C-API for inference. #1062

reyoung · 2017-01-04T07:38:31Z

jacquesqiao · 2017-01-04T11:35:18Z

paddle/capi/Arguments.cpp

@@ -0,0 +1,49 @@
+#include "PaddleCAPI.h"


看到大家都把这个目录叫做 "c", 这个作为参考

https://github.com/dmlc/mxnet/tree/master/src/c_api

一般项目目录第一层是编码语言，就比如Paddle

-- paddle |- python |- proto

所以有一些项目的会把这部分放到c目录下。而对于Paddle，一来C/CPP区分界限不明显，二来但就这个C-API来说，本身也是C与C++的混合写法。所以很难独立成C目录。

jacquesqiao · 2017-01-04T11:54:21Z

paddle/capi/PaddleCAPI.h

@@ -0,0 +1,54 @@
+#ifndef __PADDLE_PADDLE_CAPI_PADDLECAPI_H_INCLUDED__


#pragma once 我们的规范貌似是这个？

这个因为是会被用户使用的库，所以还是用 #if 作为guard把。

jacquesqiao · 2017-01-04T11:55:38Z

paddle/utils/tests/test_ClassRegistrar.cpp

@@ -0,0 +1,27 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.


后边的test和前面的关系不大是么

抱歉，这个文件是勿提交的

jacquesqiao · 2017-01-04T11:58:26Z

paddle/capi/PaddleCAPI.h

+extern "C" {
+#endif
+
+#define PD_NO_ERROR 0


是不是定义一个enum更好

typedef enum { PD_NO_ERROR = 0, PD_NULLPTR = 1 } PD_Code;

* See [here](https://gflags.github.io/gflags/#cmake).

reyoung · 2017-01-11T06:18:05Z

@wangkuiyi @hedaoyuan @jacquesqiao This C-API is only used for model inference now. And it is not part of current API design. However, neural network inference API is an urgent issue now.

It just exposes some RAW Paddle APIs to C.

exampleAdd sequence exampleAdd sequence exampleAdd sequence exampleAdd sequence exampleAdd sequence example

Also change previous design file, make concept consistant.

reyoung · 2017-03-26T09:11:15Z

麻烦 @wangkuiyi @jacquesqiao @hedaoyuan @gangliao @QiJune 来Review一下这个PR吧。设计文档在这里

Xreki · 2017-04-12T10:18:41Z

doc/design/multi_language_interface/00.why_plain_c.md

  *width = m->width();
  *height = m->height();
 }
 ```

-其中`paddle/math/matrix.hpp`文件内容为:
+其中`paddle/capi/CMatrix.hpp`文件内容为:


文件名采用哪种命名呢，代码里面是matrix.h和Matrix.cpp，也需要统一一下吧

这个命名确实是统一的。。

对于C的header，为全小写加下划线形式。
对于CPP的source，为大写间隔。

只是这两个文件面向的语言不一样，因而采用更适合那个语言的命名风格。C语言的函数命名同理。

Xreki · 2017-04-12T10:23:20Z

doc/design/multi_language_interface/01.inference_implementation.md

+
+通常，这个结构体包含两个项目。
+
+* `type`是一个类型的标志。对于每种类型，type字段均不尽相同。这样，即使C-API接受的类型全是`void *`，我们也可以确定每一个参数的类型。


type在目前的代码里面好像没有起到什么作用，建议在cast之后检查一下type

目前确实没用，但是每次cast都检查也有点难受。毕竟检查都是要耗时的。。

这个type预留的目的还是为了有一定的多态性。。。

譬如，可以写一个 paddle_destroy(void*),这样可以删除任意类型的paddle对象。或者，paddle_tensor_get_data可以接受matrix和vector都行。。

Xreki · 2017-04-12T10:23:41Z

doc/design/multi_language_interface/00.why_plain_c.md

@@ -70,20 +70,20 @@ extern "C"
 paddle_error paddle_matrix_shape(paddle_matrix matrix,


函数名是不是统一使用paddle__动词词组比较好，代码里面是paddle_matrix_get_shape，arguments里面有几个函数不是。

Xreki · 2017-04-12T10:24:48Z

doc/design/multi_language_interface/01.inference_implementation.md

+
+### libpaddle\_capi_shared.{so, dylib}
+
+`libpaddle_capi_shared`是C-API导出的动态库。这个动态库的连接参数与Paddle的其他二进制(例如`paddle_traienr`)类似。用户可以直接使用这个动态库来引入Paddle C-API。具体使用方法为`-lpaddle_capi_shared`。


typo: paddle_traienr -> paddle_trainer

Xreki · 2017-04-12T11:05:39Z

paddle/capi/Arguments.cpp

+
+using paddle::capi::cast;
+
+#define castArg(v) cast<paddle::capi::CArguments>(v)


上文提到的，castArg时，是否需要检查type是不是kARGUMENTS

Xreki · 2017-04-12T11:43:10Z

paddle/capi/examples/model_inference/dense/main.c

+    printf("%.2f ", array[i]);
+  }
+  printf("\n");
+


没有调用paddle_matrix_destroy等接口释放资源。

确实，原来想程序退出了资源总会被OS回收，但是在example里面这么写确实不好。

Xreki · 2017-04-12T11:46:44Z

paddle/capi/gradient_machine.cpp

+                                      NeuralNetwork* network) {
+  return new MyNeuralNetwork(name, network);
+}
+}  // namespace paddle


MyNeuralNetwork定义在其他地方没有用到。

NerualNetwork本身不能直接new的。他的构造函数是私有的。。这里只是为了把它的构造函数expose出来。

Xreki · 2017-04-12T11:53:58Z

paddle/capi/gradient_machine.cpp

+}
+
+paddle_error paddle_gradient_machine_create_shared_param(
+    paddle_gradient_machine origin,


CGradientMachine是否可以添加config和size这两个成员，这样这个接口可以简化成

paddle_error paddle_gradient_machine_create_shared_param(paddle_gradient_machine origin, paddle_gradient_machine* slave)

也能保证origin肯定用到的是正确的config

share参数并不一定非要配置一样。有可能是两个不同的网络结构，但是参数有共享。

Xreki · 2017-04-12T11:58:27Z

paddle/capi/tests/test_Arguments.cpp

+  paddle_arguments args = paddle_arguments_create_none();
+  ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(args, 1));
+
+  paddle_matrix mat = paddle_matrix_create(128, 64, false);


没有GPU的单测？

确实没有，这就加一下。

Xreki · 2017-04-12T11:59:43Z

paddle/capi/tests/test_Matrix.cpp

@@ -0,0 +1,46 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.


上面那个text_Init.cpp是空的，是误加还是尚未实现？

Start Doing C-API for predict.

fbb1b0e

reyoung added the developing label Jan 4, 2017

jacquesqiao reviewed Jan 4, 2017

View reviewed changes

reyoung added 5 commits January 5, 2017 15:42

Doing C-API

aa6e252

Stash

3fcd81f

Try to use standard way to import gflags.

a873a40

* See [here](https://gflags.github.io/gflags/#cmake).

Merge branch 'feature/use_std_cmake' into feature/c_api

106620e

add unittest for prediction

fdb64ac

reyoung force-pushed the feature/c_api branch from ea607e6 to fdb64ac Compare January 5, 2017 13:20

reyoung added 7 commits January 10, 2017 15:24

Merge branch 'develop' of github.com:baidu/Paddle into feature/c_api

a22c889

Merge branch 'feature/add_third_party_for_gflags' into feature/c_api

657d204

Add style check to target

873368f

Add WITH_C_API option

fe8d5ff

Merge branch 'develop' of github.com:baidu/Paddle into feature/c_api

d23bae7

Fix unittest

06b1a6a

C-API for model inference.

4fd6888

reyoung force-pushed the feature/c_api branch from eca863e to 4fd6888 Compare January 11, 2017 06:04

reyoung added 2 commits January 11, 2017 14:12

Add warning message

005ac1f

Revert unchanged files

3bc0d8b

reyoung requested review from hedaoyuan and wangkuiyi January 11, 2017 06:14

reyoung added enhancement and removed developing labels Jan 11, 2017

reyoung added 3 commits January 12, 2017 14:27

Fix a bug, should be ALL in custom_target

987a908

Add dump binary config

3b5bed6

Fix typo in API.h

0874a7e

reyoung force-pushed the feature/c_api branch from 5cfc28c to 470bbcf Compare March 26, 2017 05:48

reyoung added 4 commits March 26, 2017 14:06

Add sequence exampleAdd sequence exampleAdd sequence exampleAdd sequence

34b3ee3

exampleAdd sequence exampleAdd sequence exampleAdd sequence exampleAdd sequence exampleAdd sequence example

Add model_inference directory

852a94f

Add usage documentation of C-API.

0d73f4c

Merge branch 'develop' of github.com:baidu/Paddle into feature/c_api

6b78a11

reyoung force-pushed the feature/c_api branch from 2879107 to 88d0a2e Compare March 26, 2017 09:00

Add Implementation documentation.

6623096

Also change previous design file, make concept consistant.

reyoung force-pushed the feature/c_api branch from 88d0a2e to 6623096 Compare March 26, 2017 09:02

Add toc

505d207

reyoung requested a review from gangliao March 26, 2017 09:10

reyoung added 2 commits March 26, 2017 17:13

Revert unchanged file.

e7bc880

Find a bug about recommark.

ddbb610

reyoung requested review from gangliao and removed request for gangliao March 27, 2017 03:06

Xreki reviewed Apr 12, 2017

View reviewed changes

reyoung changed the title ~~C-API for predict.~~ C-API for inference. Apr 13, 2017

reyoung added 8 commits April 14, 2017 14:11

Merge branch 'develop' of github.com:baidu/Paddle into feature/c_api

18a3588

Follow comments

87dfc12

Add TODO for GPU unittest

bda2008

Merge branch 'develop' of github.com:baidu/Paddle into feature/c_api

28c4cee

Change name conventions.

91927cc

Merge branch 'develop' of github.com:baidu/Paddle into feature/c_api

d6a7648

Follow comments.

dfd79c8

Typo

4e0f72e

wangkuiyi approved these changes Apr 21, 2017

View reviewed changes

reyoung merged commit f979b12 into PaddlePaddle:develop Apr 21, 2017

reyoung deleted the feature/c_api branch April 25, 2017 12:34

zhhsplendid pushed a commit to zhhsplendid/Paddle that referenced this pull request Sep 25, 2019

update optimizer cn to 1.5 en official (PaddlePaddle#1062)

ef45bca

wangxicoding pushed a commit to wangxicoding/Paddle that referenced this pull request Dec 9, 2021

fix the bug of winodws load bug. (PaddlePaddle#1062)

bca4e22

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

C-API for inference. #1062

C-API for inference. #1062

reyoung commented Jan 4, 2017 •

edited

Loading

jacquesqiao Jan 4, 2017

reyoung Jan 12, 2017

jacquesqiao Jan 4, 2017

reyoung Jan 5, 2017

jacquesqiao Jan 4, 2017

reyoung Jan 5, 2017

jacquesqiao Jan 4, 2017

reyoung Jan 5, 2017

reyoung commented Jan 11, 2017

reyoung commented Mar 26, 2017

Xreki Apr 12, 2017

reyoung Apr 20, 2017

Xreki Apr 12, 2017

reyoung Apr 20, 2017

Xreki Apr 12, 2017

reyoung Apr 20, 2017

Xreki Apr 12, 2017

Xreki Apr 12, 2017

Xreki Apr 12, 2017 •

edited

Loading

reyoung Apr 14, 2017

Xreki Apr 12, 2017

reyoung Apr 14, 2017

Xreki Apr 12, 2017

reyoung Apr 14, 2017

Xreki Apr 12, 2017

reyoung Apr 14, 2017

Xreki Apr 12, 2017

reyoung Apr 14, 2017

		@@ -0,0 +1,54 @@
		#ifndef __PADDLE_PADDLE_CAPI_PADDLECAPI_H_INCLUDED__

		@@ -0,0 +1,27 @@
		/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.


		通常，这个结构体包含两个项目。

		* `type`是一个类型的标志。对于每种类型，type字段均不尽相同。这样，即使C-API接受的类型全是`void *`，我们也可以确定每一个参数的类型。

		@@ -70,20 +70,20 @@ extern "C"
		paddle_error paddle_matrix_shape(paddle_matrix matrix,


		### libpaddle\_capi_shared.{so, dylib}

		`libpaddle_capi_shared`是C-API导出的动态库。这个动态库的连接参数与Paddle的其他二进制(例如`paddle_traienr`)类似。用户可以直接使用这个动态库来引入Paddle C-API。具体使用方法为`-lpaddle_capi_shared`。


		using paddle::capi::cast;

		#define castArg(v) cast<paddle::capi::CArguments>(v)

		@@ -0,0 +1,46 @@
		/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

C-API for inference. #1062

C-API for inference. #1062

Conversation

reyoung commented Jan 4, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

reyoung commented Jan 11, 2017

reyoung commented Mar 26, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Xreki Apr 12, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

reyoung commented Jan 4, 2017 •

edited

Loading

Xreki Apr 12, 2017 •

edited

Loading