diff --git a/Demo/iOS/AICamera/README_cn.md b/Demo/iOS/AICamera/README.cn.md similarity index 93% rename from Demo/iOS/AICamera/README_cn.md rename to Demo/iOS/AICamera/README.cn.md index 036d0b6..8168ed0 100644 --- a/Demo/iOS/AICamera/README_cn.md +++ b/Demo/iOS/AICamera/README.cn.md @@ -51,9 +51,9 @@ App模型需要开启摄像头权限。启动App后，点击屏幕，将会出 ## 快速安装 -如果你想快速地体验PDCamera，可通过扫描以下二维码进行安装。成功识别二维码之后，会自动跳转到安装页面，点击“Install PDCamera”链接，App会自动下载并安装到你的iOS设备上。 +如果你想快速地体验PDCamera，可通过扫描以下二维码进行安装。成功识别二维码之后，会自动跳转到安装页面，点击**“Install PDCamera”**链接，App会自动下载并安装到你的iOS设备上。 -成功安装App后，你还需要安装如下步骤进一步设置：设置 → 通用 → 设备管理 → Baidu USA llc → 信任“Baidu USA llc”。 +成功安装App后，你还需要按照如下步骤设置你的iOS设备：**设置 → 通用 → 设备管理 → Baidu USA llc → 信任“Baidu USA llc”**。

@@ -71,6 +71,8 @@ Github上面只维护了该Demo相关的源码文件和项目配置。用户可 VGG模型的识别精度高，但由于模型较大（104.3MB），需要占用较高的内存（\~800MB），并且识别速度慢（每帧~1.5秒），因此对设备的计算能力要求较高（iPhone6s以上），默认没有添加到项目中。用户也可自行下载[vgg\_ssd\_net.paddle](http://cloud.dlnel.org/filepub/?uuid=1116a5f3-7762-44b5-82bb-9954159cb5d4)，添加到项目中，体验其高精度识别效果。 +这里，我们使用的是**合并的模型**（merged model）。如何从配置文件（例如`config.py`）和训练得到的参数文件（例如`params_pass_0.tar.gz`）得的**合并的模型**文件，请参考[如何合并模型](../../../deployment/model/merge_config_parameters/README.cn.md)。 + ### 准备PaddlePaddle库 用户可按照[iOS平台编译指南](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_ios_cn.md)，拉取[Paddle](https://github.com/PaddlePaddle/Paddle)最新的源码，编译适用于iOS平台的PaddlePaddle库。在执行`make install`之后，PaddlePaddle库将会安装在`CMAKE_INSTALL_PREFIX`所指定的目录下。该目录包含如下子目录： diff --git a/Demo/iOS/AICamera/README.md b/Demo/iOS/AICamera/README.md index 4764b50..8184d1b 100644 --- a/Demo/iOS/AICamera/README.md +++ b/Demo/iOS/AICamera/README.md @@ -1,65 +1,60 @@ # PDCamera iOS Demo with SSD Model +- [Overview](#overview) + - [Pre-trained Models](#pre-trained-models) + - [Demo Screenshot](#demo-screenshot) +- [Fast Installation through QR Code](#fast-installation-through-qr-code) +- [Build from Source Code](#build-from-source-code) + - [Prepare Models](#prepare-models) + - [Prepare PaddlePaddle Inference Library](#prepare-paddlepaddle-inference-library) + - [Directory Tree](#directory-tree) +- [Integrate Paddle C Library to iOS Project](#integrate-paddle-c-library-to-ios-project) + +## Overview + This iOS demo shows PaddlePaddle running SSD(Single Shot MultiBox Detector）Object detection on iOS devices locally and offline. It loads a pretrained model with PaddlePaddle and uses camera to capture images and call PaddlePaddle's inference ability to show detected objects to users. You can look at SSD model architecture [here](https://github.com/PaddlePaddle/models/tree/develop/ssd) and a linux demo [here](https://github.com/PaddlePaddle/Mobile/tree/develop/Demo/linux) +### Pre-trained Models -## Download and run the app +`pascal_mobilenet_300_66` and `vgg_ssd_net` models can classify 20 objects. +`face_mobilenet_160_91` can only classify human's face. -To simply run the demo with iPhone/iPad, scan the QR code below, click "Install PDCamera" in the link and the app will be downloaded in the background. -After installed, go to Settings -> General -> Device Management -> Baidu USA llc -> Trust "Baidu USA llc" +| Model | Dimensions | Accuracy | Size | +| ------------------------ |:----------:| --------:|------:| +| [pascal\_mobilenet\_300\_66.paddle](http://cloud.dlnel.org/filepub/?uuid=39c325d9-b468-4940-ba47-d50c8ec5fd5b) | 300 x 300 | 66% | 23.2MB | +| [vgg\_ssd\_net.paddle](http://cloud.dlnel.org/filepub/?uuid=1116a5f3-7762-44b5-82bb-9954159cb5d4) | 300 x 300 | 71% | 104.3MB | +| [face\_mobilenet\_160\_91.paddle](http://cloud.dlnel.org/filepub/?uuid=038c1dbf-08b3-42a9-b2dc-efccd63859fb) | 160 x 160 | 91% | 18.4MB | +### Demo Screenshot -### QR code link +Simply tap on the screen to toggle settings. - - -### Demo screenshot +- Models: Select Pascal MobileNet 300 or Face MobileNet 160, App will exit, need to launch to restart. +- Camera: Toggle Front/Back Camera. App will exit, need to launch to restart. +- Accuracy Threshold: Adjust threshold to filter more/less objects based on probability. +- Time Refresh Rate: Adjust the time to refresh bounding box more/less frequently. - +

+ + +
+Figure-1 +

Detected object will be highlighted as a bounding box with a classified object label and probability. +## Fast Installation through QR Code -## Classifications -`pascal_mobilenet_300_66` and `vgg_ssd_net` models can only classify following 20 objects: - -- aeroplane -- bicycle -- background -- boat -- bottle -- bus -- car -- cat -- chair -- cow -- diningtable -- dog -- horse -- motorbike -- person -- pottedplant -- sheep -- sofa -- train -- tvmonitor - -`face_mobilenet_160_91` can only classify human's face - - -## Settings - -Simply tap on the screen to toggle settings - -- Models: Select Pascal MobileNet 300 or Face MobileNet 160, App will exit, need to launch to restart. -- Camera: Toggle Front/Back Camera. App will exit, need to launch to restart. -- Accuracy Threshold: Adjust threshold to filter more/less objects based on probability -- Time Refresh Rate: Adjust the time to refresh bounding box more/less frequently +To simply run the demo with iPhone/iPad, scan the QR code below, click "Install PDCamera" in the link and the app will be downloaded in the background. +After installed, go to Settings -> General -> Device Management -> Baidu USA llc -> Trust "Baidu USA llc" +

+ +

-## Development or modify +## Build from Source Code Use latest XCode for development. This demo requires a camera for object detection, therefore you must use a device (iPhone or iPad) for development and testing. Simulators will not work as they cannot access camera. @@ -67,23 +62,7 @@ For developers, feel free to use this as a reference to start a new project. Thi Swift cannot directly call C API, in order to have client in Swift work, create Objective-C briding header and a Objective-C++ wrapper (.mm files) to access paddle APIs. - -## Integrate Paddle C Library to iOS - --Follow this guide [Build PaddlePaddle for iOS](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_ios_cn.md) to generate paddle libs(include, lib, third_party). --Create a folder paddle-ios and add to project root. Put the 3 paddle libs folder under paddle-ios. -- Add the `include` directory to **Header Search Paths** -![image](https://user-images.githubusercontent.com/12538138/32491809-b215cf7a-c37d-11e7-87f8-3d45f07bc63e.png) - -- Add the `Accelerate.framework` or `veclib.framework` to your project, if your PaddlePaddle is built with `IOS_USE_VECLIB_FOR_BLAS=ON` -- Add the libraries of paddle, `libpaddle_capi_layers.a` and `libpaddle_capi_engine.a`, and all the third party libraries to your project - - -- Set `-force_load` for `libpaddle_capi_layers.a` -![image](https://user-images.githubusercontent.com/12538138/32492328-8504ebae-c37f-11e7-98b5-41615519fbb3.png) - - -## Download Models +### Prepare Models Our models are too large to upload to Github. Create a model folder and add to project root. Download [face_mobilenet_160_91.paddle](http://cloud.dlnel.org/filepub/?uuid=038c1dbf-08b3-42a9-b2dc-efccd63859fb) and [pascal_mobilenet_300_66.paddle](http://cloud.dlnel.org/filepub/?uuid=39c325d9-b468-4940-ba47-d50c8ec5fd5b) to the model folder. @@ -92,11 +71,53 @@ Note: Only runs on iPhone6s or above (iPhone 6 or below will crash due to memory If you want to try it out, download [vgg_ssd_net.paddle](http://cloud.dlnel.org/filepub/?uuid=1116a5f3-7762-44b5-82bb-9954159cb5d4), then go to XCode target -> Bulid Phases -> Copy Bundle Resources, click '+' to add vgg_ssd_net.paddle +### Prepare PaddlePaddle Inference Library + +Follow this guide [Build PaddlePaddle for iOS](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_ios_cn.md) to generate paddle libs(include, lib, third_party). +Create a folder paddle-ios and add to project root. Put the 3 paddle libs folder under paddle-ios. + +### Directory Tree + +``` +$ git clone https://github.com/PaddlePaddle/Mobile.git +$ cd Mobile/Demo/iOS/AICamera +$ tree +. +├── AICamera # sources codes +├── PDCamera.xcodeproj +├── README.md +├── README_cn.md +├── assets +├── models # models +│   ├── face_mobilenet_160_91.paddle +│   ├── pascal_mobilenet_300_66.paddle +│   └── vgg_ssd_net.paddle +└── paddle-ios # PaddlePaddle inference library + ├── include + ├── lib + │   ├── libpaddle_capi_engine.a + │   ├── libpaddle_capi_layers.a + │   └── libpaddle_capi_whole.a + └── third_party +``` + +## Integrate Paddle C Library to iOS Project -## Accuracy +- Add the `include` directory to **Header Search Paths** + +

+ +

+ +- Add the `Accelerate.framework` or `veclib.framework` to your project, if your PaddlePaddle is built with `IOS_USE_VECLIB_FOR_BLAS=ON` +- Add the libraries of paddle, `libpaddle_capi_layers.a` and `libpaddle_capi_engine.a`, and all the third party libraries to your project + +

+ +

+ +- Set `-force_load` for `libpaddle_capi_layers.a` -| Model | Dimensions | Accuracy | -| ------------------------ |:----------:| --------:| -| face_mobilenet_160_91 | 160x160 | 91% | -| pascal_mobilenet_300_66 | 300x300 | 66% | -| vgg_ssd_net | 300x300 | 71% | +

+ +

diff --git a/README.cn.md b/README.cn.md new file mode 100644 index 0000000..ba0d6a1 --- /dev/null +++ b/README.cn.md @@ -0,0 +1,47 @@ +# 移动PaddlePaddle + +[![Build Status](https://travis-ci.org/PaddlePaddle/Mobile.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Mobile) +[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/mobile/README.html) +[![Download](https://img.shields.io/badge/download-latest-brightgreen.svg?style=flat)](https://github.com/PaddlePaddle/Mobile/wiki) +[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE) + +PaddlePaddle支持在移动设备上，使用训练好的模型进行离线推断。这里，我们主要介绍如何在移动设备上部署PaddlePaddle推断库，以及移动设备上可以使用到的一些优化方法。 + +## 构建PaddlePaddle库 +PaddlePaddle可以通过原生编译、交叉编译的方式，构建多种移动平台上的推断库。 + +- [Android平台编译指南](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_android_cn.md) +- [iOS平台编译指南](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_ios_cn.md) +- [Rapsberry Pi3平台编译指南](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_raspberry_cn.md) +- NVIDIA Driver PX2平台，采用原生编译的方式，可直接依照[PaddlePaddle源码编译指南](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/build_from_source_cn.rst)进行编译 + +## 使用示例 + +- [命令行示例程序](./benchmark/tool/C/README.cn.md) +- [iOS示例应用：PDCamera](./Demo/iOS/AICamera/README.cn.md) + +## 部署优化方法 +移动端对接入库的大小通常都有要求，在编译PaddlePaddle库时，用户可以通过设置一些编译选项来进行优化。 + +- [如何构建最小的PaddlePaddle推断库](./deployment/library/build_for_minimum_size.md) + +训练得到的模型，可在不降低或者轻微降低模型推断精度的前提下，进行一些变换，优化移动设备上的内存使用和执行效率。 + +- [合并网络中的BN层](./deployment/model/merge_batch_normalization/README.md) +- [压缩模型大小的rounding方法](./deployment/model/rounding/README.md) +- [如何合并模型](./deployment/model/merge_config_parameters/README.cn.md) +- INT8量化方法 + +## 模型压缩 +基于PaddlePaddle框架，可以使用模型压缩训练进一步裁剪模型的大小。 + +- [Pruning稀疏化方法](./model_compression/pruning/README.md) + +## 性能数据 +我们列出一些移动设备上的性能测试数据，给用户参考和对比。 + +- [Mobilenet模型性能数据](./benchmark/README.md) +- ENet模型性能数据 +- [DepthwiseConvolution优化效果](https://github.com/hedaoyuan/Function/blob/master/src/conv/README.md) + +本教程由[PaddlePaddle](https://github.com/PaddlePaddle/Paddle)创作，采用[Apache-2.0 license](LICENSE)许可协议进行许可。 diff --git a/README.md b/README.md index 5ebe531..f6a7b42 100644 --- a/README.md +++ b/README.md @@ -1,33 +1,40 @@ # Mobile [![Build Status](https://travis-ci.org/PaddlePaddle/Mobile.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Mobile) +[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/mobile/README.html) +[![Download](https://img.shields.io/badge/download-latest-brightgreen.svg?style=flat)](https://github.com/PaddlePaddle/Mobile/wiki) [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE) Here mainly describes how to deploy PaddlePaddle to the mobile end, as well as some deployment optimization methods and some benchmark. -## How to build PaddlePaddle for mobile -- Build PaddlePaddle for Android [[Chinese](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_android_cn.md)] [[English](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_android_en.md)] -- Build PaddlePaddle for IOS [[Chinese](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_ios_cn.md)] [[English](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_ios_en.md)] -- Build PaddlePaddle for Raspberry Pi3 [[Chinese](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_raspberry_cn.md)] [[English](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_raspberry_en.md)] -- Build PaddlePaddle for PX2 -- [How to build PaddlePaddle mobile inference library with minimum size.](./deployment/library/build_for_minimum_size.md) +## Build PaddlePaddle +- [Build PaddlePaddle for Android](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_android_en.md) +- [Build PaddlePaddle for IOS](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_ios_en.md) +- [Build PaddlePaddle for Raspberry Pi3](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_raspberry_en.md) +- Build PaddlePaddle for NVIDIA Driver PX2 ## Demo - [A command-line inference demo.](./benchmark/tool/C/README.md) -- [iOS demo of AICamera](./Demo/iOS/AICamera/README.md) +- [iOS demo of PDCamera](./Demo/iOS/AICamera/README.md) ## Deployment optimization methods -- [Merge batch normalization before deploying the model to the mobile.](./deployment/model/merge_batch_normalization/README.md) -- [Compress the model before deploying the model to the mobile.](./deployment/model/rounding/README.md) -- [Merge model config and parameter files into one file.](./deployment/model/merge_config_parameters/README.md) -- How to deploy int8 model in mobile inference with PaddlePaddle. +Optimization for the library: + +- [How to build PaddlePaddle mobile inference library with minimum size.](./deployment/library/build_for_minimum_size.md) + +Optimization for models： + +- [Merge batch normalization layers](./deployment/model/merge_batch_normalization/README.md) +- [Compress the model based on rounding](./deployment/model/rounding/README.md) +- [Merge model's config and parameters](./deployment/model/merge_config_parameters/README.md) +- How to deploy int8 model in mobile inference with PaddlePaddle ## Model compression -- [How to use pruning to train smaller model](./model_compression/pruning/) +- [How to use pruning to train smaller model](./model_compression/pruning/README.md) ## PaddlePaddle mobile benchmark - [Benchmark of Mobilenet](./benchmark/README.md) - Benchmark of ENet -- [Benchmark of DepthwiseConvolution in PaddlePaddle](https://github.com/hedaoyuan/Function/blob/master/src/conv/README.md) +- [Benchmark of DepthwiseConvolution](https://github.com/hedaoyuan/Function/blob/master/src/conv/README.md) This tutorial is contributed by [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) and licensed under the [Apache-2.0 license](LICENSE). diff --git a/benchmark/tool/C/README_cn.md b/benchmark/tool/C/README.cn.md similarity index 81% rename from benchmark/tool/C/README_cn.md rename to benchmark/tool/C/README.cn.md index bd4264b..37cf3f9 100644 --- a/benchmark/tool/C/README_cn.md +++ b/benchmark/tool/C/README.cn.md @@ -9,13 +9,15 @@ - **Step 1，编译Android平台上适用的PaddlePaddle库。** - 用户需要按照[Android平台编译指南](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_android_cn.md)，编译Android平台上适用的PaddlePaddle库。在执行`make install`之后，PaddlePaddle库将会安装在`CMAKE_INSTALL_PREFIX`所指定的目录下。该目录包含如下几个子目录： + 用户可以按照[Android平台编译指南](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/mobile/cross_compiling_for_android_cn.md)，拉取PaddlePaddle最新代码，编译Android平台上适用的PaddlePaddle库。在执行`make install`之后，PaddlePaddle库将会安装在`CMAKE_INSTALL_PREFIX`所指定的目录下。该目录包含如下几个子目录： - `include`，其中包含使用PaddlePaddle所需要引入的头文件，通常代码中加入`#include `即可。 - `lib`，其中包含了PaddlePaddle对应架构的库文件。其中包括： - 动态库，`libpaddle_capi_shared.so`。 - 静态库，`libpaddle_capi_layers.a`和`libpaddle_capi_engine.a`。 - `third_party`，PaddlePaddle所依赖的第三方库。 + 你也可以从[wiki](https://github.com/PaddlePaddle/Mobile/wiki)下载编译好的版本。 + - **Step 2，编译示例程序。** 示例程序项目使用CMake管理，可按照以下步骤，编译Android设备上运行的可执行程序。 @@ -58,8 +60,8 @@ - **Step 3，准备模型。** - Android设备上推荐使用`merged model`。以Mobilenet为例，要生成`merged model`，首先你需要准备以下文件： - - 模型配置文件[mobilenet.py](https://github.com/PaddlePaddle/Mobile/blob/develop/models/mobilenet.py)，它是使用PaddlePaddle的v2 api编写的`Mobilenet`模型的网络结构。用户可在[models](https://github.com/PaddlePaddle/Mobile/tree/develop/models)获取更多PaddlePaddle常用的网络配置，该repo下面同时提供了使用PaddlePaddle训练模型的方法。 + Android设备上推荐使用**合并的模型（merged model）**。以Mobilenet为例，要生成**合并的模型**文件，首先你需要准备以下文件： + - 模型配置文件[mobilenet.py](https://github.com/PaddlePaddle/Mobile/tree/develop/models/standard_network/mobilenet.py)，它是使用PaddlePaddle的v2 api编写的`Mobilenet`模型的网络结构。当前repo的[models](https://github.com/PaddlePaddle/Mobile/tree/develop/models)目录下维护了一些移动端常用的PaddlePaddle网络配置。同时，用户可在[models](https://github.com/PaddlePaddle/models)repo下面找到更多PaddlePaddle常用的网络配置，该repo下面同时提供了使用PaddlePaddle训练模型的方法。 - 模型参数文件。使用PaddlePaddle v2 api训练得到的参数将会存储成`.tar.gz`文件。比如，我们提供了一个使用[flowers102](http://www.robots.ox.ac.uk/~vgg/data/flowers/102/)数据集训练`Mobilnet`分类模型的参数文件[mobilenet\_flowers102.tar.gz](http://cloud.dlnel.org/filepub/?uuid=4a3fcd7a-719c-479f-96e1-28a4c3f2195e)。你也可以使用以下命令下载该参数文件： ```bash @@ -71,7 +73,7 @@ 在准备好模型配置文件（.py）和参数文件（.tar.gz）之后，且所在机器已经成功安装了PaddlePaddle的Python包之后，我们可以通过执行以下脚本生成需要的`merged model`。 ```bash - $ cd Mobile/tools/merge_config_parameters + $ cd Mobile/deployment/model/merge_config_parameters $ python merge_model.py ``` @@ -81,7 +83,7 @@ wget -C http://cloud.dlnel.org/filepub/?uuid=d3b95cf9-4dc3-476f-bdc7-98ac410c4f71 -O mobilenet_flowers102.paddle ``` - 更多有关于生成`merged model`的详情，请参考[merge\_config\_parameters](https://github.com/PaddlePaddle/Mobile/tree/develop/tools/merge_config_parameters)。 + 更多有关于生成`merged model`的详情，请参考[merge\_config\_parameters](https://github.com/PaddlePaddle/Mobile/tree/develop/deployment/model/merge_config_parameters/README.cn.md)。 - **Step 4，在Android设备上测试。** diff --git a/deployment/model/merge_config_parameters/README.cn.md b/deployment/model/merge_config_parameters/README.cn.md new file mode 100644 index 0000000..9c86236 --- /dev/null +++ b/deployment/model/merge_config_parameters/README.cn.md @@ -0,0 +1,56 @@ +# 如何合并模型 + +由PaddlePaddle训练得到的模型，通常包含两个部分：模型配置文件和参数文件。PaddlePaddle提供工具，将配置文件和参数文件合并成一个文件，即这里所说的**合并的模型**文件，方便在移动端上的离线推断应用中使用。 + +针对PaddlePaddle的v1和v2 api，我们分别提供了两套合并模型的工具，下面将分别介绍。 + +## merge\_v2\_model + +这个工具适用于使用v2 api训练的模型。`merge_v2_model`是PaddlePaddle提供的一个python函数，使用该工具，首先你需要安装PaddlePaddl的python包。我们以移动端上常用的`Mobilenet`为例，来介绍这个工具的使用。 + +- **Step 1，准备工作。** + - 准备**模型配置文件：** 用于推断任务的模型配置文件，必须只包含`inference`网络，即不能包含训练网络中需要的`label`、`loss`以及`evaluator`层。我们使用的基于`Mobilenet`的图像分类任务配置文件见[mobilenet.py](../../../models/standard_network/mobilenet.py)。 + + - 准备**参数文件：** 使用PaddlePaddle v2 api训练得到的参数将会存储成`.tar.gz`文件，可直接用于合并模型。我们提供一个使用[flowers102](http://www.robots.ox.ac.uk/~vgg/data/flowers/102/)数据集训练`Mobilenet`分类模型的参数文件[mobilenet_flowers102.tar.gz](http://cloud.dlnel.org/filepub/?uuid=4a3fcd7a-719c-479f-96e1-28a4c3f2195e)。用户可点击参数文件名字通过浏览器下载，或者使用以下命令下载： + + ```bash + wget -C http://cloud.dlnel.org/filepub/?uuid=4a3fcd7a-719c-479f-96e1-28a4c3f2195e -O mobilenet_flowers102.tar.gz + ``` + +- **Step 2，合并模型。** + + 运行python脚本[merge_model.py](./merge_model.py)，即可得到合并的模型`mobilenet_flowers102.paddle`。 + + ```bash + $ cat merge_model.py + import paddle.v2 as paddle + from paddle.utils.merge_model import merge_v2_model + + # import network configuration + from mobilenet import mobile_net + + if __name__ == "__main__": + image_size = 224 + num_classes = 102 + net = mobile_net(3 * image_size * image_size, num_classes, 1.0) + param_file = './mobilenet_flowers102.tar.gz' + output_file = './mobilenet_flowers102.paddle' + merge_v2_model(net, param_file, output_file) + ``` + +## paddle\_merge\_model + +这个工具适用于使用v1 api训练的模型。`paddle_merge_model`是PaddlePaddle提供的一个可执行文件。假设PaddlePaddle的安装目录位于`PADDLE_ROOT`，该工具的使用方法如下： + +```bash +$PADDLE_ROOT/opt/paddle/bin/paddle_merge_model \ + --model_dir="pass-00000" \ + --config_file="config.py" \ + --model_file="output.paddle" +``` + +该工具需要三个参数： + +- `--model_dir`，参数文件所在目录。 +- `--config_file`，`inference`网络配置文件的路径。 +- `--model_file`，生成的**合并的模型**文件的路径。 diff --git a/deployment/model/merge_config_parameters/README.md b/deployment/model/merge_config_parameters/README.md index 4edd515..fa43a7a 100644 --- a/deployment/model/merge_config_parameters/README.md +++ b/deployment/model/merge_config_parameters/README.md @@ -10,8 +10,8 @@ This applies to all PaddlePaddle v2 models, we show a demo of mobilenet. ### Step 1: Prepartions -**Model Config :** [Mobilenet model config](../../models/mobilenet.py). -**Model Parameters:** [Mobilenet model param pretrained on flower102 download](https://pan.baidu.com/s/1geHkrw3) +**Model Config :** [Mobilenet model config](../../../models/standard_network/mobilenet.py). +**Model Parameters:** [Mobilenet model param pretrained on flower102 download](http://cloud.dlnel.org/filepub/?uuid=4a3fcd7a-719c-479f-96e1-28a4c3f2195e). ### Step2: Merge