diff --git a/doc/howto/index_cn.rst b/doc/howto/index_cn.rst index ccd909770253bb85dbc8a5a2560594076c2f68b0..e0c69f7a6a4043abe999af6c8dd2555178b68424 100644 --- a/doc/howto/index_cn.rst +++ b/doc/howto/index_cn.rst @@ -9,6 +9,7 @@ usage/cmd_parameter/index_cn.rst usage/cluster/cluster_train_cn.md + usage/capi/index_cn.rst 开发标准 -------- diff --git a/doc/howto/usage/capi/compile_paddle_lib_cn.md b/doc/howto/usage/capi/compile_paddle_lib_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..ac5ecffe2ea8ddc3703a32e9a0a8ee83bbe5dd14 --- /dev/null +++ b/doc/howto/usage/capi/compile_paddle_lib_cn.md @@ -0,0 +1,122 @@ +## 编译 PaddlePaddle 预测库 + +### 概述 + +使用 C-API 进行预测依赖于将 PaddlePaddle 核心代码编译成链接库,只需在编译时需配制下面这些编译选项: + +必须配置选项: +- `WITH_C_API`,必须配置为`ON`。 + +推荐配置选项: +- `WITH_PYTHON`,推荐配置为`OFF` +- `WITH_SWIG_PY`,推荐配置为`OFF` +- `WITH_GOLANG`,推荐设置为`OFF` + +可选配置选项: +- `WITH_GPU`,可配置为`ON/OFF` +- `WITH_MKL`,可配置为`ON/OFF` + +对推荐配置中的选项建议按照设置,以避免链接不必要的库。其它可选编译选项按需进行设定。 + +下面的代码片段从github拉取最新代码,配制编译选项(需要将PADDLE_ROOT替换为PaddlePaddle预测库的安装路径): + +```shell +PADDLE_ROOT=/path/of/capi +git clone https://github.com/PaddlePaddle/Paddle.git +cd Paddle +mkdir build +cd build +cmake -DCMAKE_INSTALL_PREFIX=$PADDLE_ROOT \ + -DCMAKE_BUILD_TYPE=Release \ + -DWITH_C_API=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_GOLANG=OFF \ + -DWITH_PYTHON=OFF \ + -DWITH_MKL=OFF \ + -DWITH_GPU=OFF \ + .. +``` + +执行上述代码生成Makefile文件后,执行:`make && make install`。成功编译后,使用C-API所需的依赖(包括:(1)编译出的PaddlePaddle预测库和头文件;(2)第三方链接库和头文件)均会存放于`PADDLE_ROOT`目录中。 + +编译成功后在 `PADDLE_ROOT` 下会看到如下目录结构(包括了编译出的PaddlePaddle头文件和链接库,以及第三方依赖链接库和头文件(如果需要,由链接方式决定)): + +```text +├── include +│   └── paddle +│   ├── arguments.h +│   ├── capi.h +│   ├── capi_private.h +│   ├── config.h +│   ├── error.h +│   ├── gradient_machine.h +│   ├── main.h +│   ├── matrix.h +│   ├── paddle_capi.map +│   └── vector.h +├── lib +│   ├── libpaddle_capi_engine.a +│   ├── libpaddle_capi_layers.a +│   ├── libpaddle_capi_shared.so +│   └── libpaddle_capi_whole.a +└── third_party + ├── gflags + │   ├── include + │   │   └── gflags + │   │   ├── gflags_completions.h + │   │   ├── gflags_declare.h + │   │   ... + │   └── lib + │   └── libgflags.a + ├── glog + │   ├── include + │   │   └── glog + │   │   ├── config.h + │   │   ... + │   └── lib + │   └── libglog.a + ├── openblas + │   ├── include + │   │   ├── cblas.h + │   │   ... + │   └── lib + │   ... + ├── protobuf + │   ├── include + │   │   └── google + │   │   └── protobuf + │   │   ... + │   └── lib + │   └── libprotobuf-lite.a + └── zlib + ├── include + │   ... + └── lib + ... + +``` + +### 链接说明 + +目前提供三种链接方式: + +1. 链接`libpaddle_capi_shared.so` 动态库 + - 使用 PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_shared.so`时,需注意: + 1. 如果编译时指定编译CPU版本,且使用`OpenBLAS`数学库,在使用C-API开发预测程序时,只需要链接`libpaddle_capi_shared.so`这一个库。 + 1. 如果是用编译时指定CPU版本,且使用`MKL`数学库,由于`MKL`库有自己独立的动态库文件,在使用PaddlePaddle C-API开发预测程序时,需要自己链接MKL链接库。 + 1. 如果编译时指定编译GPU版本,CUDA相关库会在预测程序运行时动态装载,需要将CUDA相关的库设置到`LD_LIBRARY_PATH`环境变量中。 + - 这种方式最为简便,链接相对容易,**在无特殊需求情况下,推荐使用此方式**。 + +2. 链接静态库 `libpaddle_capi_whole.a` + - 使用PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_whole.a`时,需注意: + 1. 需要指定`-Wl,--whole-archive`链接选项。 + 1. 需要显式地链接 `gflags`、`glog`、`libz`、`protobuf` 等第三方库,可在`PADDLE_ROOT/third_party`下找到。 + 1. 如果在编译 C-API 时使用OpenBLAS数学库,需要显示地链接`libopenblas.a`。 + 1. 如果在编译 C-API 是使用MKL数学库,需要显示地链接MKL的动态库。 + +3. 链接静态库 `libpaddle_capi_layers.a`和`libpaddle_capi_engine.a` + - 使用PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_whole.a`时,需注意: + 1. 这种链接方式主要用于移动端预测。 + 1. 为了减少生成链接库的大小把`libpaddle_capi_whole.a`拆成以上两个静态链接库。 + 1. 需指定`-Wl,--whole-archive -lpaddle_capi_layers` 和 `-Wl,--no-whole-archive -lpaddle_capi_engine` 进行链接。 + 1. 第三方依赖库需要按照与方式2同样方法显示地进行链接。 diff --git a/doc/howto/usage/capi/images/csr.png b/doc/howto/usage/capi/images/csr.png new file mode 100644 index 0000000000000000000000000000000000000000..3dc10b8de4f6d3f517624956b1694b689405a031 Binary files /dev/null and b/doc/howto/usage/capi/images/csr.png differ diff --git a/doc/howto/usage/capi/images/sequence_data.png b/doc/howto/usage/capi/images/sequence_data.png new file mode 100644 index 0000000000000000000000000000000000000000..6e47a46b8955dfe977e85898fe3c9f33ed28de7e Binary files /dev/null and b/doc/howto/usage/capi/images/sequence_data.png differ diff --git a/doc/howto/usage/capi/images/workflow_of_CAPI.png b/doc/howto/usage/capi/images/workflow_of_CAPI.png new file mode 100644 index 0000000000000000000000000000000000000000..a4399ade048b3fe10d2d9c714bc34333ca068edb Binary files /dev/null and b/doc/howto/usage/capi/images/workflow_of_CAPI.png differ diff --git a/doc/howto/usage/capi/index_cn.rst b/doc/howto/usage/capi/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..fd774fbc742671c5a8009cb742f2c9d06a525199 --- /dev/null +++ b/doc/howto/usage/capi/index_cn.rst @@ -0,0 +1,9 @@ +PaddlePaddle C-API +================== + +.. toctree:: + :maxdepth: 1 + + compile_paddle_lib_cn.md + organization_of_the_inputs_cn.md + workflow_of_capi_cn.md diff --git a/doc/howto/usage/capi/organization_of_the_inputs_cn.md b/doc/howto/usage/capi/organization_of_the_inputs_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..563ec5ca21ec5d75800fa201943d65e6d6fe51ea --- /dev/null +++ b/doc/howto/usage/capi/organization_of_the_inputs_cn.md @@ -0,0 +1,285 @@ +## 输入/输出数据组织 + +这篇文档介绍在使用 PaddlePaddle C-API 时如何组织输入数据,以及如何解析神经网络前向计算的输出结果。 + +### 输入/输出数据类型 +在C-API中,按照基本数据类型在PaddlePaddle内部的定义和实现,输入数据可分为: +1. 一维整型数组 +1. 二维浮点型矩阵 + - 稠密矩阵 + - 稀疏矩阵 + +说明: +1. 一维数组**仅支持整型值**; + - 常用于自然语言处理任务,例如:表示词语在词典中的序号; + - 分类任务中类别标签; +1. 逻辑上高于二维的数据(例如含有多个通道的图片,视频等)在程序实现中都会转化为二维矩阵,转化方法在相应的领域都有通用解决方案,需要使用者自己了解并完成转化; +1. 二维矩阵可以表示行向量和列向量,任何时候如果需要浮点型数组(向量),都应使用C-API中的矩阵来表示,而不是C-API中的一维数组。 +1. 不论是一维整型数组还是二维浮点数矩阵,**为它们附加上序列信息将变成序列输入。PaddlePaddle 会通过判数据是否附带有序列信息来判断一个向量/矩阵是否是一个序列**。当非序列输入时,无需关心和处理序列信息。关于什么是“序列信息”,下文会详细进行介绍。 + +### 基本使用概念 + +- 在PaddlePaddle内部,神经网络中一个计算层的输入/输出被组织为一个 `Argument` 结构体,如果神经网络有多个输入或者多个输入,每一个输入/输入都会对应有自己的`Argument`。 +- `Argument` 并不真正“存储”数据,而是将输入/输出信息有机地组织在一起。 +- 在`Argument`内部由`IVector`(对应着上文提到的一维整型数组)和`Matrix`(对应着上文提到的二维浮点型矩阵)来实际存储数据;由 `Sequence Start Positions` (下文详细解释) 来描述输入/输出的序列信息。 + +- **注**: + 1. 这篇文档之后部分将会统一使用`argument`来特指PaddlePaddle中神经网络计算层一个输入/输出数据。 + 1. 使用`paddle_ivector`来特指PaddlePaddle中的一维整型数组。 + 1. 使用`paddle_matrix`来特指PaddlePaddle中的二维浮点型矩阵。 + +### 组织输入数据 +- 一维整型数组 + + 概念上可以将`paddle_ivector`理解为一个一维的整型数组,通常用于表示离散的类别标签,或是在自然语言处理任务中表示词语在字典中的序号。下面的代码片段创建了含有三个元素`1`、`2`、`3`的`paddle_ivector`。 + ```c + int ids[] = {1, 2, 3}; + paddle_ivector ids_array = + paddle_ivector_create(ids, sizeof(ids) / sizeof(int), false, false); + CHECK(paddle_arguments_set_ids(in_args, 0, ids_array)); + ``` + +- **稠密矩阵** + - 一个`m×n`的稠密矩阵是一个由`m`行`n`列元素排列成的矩形阵列,矩阵里的元素是浮点数。对神经网络来说,矩阵的高度`m`是一次预测接受的样本数目,宽度$n$是神经网络定义时,`paddle.layer.data`的`size`。 + - 下面的代码片段创建了一个高度为1,宽度为`layer_size`的稠密矩阵,矩阵中每个元素的值随机生成。 + + ```c + paddle_matrix mat = paddle_matrix_create( + /* height = batch size */ 1, + /* width = dimensionality of the data layer */ layer_size, + /* whether to use GPU */ false); + + paddle_real* array; + // Get the pointer pointing to the start address of the first row of the + // created matrix. + CHECK(paddle_matrix_get_row(mat, 0, &array)); + + // Fill the matrix with a randomly generated test sample. + srand(time(0)); + for (int i = 0; i < layer_size; ++i) { + array[i] = rand() / ((float)RAND_MAX); + } + + // Assign the matrix to the argument. + CHECK(paddle_arguments_set_value(in_args, 0, mat)); + ``` + +- **稀疏矩阵** + + PaddlePaddle C-API 中 稀疏矩阵使用[CSR(Compressed Sparse Row Format)](https://en.wikipedia.org/wiki/Sparse_matrix#Compressed_sparse_row_(CSR,_CRS_or_Yale_format))格式存储。下图是CSR存储稀疏矩阵的示意图。 +

+
图1. 稀疏矩阵存储示意图 +

+ + CSR存储格式通过:(1)非零元素的值(上图中的`values`);(2)行偏移(上图中的`row offsets`):每一行元素在`values`中的起始偏移,`row offsets`中元素个数总是等于行数 + 1;(3)非零元素的列号(上图中的`column indices`)来确定稀疏矩阵的内容。 + + 在PaddlePaddle C-API中,通过调用以下接口创建稀疏矩阵: + + ```c + PD_API paddle_matrix paddle_matrix_create_sparse( + uint64_t height, uint64_t width, uint64_t nnz, bool isBinary, bool useGpu); + ``` + + 1. 创建稀疏矩阵时需要显示地指定矩阵的(1)高度(`height`,在神经网络中等于一次预测处理的样本数)(2)宽度(`width`,`paddle.layer.data`的`size`)以及(3)非零元个数(`nnz`)。 + 1. 当上述接口第4个参数`isBinary`指定为`true`时,**只需要设置行偏移(`row_offset`)和列号(`colum indices`),不需要提供元素值(`values`)**,这时行偏移和列号指定的元素默认其值为1。 + + 下面的代码片段创建了一个CPU上的二值稀疏矩阵: + + ```c + paddle_matrix mat = paddle_matrix_create_sparse(1, layer_size, nnz, true, false); + int colIndices[] = {9, 93, 109}; // layer_size here is greater than 109. + int rowOffset[] = {0, sizeof(colIndices) / sizeof(int)}; + + CHECK(paddle_matrix_sparse_copy_from(mat, + rowOffset, + sizeof(rowOffset) / sizeof(int), + colIndices, + (colIndices) / sizeof(int), + NULL /*values array is NULL.*/, + 0 /*size of the value arrary is 0.*/)); + CHECK(paddle_arguments_set_value(in_args, 0, mat)); + ``` + 下面的代码片段在创建了一个CPU上的带元素值的稀疏矩阵: + ```c + paddle_matrix mat = paddle_matrix_create_sparse(1, layer_size, nnz, false, false); + int colIndices[] = {9, 93, 109}; // layer_size here is greater than 109. + int rowOffset[] = {0, sizeof(colIndices) / sizeof(int)}; + float values[] = {0.5, 0.5, 0.5}; + + CHECK(paddle_matrix_sparse_copy_from(mat, + rowOffset, + sizeof(rowOffset) / sizeof(int), + colIndices, + sizeof(colIndices) / sizeof(int), + values, + sizeof(values) / sizeof(float))); + ``` + 注意事项: + 1. 移动端预测**不支持**稀疏矩阵及相关的接口。 + +### 组织序列信息 + +多个排成一列的元素(可以是整型、浮点数、浮点数向量等)构成一个序列,元素之间的顺序是序列所携带的重要信息。不同序列可能会含有不同数目个元素。在 PaddlePaddle 中,序列输入/输出数据是在上文介绍的**数据输入(一维整型数组,二维浮点数矩阵)基础上,附加上序列信息**。下面详细解释什么是“序列信息”。 + +我们将神经网络一次计算接受的所有输入样本称之为一个`batch`(可以含有一条或多条样本),每一个序列在整个`batch`中的偏移,就是PaddlePaddle中所指的**序列信息**,称之为“sequence start positions”。PaddlePaddle 支持两种序列类型: + +1. 单层序列 + - 序列中的每一个元素是非序列,是进行计算的基本单位,不可再进行拆分。 + - 例如:自然语言中的句子是一个序列,序列中的元素是词语; +1. 双层序列 + - 序列中的每一个元素又是一个序列。 + - 例如:自然语言中的段落是一个双层序列;段落是由句子构成的序列;句子是由词语构成的序列。 + - 双层序列在处理长序列的任务或是构建层级模型时会发挥作用。 + +这篇文档之后部分会统一使用`sequence_start_positions`来特指:PaddlePaddle中神经网络计算层输入/输出所携带的序列信息。 + +对双层序列来讲,不仅要提供每一个外层序列在整个`batch`中的偏移,每一个外层序列又含有若干个内层序列,需要同时提供每一个内层序列在整个`batch`中的偏移。也就是说:**双层序列需要设置分别为外层序列和内层序列分别设置`sequence_start_positions`信息**。 + +**注:** +1. 不论序列中的元素在内存中占用多少实际存储空间,`sequence_start_positions`表示的偏移是以“序列中的一个元素”作为统计的基本单位,而不是相对`batch`起始存储地址以数据的存储大小为单位的偏移。 +1. 非序列输入不携带`sequence_start_positions`,非序列输入无需构造`sequence_start_positions`。 +1. **不论是单层序列还是双层序列的序列信息,都使用`paddle_ivector`(也就是PaddlePaddle中的一维整型数组)来存储。** + +图2 是PaddlePaddle中单层序列和双层序列存储示意图。 +

+
图2. 序列输入示意图 +

+ +- 单层序列 + + 图2 (a) 展示了一个含有4个序列的`batch`输入: + 1. 4个序列的长度分别为:5、3、2、4; + 1. 这时的`sequence_start_positions`为:`[0, 5, 8, 10, 14]`; + 1. 本地训练. 不论数据域是`paddle_ivector`类型还是`paddle_matrix`类型,都可以通过调用下面的接口为原有的数据输入附加上序列信息,使之变为一个单层序列输入,代码片段如下: + + ```c + int seq_pos_array[] = {0, 5, 8, 10, 14}; + paddle_ivector seq_pos = paddle_ivector_create( + seq_pos_array, sizeof(seq_pos_array) / sizeof(int), false, false); + // Suppose the network only has one input data layer. + CHECK(paddle_arguments_set_sequence_start_pos(in_args, 0, 0, seq_pos)); + ``` + +- 双层序列 + + 图2 (b) 展示了一个含有4个序列的`batch`输入; + 1. 4个序列的长度分别为:5、3、2、4;这四个序列又分别含有3、2、1、2个子序列; + 1. 这时的需要同时提供: + - 外层序列在`batch`中的起始偏移`:[0, 5, 8, 10, 14]`; + - 内层序列在`batch`中的起始偏移:`[0, 2, 3, 5, 7, 8, 10, 13, 14]`; + 1. 不论数据域是`paddle_ivector`类型还是`paddle_matrix`类型,这时需要调用创建序列信息和为`argument`设置序列信息的接口**两次**,分别为数据输入添加外层序列和内层序列的序列信息,使之变为一个双层序列输入,代码片段如下: + ```c + // set the sequence start positions for the outter sequences. + int outter_seq_pos_array[] = {0, 5, 8, 10, 14}; + paddle_ivector seq_pos = + paddle_ivector_create(outter_seq_pos_array, + sizeof(outter_pos_array) / sizeof(int), + false, + false); + // The third parameter of this API indicates the sequence level. + // 0 for the outter sequence. 1 for the inner sequence. + // If the input is a sequence not the nested sequence, the third parameter is + // fixed to be 0. + CHECK(paddle_arguments_set_sequence_start_pos(in_args, 0, 0, seq_pos)); + + // set the sequence start positions for the outter sequences. + int inner_seq_pos_array[] = {0, 2, 3, 5, 7, 8, 10, 13, 14}; + paddle_ivector seq_pos = paddle_ivector_create( + inner_pos_array, sizeof(inner_pos_array) / sizeof(int), false, false); + // The third parameter of this API indicates the sequence level. + // 0 for the outter sequence. 1 for the inner sequence. + CHECK(paddle_arguments_set_sequence_start_pos(in_args, 0, 1, seq_pos)); + ``` + +注意事项: +1. 当一个`batch`中含有多个序列,**不支持序列长度为`0`的序列(也就是空输入)** 作为输入。不同计算层对空输入的处理策略有可能不同,潜在会引起未定义行为,或者引起行时错误,请在输入时进行合法性检查。 + +### Python 端数据类型说明 + +下表列出了Python端训练接口暴露的数据类型(`paddle.layer.data`函数`type`字段的取值)对应于调用C-API需要创建的数据类型: + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Python 端数据类型C-API 输入数据类型
paddle.data_type.integer_value整型数组,无需附加序列信息
paddle.data_type.dense_vector浮点型稠密矩阵,无需附加序列信息
paddle.data_type.sparse_binary_vector浮点型稀疏矩阵,无需提供非零元的值,默认为1,无需附加序列信息
paddle.data_type.sparse_vector浮点型稀疏矩阵,需提供非零元的值,无需附加序列信息
paddle.data_type.integer_value_sequence整型数组,需附加序列信息
paddle.data_type.dense_vector_sequence浮点型稠密矩阵,需附加序列信息
paddle.data_type.sparse_binary_vector_sequence浮点型稀疏矩阵,无需提供非零元的值,默认为1,需附加序列信息
paddle.data_type.sparse_vector_sequence浮点型稀疏矩阵,需提供非零元的值,需附加序列信息
paddle.data_type.integer_value_sub_sequence整型数组,需附加双层序列信息
paddle.data_type.dense_vector_sub_sequence浮点型稠密矩阵,需附加双层序列信息
paddle.data_type.sparse_binary_vector_sub_sequence浮点型稀疏矩阵,无需提供非零元的值,默认为1,需附加双层序列信息
paddle.data_type.sparse_vector_sub_sequence浮点型稀疏矩阵,需提供非零元的值,需附加双层序列信息
+ +
+ + +### 输出数据 + +PaddlePaddle中一个计算层的输出数据组织方式和输入数据组织方式完全相同。一个输出数据同样被组织为一个`argument`,`argument`通过`paddle_matrix`或`paddle_ivector`存数数据,如果输出是一个序列,那么会携带有`sequence_start_positions`信息。调用C-API相关接口,读取需要的结果即可。 + +### 总结 + +- 在PaddlePaddle内部,神经网络中一个计算层的输入/输出被组织为`argument`。 +- `argument`并不真正“存储”数据,而是将输入/输出信息有机地组织在一起。 +- 在`argument`内部由`paddle_ivector`(一维整型数组)和`paddle_matrix`(二维浮点型矩阵)来实际存储数据。 +如果是一个序列输入/输出由 `sequence start positions` 来记录输入/输出的序列信息。 + +于是,在组织神经网络输入时,需要思考完成以下工作: +1. 为每一个输入/输出创建`argument`。 + - C-API 中操作`argument`的接口请查看[argument.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h)。 +1. 为每一个`argument`创建`paddle_matrix`或者`paddle_ivector`来存储数据。 + - C-API 中操作`paddle_ivector`的接口请查看 [vector.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/vector.h)。 + - C-API 中操作`paddle_matrix`的接口请查看[matrix.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/matrix.h)。 +1. 如果输入是序列数据,需要创建并填写`sequence_start_positions`信息。 + - 通过调用 [`paddle_arguments_set_sequence_start_pos`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h#L137) 来为一个`argument`添加序列信息。 + - 通过调用 [`paddle_arguments_get_sequence_start_pos`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h#L150) 来读取一个`argument`添加序列信息。 + - 接口说明请查看 [argument.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h) 文件。 diff --git a/doc/howto/usage/capi/workflow_of_capi_cn.md b/doc/howto/usage/capi/workflow_of_capi_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..c1c2c86d0c5284dd396fdd815b69c47c5f937bf6 --- /dev/null +++ b/doc/howto/usage/capi/workflow_of_capi_cn.md @@ -0,0 +1,120 @@ +## C-API 使用流程 + +这篇文档介绍 PaddlePaddle C-API 整体使用流程。 + +### 使用流程 + +使用 C-API 的工作流程如图1所示,分为(1)准备预测模型和(2)预测程序开发两大部分。 + +

+
图1. C-API使用流程示意图 +

+ +- 准备预测模型 + 1. 只将神经网络结构进行序列化。 + - 只对神经网络结构进行序列化,加载模型需同时指定:网络结构的序列化结果和模型参数存储目录。 + 1. 将网络结构定义和训练结束存储下来的模型参数文件(多个)合并入一个文件。 + - 神经网络模型结构和训练好的模型将被序列化合并入一个文件。 + - 预测时只需加载一个文件便于发布。 + - **注意**:以上两种方式只需选择其一即可。 +- 调用 C-API 开发预测序 + 1. 初始化PaddlePaddle运行环境。 + 1. 加载预测模型。 + 1. 创建神经网络输入,组织输入数据。 + 1. 进行前向计算,获得计算结果。 + 1. 清理和结束。 + +### 准备预测模型 + +在准备预测模型部分的介绍,我们以手写数字识别任务为例。手写数字识别任务定义了一个含有[两个隐层的简单全连接网络](https://github.com/PaddlePaddle/book/blob/develop/02.recognize_digits/README.cn.md#softmax回归softmax-regression),网络接受一幅图片作为输入,将图片分类到 0 ~ 9 类别标签之一。完整代码可以查看[此目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense) 中的相关脚本。 + +调用C-API开发预测程序需要一个训练好的模型,在终端执行`python mnist_v2.py` +运行[目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense) 会使用 PaddlePaddle 内置的 [MNIST 数据集](http://yann.lecun.com/exdb/mnist/)进行训练。训练好的模型默认保存在当前运行目录下的`models`目录中。 + +下面,我们将训练结束后存储下来的模型转换成预测模型。 + +1. 序列化神经网络模型配置 + + PaddlePaddle 使用 protobuf 来传输网络配置文件中定义的网络结构和相关参数,使用 C-API 进行预测时,需要将网络结构使用 protobuf 进行序列化,写入文件中。 + + 调用[`paddle.utils.dump_v2_config`](https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/utils/dump_v2_config.py)中的`dump_v2_config`函数能够将使用 PaddlePaddle V2 API 定义的神经网络结构 dump 到指定文件中,示例代码如下: + + ```python + from paddle.utils.dump_v2_config import dump_v2_config + from mnist_v2 import network + + predict = network(is_infer=True) + dump_v2_config(predict, "trainer_config.bin", True) + ``` + + 对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例,[`mnist_v2.py`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/mnist_v2.py)脚本集成了序列化神经网络结构的过程,可以直接运行 `python mnist_v2.py --task dump_config` 对神经网络结构进行序列化,结果会写入当前运行目录下的`trainer_config.bin`文件中。 + + 使用这种方式,需要**在运行时将神经网络的多个可学习参数放在同一个目录中**,C-API可以通过分别指定序列化后的网络结构文件和参数目录来加载训练好的模型。 + +2. 合并模型文件(可选) + + 一些情况为了便于发布,希望能够将序列化后的神经网络结构和训练好的模型参数打包进一个文件。对于这样的需求,可以使用`paddle.utils.merge_model`中的`merge_v2_model`接口对神经网络结构和训练好的参数进行序列化,将序列化结果写入一个文件内。 + + 代码示例如下: + + ```python + from paddle.utils.merge_model import merge_v2_modelss + from mnist_v2 import network + + net = network(is_infer=True) + param_file = "models/params_pass_4.tar" + output_file = "output.paddle.model" + merge_v2_model(net, param_file, output_file) + ``` + 对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例,可直接运行 `python` [merge_v2_model.py](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/merge_v2_model.py)。序列化结果会写入当前运行目录下的`output.paddle.model`文件中。使用这种方式,运行时C-API可以通过指定`output.paddle.model`文件的路径来加载预测模型。 + +#### 注意事项 +1. 为使用C-API,在调用`dump_v2_config`序列化神经网络结构时,参数`binary`必须指定为`True`。 +1. **预测使用的网络结构往往不同于训练**,通常需要去掉网络中的:(1)类别标签层;(2)损失函数层;(3)`evaluator`等,只留下核心计算层,请注意是否需要修改网络结构。 +1. 预测时,可以获取网络中定义的任意多个(大于等于一个)层前向计算的结果,需要哪些层的计算结果作为输出,就将这些层加入一个Python list中,作为调用`dump_v2_config`的第一个参数。 + +### 编写预测代码 + +预测代码更多详细示例代码请参考[C-API使用示例](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference) 目录下的代码示例。这一节对图1中预测代码编写的5个步骤进行介绍和说明。 + +#### step 1. 初始化PaddlePaddle运行环境 +第一步需调用[`paddle_init`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/main.h#L27) 初始化PaddlePaddle运行环境,该接口接受两个参数:参数的个数和参数列表。 + +#### step2. 加载模型 + +这里介绍C-API使用中的一个重要概念:Gradient Machine。 + +概念上,在 PaddlePaddle 内部,一个GradientMachine类的对象管理着一组计算层(PaddlePaddle Layers)来完成前向和反向计算,并处理与之相关的所有细节。在调用C-API预测时,只需进行前向计算而无需调用反向计算。这篇文档之后部分会使用`gradient machine`来特指调用PaddlePaddle C-API创建的GradientMachine类的对象。每一个 `gradient machine` 都会管理维护一份训练好的模型,下面是C-API提供的,两种常用的模型加载方式: + +1. 调用[`paddle_gradient_machine_load_parameter_from_disk`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L61)接口,从磁盘加载预测模型。这时`gradient machine`会独立拥有一份训练好的模型; +1. 调用[`paddle_gradient_machine_create_shared_param`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L88)接口,与其它`gradient machine`的共享已经加载的预测模型。这种情况多出现在使用多线程预测时,通过多个线程共享同一个模型来减少内存开销。可参考[此示例](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/examples/model_inference/multi_thread/main.c)。 + +- 注意事项 + 1. 使用PaddlePaddle V2 API训练,模型中所有可学习参数会被存为一个压缩文件,需要手动进行解压,将它们放在同一目录中,C-API不会直接加载 V2 API 存储的压缩文件。 + 1. 如果使用`merge model`方式将神经网络结构和训练好的参数序列化到一个文件,请参考此[示例](https://github.com/PaddlePaddle/Mobile/blob/develop/Demo/linux/paddle_image_recognizer.cpp#L59)。 + 1. 通过灵活使用以上两个接口,加载模型可其它多种方式,例如也可在程序运行过程中再加载另外一个模型。 + +#### step 3. 创建神经网络输入,组织输入数据 + +基本使用概念: +- 在PaddlePaddle内部,神经网络中一个计算层的输入输出被组织为一个 `Argument` 结构体,如果神经网络有多个输入或者多个输出,每一个输入/输出都会对应有自己的`Argument`。 +- `Argument` 并不真正“存储”数据,而是将输入/输出数据有机地组织在一起。 +- 在`Argument`内部由:1. `Matrix`(二维矩阵,存储浮点类型输入/输出);2. `IVector`(一维数组,**仅用于存储整型值**,多用于自然语言处理任务)来实际存储数据。 + +C-API支持的所有输入数据类型和他们的组织方式,请参考“输入/输出数据组织”一节。 + +这篇文档的之后部分会使用`argument`来特指PaddlePaddle C-API中神经网络的一个输入/输出,使用`paddle_matrix`**特指**`argument`中用于存储数据的`Matrix`类的对象。 + +在组织神经网络输入,获取输出时,需要思考完成以下工作: +1. 为每一个输入/输出创建`argument`; +1. 为每一个`argument`创建`paddle_matrix`来存储数据; + +与输入不同的是,不需在使用C-API时为输出`argument`的`paddle_matrix`对象分配空间。前向计算之后PaddlePaddle内部已经分配/管理了每个计算层输出的存储空间。 + +#### step 4. 前向计算 + +完成上述准备之后,通过调用 `[paddle_gradient_machine_forward](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L73)` 接口完成神经网络的前向计算。 + +#### step 5. 清理 + +结束预测之后,对使用的中间变量和资源进行清理和释放。 diff --git a/paddle/capi/examples/model_inference/dense/main.c b/paddle/capi/examples/model_inference/dense/main.c index 5eeaf7e31fac7c9ed0b9269e74a7e467bde155ef..376cd46fb09a156d426453986c87dcff6e2f71dd 100644 --- a/paddle/capi/examples/model_inference/dense/main.c +++ b/paddle/capi/examples/model_inference/dense/main.c @@ -3,59 +3,82 @@ #include "../common/common.h" +// Modify this path as needed. #define CONFIG_BIN "./trainer_config.bin" +// Modify this path as needed. +// This demo assumes that merged model is not used, then this path is the +// directory storing all the trained parameters. +// If the model is trained by PaddlePaddle V2 API, the model is saved as +// a compressed file. You need to uncompress the compressed file first. +#define MODEL_PATH "models/pass_4" int main() { - // Initalize Paddle + // Initalize the PaddlePaddle runtime environment. char* argv[] = {"--use_gpu=False"}; CHECK(paddle_init(1, (char**)argv)); - // Reading config binary file. It is generated by `convert_protobin.sh` + // Read the binary configuration file generated by `convert_protobin.sh` long size; void* buf = read_config(CONFIG_BIN, &size); - // Create a gradient machine for inference. + // Create the gradient machine for inference. paddle_gradient_machine machine; CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size)); - CHECK(paddle_gradient_machine_randomize_param(machine)); - // Loading parameter. Uncomment the following line and change the directory. - // CHECK(paddle_gradient_machine_load_parameter_from_disk(machine, - // "./some_where_to_params")); + // Load the trained model. Modify the parameter MODEL_PATH to set the correct + // path of the trained model. + CHECK(paddle_gradient_machine_load_parameter_from_disk(machine, MODEL_PATH)); + + // Inputs and outputs of the network are organized as paddle_arguments object + // in C-API. In the comments below, "argument" specifically means one input of + // the neural network in PaddlePaddle C-API. paddle_arguments in_args = paddle_arguments_create_none(); - // There is only one input of this network. + // There is only one data layer in this demo MNIST network, invoke this + // function to create one argument. CHECK(paddle_arguments_resize(in_args, 1)); - // Create input matrix. - paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1, - /* size */ 784, - /* useGPU */ false); - srand(time(0)); + // Each argument needs one matrix or one ivector (integer vector, for sparse + // index input, usually used in NLP task) to holds the real input data. + // In the comments below, "matrix" specifically means the object needed by + // argument to hold the data. Here we create the matrix for the above created + // agument to store the testing samples. + paddle_matrix mat = + paddle_matrix_create(/* height = batch size */ 1, + /* width = dimensionality of the data layer */ 784, + /* whether to use GPU */ false); paddle_real* array; - - // Get First row. + // Get the pointer pointing to the start address of the first row of the + // created matrix. CHECK(paddle_matrix_get_row(mat, 0, &array)); + // Fill the matrix with a randomly generated test sample. + srand(time(0)); for (int i = 0; i < 784; ++i) { array[i] = rand() / ((float)RAND_MAX); } + // Assign the matrix to the argument. CHECK(paddle_arguments_set_value(in_args, 0, mat)); + // Create the output argument. paddle_arguments out_args = paddle_arguments_create_none(); + + // Invoke the forward computation. CHECK(paddle_gradient_machine_forward(machine, in_args, out_args, - /* isTrain */ false)); - paddle_matrix prob = paddle_matrix_create_none(); + /* is train taks or not */ false)); + // Create the matrix to hold the forward result of the neural network. + paddle_matrix prob = paddle_matrix_create_none(); + // Access the matrix of the output argument, the predicted result is stored in + // which. CHECK(paddle_arguments_get_value(out_args, 0, prob)); uint64_t height; uint64_t width; - CHECK(paddle_matrix_get_shape(prob, &height, &width)); CHECK(paddle_matrix_get_row(prob, 0, &array)); @@ -68,6 +91,7 @@ int main() { } printf("\n"); + // The cleaning up. CHECK(paddle_matrix_destroy(prob)); CHECK(paddle_arguments_destroy(out_args)); CHECK(paddle_matrix_destroy(mat)); diff --git a/paddle/capi/examples/model_inference/dense/merge_v2_model.py b/paddle/capi/examples/model_inference/dense/merge_v2_model.py new file mode 100644 index 0000000000000000000000000000000000000000..c030d572cbdb15cb5e90f2685723a81efb230f81 --- /dev/null +++ b/paddle/capi/examples/model_inference/dense/merge_v2_model.py @@ -0,0 +1,8 @@ +from paddle.utils.merge_model import merge_v2_model + +from mnist_v2 import network + +net = network(is_infer=True) +param_file = "models/params_pass_4.tar" +output_file = "output.paddle.model" +merge_v2_model(net, param_file, output_file) diff --git a/paddle/capi/examples/model_inference/dense/mnist_v2.py b/paddle/capi/examples/model_inference/dense/mnist_v2.py new file mode 100644 index 0000000000000000000000000000000000000000..ee28111153ca2cf24b9789452c65a0f4c7b64538 --- /dev/null +++ b/paddle/capi/examples/model_inference/dense/mnist_v2.py @@ -0,0 +1,117 @@ +import os +import sys +import gzip +import logging +import argparse +from PIL import Image +import numpy as np + +import paddle.v2 as paddle +from paddle.utils.dump_v2_config import dump_v2_config + +logger = logging.getLogger("paddle") +logger.setLevel(logging.INFO) + + +def multilayer_perceptron(img, layer_size, lbl_dim): + for idx, size in enumerate(layer_size): + hidden = paddle.layer.fc(input=(img if not idx else hidden), + size=size, + act=paddle.activation.Relu()) + return paddle.layer.fc(input=hidden, + size=lbl_dim, + act=paddle.activation.Softmax()) + + +def network(input_dim=784, lbl_dim=10, is_infer=False): + images = paddle.layer.data( + name='pixel', type=paddle.data_type.dense_vector(input_dim)) + + predict = multilayer_perceptron( + images, layer_size=[128, 64], lbl_dim=lbl_dim) + + if is_infer: + return predict + else: + label = paddle.layer.data( + name='label', type=paddle.data_type.integer_value(lbl_dim)) + return paddle.layer.classification_cost(input=predict, label=label) + + +def main(task="train", use_gpu=False, trainer_count=1, save_dir="models"): + if task == "train": + if not os.path.exists(save_dir): + os.mkdir(save_dir) + + paddle.init(use_gpu=use_gpu, trainer_count=trainer_count) + cost = network() + parameters = paddle.parameters.create(cost) + optimizer = paddle.optimizer.Momentum( + learning_rate=0.1 / 128.0, + momentum=0.9, + regularization=paddle.optimizer.L2Regularization(rate=0.0005 * 128)) + + trainer = paddle.trainer.SGD(cost=cost, + parameters=parameters, + update_equation=optimizer) + + def event_handler(event): + if isinstance(event, paddle.event.EndIteration): + if event.batch_id % 100 == 0: + logger.info("Pass %d, Batch %d, Cost %f, %s" % + (event.pass_id, event.batch_id, event.cost, + event.metrics)) + if isinstance(event, paddle.event.EndPass): + with gzip.open( + os.path.join(save_dir, "params_pass_%d.tar" % + event.pass_id), "w") as f: + trainer.save_parameter_to_tar(f) + + trainer.train( + reader=paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=8192), + batch_size=128), + event_handler=event_handler, + num_passes=5) + elif task == "dump_config": + predict = network(is_infer=True) + dump_v2_config(predict, "trainer_config.bin", True) + else: + raise RuntimeError(("Error value for parameter task. " + "Available options are: train and dump_config.")) + + +def parse_cmd(): + parser = argparse.ArgumentParser( + description="PaddlePaddle MNIST demo for CAPI.") + parser.add_argument( + "--task", + type=str, + required=False, + help=("A string indicating the taks type. " + "Available options are: \"train\", \"dump_config\"."), + default="train") + parser.add_argument( + "--use_gpu", + type=bool, + help=("A bool flag indicating whether to use GPU device or not."), + default=False) + parser.add_argument( + "--trainer_count", + type=int, + help=("This parameter is only used in training task. It indicates " + "how many computing threads are created in training."), + default=1) + parser.add_argument( + "--save_dir", + type=str, + help=("This parameter is only used in training task. It indicates " + "path of the directory to save the trained models."), + default="models") + return parser.parse_args() + + +if __name__ == "__main__": + args = parse_cmd() + main(args.task, args.use_gpu, args.trainer_count, args.save_dir) diff --git a/paddle/capi/examples/model_inference/sparse_binary/main.c b/paddle/capi/examples/model_inference/sparse_binary/main.c index 8ba67aee560239d3050c7f40198d20df99ec370e..029b94ee63ba282aa48193ffd4f625657ddc3a60 100644 --- a/paddle/capi/examples/model_inference/sparse_binary/main.c +++ b/paddle/capi/examples/model_inference/sparse_binary/main.c @@ -1,5 +1,6 @@ #include #include + #include "../common/common.h" #define CONFIG_BIN "./trainer_config.bin" @@ -9,16 +10,18 @@ int main() { char* argv[] = {"--use_gpu=False"}; CHECK(paddle_init(1, (char**)argv)); - // Reading config binary file. It is generated by `convert_protobin.sh` + // Read the binary configuration file which is generated by + // `convert_protobin.sh` long size; void* buf = read_config(CONFIG_BIN, &size); - // Create a gradient machine for inference. + // Create the gradient machine for inference. paddle_gradient_machine machine; CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size)); CHECK(paddle_gradient_machine_randomize_param(machine)); - // Loading parameter. Uncomment the following line and change the directory. + // Load the trained parameters. Uncomment the following line and change the + // directory as needed. // CHECK(paddle_gradient_machine_load_parameter_from_disk(machine, // "./some_where_to_params")); paddle_arguments in_args = paddle_arguments_create_none(); @@ -26,7 +29,7 @@ int main() { // There is only one input of this network. CHECK(paddle_arguments_resize(in_args, 1)); - // Create input matrix. + // Create the input matrix. paddle_matrix mat = paddle_matrix_create_sparse(1, 784, 3, true, false); srand(time(0)); paddle_real* array; diff --git a/python/paddle/utils/dump_v2_config.py b/python/paddle/utils/dump_v2_config.py new file mode 100644 index 0000000000000000000000000000000000000000..5dc2111e379fd39b40e1e9bcf2e577b57b101a68 --- /dev/null +++ b/python/paddle/utils/dump_v2_config.py @@ -0,0 +1,62 @@ +# Copyright (c) 2016 PaddlePaddle Authors. 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. +import collections + +from paddle.trainer_config_helpers.layers import LayerOutput +from paddle.v2.layer import parse_network +from paddle.proto import TrainerConfig_pb2 + +__all__ = ["dump_v2_config"] + + +def dump_v2_config(topology, save_path, binary=False): + """ Dump the network topology to a specified file. + + This function is only used to dump network defined by using PaddlePaddle V2 + APIs. This function will NOT dump configurations related to PaddlePaddle + optimizer. + + :param topology: The output layers (can be more than one layers given in a + Python List or Tuple) of the entire network. Using the + specified layers (if more than one layer is given) as root, + traversing back to the data layer(s), all the layers + connected to the specified output layers will be dumped. + Layers not connceted to the specified will not be dumped. + :type topology: LayerOutput|List|Tuple + :param save_path: The path to save the dumped network topology. + :type save_path: str + :param binary: Whether to dump the serialized network topology or not. + The default value is false. NOTE that, if you call this + function to generate network topology for PaddlePaddle C-API, + a serialized version of network topology is required. When + using PaddlePaddle C-API, this flag MUST be set to True. + :type binary: bool + """ + + if isinstance(topology, LayerOutput): + topology = [topology] + elif isinstance(topology, collections.Sequence): + for out_layer in topology: + assert isinstance(out_layer, LayerOutput), ( + "The type of each element in the parameter topology " + "should be LayerOutput.") + else: + raise RuntimeError("Error input type for parameter topology.") + + model_str = parse_network(topology) + with open(save_path, "w") as fout: + if binary: + fout.write(model_str.SerializeToString()) + else: + fout.write(str(model_str)) diff --git a/python/paddle/utils/merge_model.py b/python/paddle/utils/merge_model.py index 421e953d2775f145800cf7179ec644697a265060..2b100207728a8532e900992f7db4d3910e893dea 100644 --- a/python/paddle/utils/merge_model.py +++ b/python/paddle/utils/merge_model.py @@ -30,7 +30,8 @@ def merge_v2_model(net, param_file, output_file): which ends with .tar.gz. @param net The output layer of the network for inference. - @param param_file Path of the parameters (.tar.gz) which is stored by v2 api. + @param param_file Path of the parameters (.tar.gz) which is stored by + v2 api. @param output_file Path of the merged file which will be generated. Usage: