成文件清除。
```bash
$ make distclean
```
- 2.1.1 交叉编译Android`armeabi-v7a`的protobuf,注意设置ANDROID_NDK的路径,以及ARCH_ABI、HOSTOSN的值,
```
- 2.1.1 交叉编译Android`armeabi-v7a`的protobuf,注意设置ANDROID_NDK的路径,以及ARCH_ABI、HOSTOSN的值
```bash
$ export ANDROID_NDK=your_ndk_path
$ export ANDROID_NDK=your_ndk_path
$ ARCH_ABI="arm-linux-androideabi-4.9"
$ HOSTOSN="darwin-x86_64"
$ export SYSROOT=$ANDROID_NDK/platforms/android-9/arch-arm
$ export SYSROOT=$ANDROID_NDK/platforms/android-9/arch-arm
$ export PREBUILT=$ANDROID_NDK/toolchains/$ARCH_ABI
$ export LDFLAGS="--sysroot=$SYSROOT"
$ export LD="$ANDROID_NDK/toolchains/$ARCH_ABI/prebuilt/$HOSTOSN/arm-linux-androideabi/bin/ld $LDFLAGS"
......@@ -53,34 +60,38 @@
$ export CCFLAGS="$CXXFLAGS"
$ export CXX="$PREBUILT/prebuilt/$HOSTOSN/bin/arm-linux-androideabi-g++ $CXXFLAGS"
$ export CC="$CXX"
$ export RANLIB="$ANDROID_NDK/toolchains/$ARCH_ABI/prebuilt/$HOSTOSN/bin/arm-linux-androideabi-ranlib"
$ ./autogen.sh
$ ./configure --host=arm-linux-androideabi --with-sysroot=$SYSROOT --enable-cross-compile --with-protoc=protoc --disable-shared CXX="$CXX" CC="$CC" LD="$LD"
$ export RANLIB="$ANDROID_NDK/toolchains/$ARCH_ABI/prebuilt/$HOSTOSN/bin/arm-linux-androideabi-ranlib"
$ ./autogen.sh
$ ./configure --host=arm-linux-androideabi --with-sysroot=$SYSROOT --enable-cross-compile --with-protoc=protoc --disable-shared CXX="$CXX" CC="$CC" LD="$LD"
$ make
```
编译生成 *.a 静态库,若希望编译*.so 动态链接库 ,请在./configure参数中改--disable-shared为--disable-static --enable-shared。
生成文件在src/.libs/下,将生成的文件拷贝至Anakin/third-party/arm-android/protobuf/lib下。
在[cmake](../../cmake/find_modules.cmake)中更新`ARM_RPOTO_ROOT`的路径。
```cmake
```
编译生成 *.a 静态库,若希望编译*.so 动态链接库 ,请在./configure参数中改--disable-shared为--disable-static --enable-shared。
生成文件在src/.libs/下,将生成的文件拷贝至Anakin/third-party/arm-android/protobuf/lib下。
[cmake](../../cmake/find_modules.cmake)中更新`ARM_RPOTO_ROOT`的路径。
```cmake
set(ARM_RPOTO_ROOT "${CMAKE_SOURCE_DIR}/third-party/arm-android/protobuf")
```
- 2.2 opencv 2.4.3+(optional)
Anakin只在examples示例中使用opencv
Android系统的opencv从[这里下载](https://opencv.org/releases.html)
解压后将 `3rdparty/libs/armeabi-v7a`中的库文件拷贝到`libs/armeabi-v7a`
在[cmake](../../cmake/find_modules.cmake)中搜索`anakin_find_opencv`,
并设置 `include_directories` 和 `LINK_DIRECTORIES`为自己安装的库的路径。
```cmake
```
- 2.2 opencv 2.4.3+(optional)
Anakin只在examples示例中使用opencv
Android系统的opencv从[这里下载](https://opencv.org/releases.html)
解压后将 `3rdparty/libs/armeabi-v7a`中的库文件拷贝到`libs/armeabi-v7a`
[cmake](../../cmake/find_modules.cmake)中搜索`anakin_find_opencv`,
并设置 `include_directories``LINK_DIRECTORIES`为自己安装的库的路径。
```cmake
include_directories(${CMAKE_SOURCE_DIR}/third-party/arm-android/opencv/sdk/native/jni/include/)
LINK_DIRECTORIES(${CMAKE_SOURCE_DIR}/third-party/arm-android/opencv/sdk/native/libs/armeabi-v7a/)
```
```
### <span id = '0003'> 3. Anakin源码编译 </span> ###
#### 编译Android版本
克隆[源码](https://github.com/PaddlePaddle/Anakin/tree/arm)
克隆[源码](https://github.com/PaddlePaddle/Anakin/tree/arm)
```bash
cd your_dir
git clone https://github.com/PaddlePaddle/Anakin.git
......@@ -88,64 +99,87 @@
git fetch origin arm
git checkout arm
```
修改`android_build.sh`
- 修改NDK路径
修改`android_build.sh`
- 修改NDK路径
```bash
#modify "your_ndk_path" to your NDK path
export ANDROID_NDK=your_ndk_path
```
- 修改ARM 处理器架构
对于32位ARM处理器, 将ANDROID_ABI 设置为 `armeabi-v7a with NEON`
对于64位ARM处理器, 可以将ANDROID_ABI 设置为 `armeabi-v7a with NEON`或者`arm64-v8a`
目前我们只支持 `armeabi-v7a with NEON``arm64-v8a` 还在开发中。
- 修改ARM 处理器架构
对于32位ARM处理器, 将ANDROID_ABI 设置为 `armeabi-v7a with NEON`
对于64位ARM处理器, 可以将ANDROID_ABI 设置为 `armeabi-v7a with NEON`或者`arm64-v8a`
目前我们只支持 `armeabi-v7a with NEON``arm64-v8a` 还在开发中。
```bash
-DANDROID_ABI="armeabi-v7a with NEON"
```
- 设置Android API
根据Android系统的版本设置API level, 例如API Level 21 -> Android 5.0.1
- 设置Android API
根据Android系统的版本设置API level, 例如API Level 21 -> Android 5.0.1
```bash
-DANDROID_NATIVE_API_LEVEL=21
```
- 选择编译静态库或动态库
设置`BUILD_SHARED=NO`编译静态库
设置`BUILD_SHARED=YES`编译动态库
- 选择编译静态库或动态库
设置`BUILD_SHARED=NO`编译静态库
设置`BUILD_SHARED=YES`编译动态库
```bash
-DBUILD_SHARED=NO
```
- OpenMP多线程支持
设置`USE_OPENMP=YES`开启OpenMP多线程
- OpenMP多线程支持
设置`USE_OPENMP=YES`开启OpenMP多线程
```bash
-DUSE_OPENMP=YES
```
- 编译单测文件
设置`BUILD_WITH_UNIT_TEST=YES`将会编译单测文件
```bash
-DBUILD_WITH_UNIT_TEST=YES
```
- 编译示例文件
设置`BUILD_EXAMPLES=YES`将会编译示例文件
```bash
-DBUILD_EXAMPLES=YES
```
- 开启opencv
如果使用opencv,设置`USE_OPENCV=YES`
```bash
-DUSE_OPENCV=YES
```
- 开始编译
运行脚本 `android_build.sh` 将自动编译Anakin
- 编译单测文件
设置`BUILD_WITH_UNIT_TEST=YES`将会编译单测文件
```bash
-DBUILD_WITH_UNIT_TEST=YES
```
- 编译示例文件
设置`BUILD_EXAMPLES=YES`将会编译示例文件
```bash
-DBUILD_EXAMPLES=YES
```
- 开启opencv
如果使用opencv,设置`USE_OPENCV=YES`
```bash
-DUSE_OPENCV=YES
```
- 开始编译
运行脚本 `android_build.sh` 将自动编译Anakin
```bash
./android_build.sh
```
### <span id = '0004'> 4. 验证安装 </span> ###
编译好的库会放在目录`${Anakin_root}/output`下;
编译好的单测文件会放在`${Anakin_root}/output/unit_test`目录下;
编译好的示例文件会放在`${Anakin_root}/output/examples`目录下。
对于Android系统,打开设备的调试模式,通过ADB可以访问的目录是`data/local/tmp`,通过ADB push将测试文件、模型和数据发送到设备目录, 运行测试文件。
### <span id = '0004'> 4. 验证安装 </span> ###
编译好的库会放在目录`${Anakin_root}/output`
编译好的单测文件会放在`${Anakin_root}/output/unit_test`目录下
编译好的示例文件会放在`${Anakin_root}/output/examples`目录下
对于Android系统,打开设备的调试模式,通过ADB可以访问的目录是`data/local/tmp`,通过ADB push将测试文件、模型和数据发送到设备目录,运行测试文件。
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