Merge branch 'develop' into update_dia_model_in_readme

.gitmodules

-[submodule "fluid/PaddleNLP/LAC"]
-	path = fluid/PaddleNLP/LAC
+[submodule "PaddleNLP/LAC"]
+	path = PaddleNLP/LAC
 	url = https://github.com/baidu/lac.git
-[submodule "fluid/PaddleNLP/SimNet"]
-	path = fluid/PaddleNLP/SimNet
+[submodule "PaddleNLP/SimNet"]
+	path = PaddleNLP/SimNet
 	url = https://github.com/baidu/AnyQ.git
-[submodule "fluid/PaddleNLP/Senta"]
-	path = fluid/PaddleNLP/Senta
+[submodule "PaddleNLP/Senta"]
+	path = PaddleNLP/Senta
 	url = https://github.com/baidu/Senta.git
-[submodule "fluid/PaddleNLP/LARK"]
-	path = fluid/PaddleNLP/LARK
-	url = https://github.com/PaddlePaddle/LARK
-[submodule "fluid/PaddleNLP/knowledge-driven-dialogue"]
-	path = fluid/PaddleNLP/knowledge-driven-dialogue
+[submodule "PaddleNLP/LARK"]
+	path = PaddleNLP/LARK
+	url = https://github.com/PaddlePaddle/LARK.git
+[submodule "PaddleNLP/knowledge-driven-dialogue"]
+	path = PaddleNLP/knowledge-driven-dialogue
 	url = https://github.com/baidu/knowledge-driven-dialogue

AutoDL/HiNAS_models/README.md

+# Image Classification Models
+This directory contains six image classification models, which are models automatically discovered by Baidu Big Data Lab (BDL) Hierarchical Neural Architecture Search project (HiNAS), achieving 96.1% accuracy on CIFAR-10 dataset. These models are divided into two categories. The first three have no skip link, named HiNAS 0-2, and the last three networks contain skip links, which are similar to the shortcut connections in Resnet, named HiNAS 3-5.
+
+---
+## Table of Contents
+- [Installation](#installation)
+- [Data preparation](#data-preparation)
+- [Training a model](#training-a-model)
+- [Model performances](#model-performances)
+
+## Installation
+Running the trainer in current directory requires:
+
+- PadddlePaddle Fluid >= v0.15.0
+- CuDNN >=6.0
+
+If PaddlePaddle and CuDNN in your runtime environment do not meet the requirements, please follow the instructions in [installation document](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html) and make an update.
+
+## Data preparation
+
+When you run the sample code for the first time, the trainer will automatically download the cifar-10 dataset. Please make sure your environment has an internet connection.
+
+The dataset will be downloaded to `dataset/cifar/cifar-10-python.tar.gz` in the same directory as the Trainer. If automatic download fails, you can go to https://www.cs.toronto.edu/~kriz/cifar.html and download cifar-10-python.tar.gz to the location mentioned above.
+
+## Training a model
+
+After the environment is ready, you can train the model. There are two entrances: `train_hinas.py` and `train_hinas_res.py`. The former is used to train Model 0-2 (without skip link), and the latter is used to train Model 3-5 (contains skip link).
+
+Train Model 0~2 (without skip link)：
+```
+python train_hinas.py --model=m_id       # m_id can be 0, 1 or 2.
+```
+Train Model 3~5 (with skip link)：
+```
+python train_hinas_res.py --model=m_id    # m_id can be 0, 1 or 2.
+```
+
+In addition, both `train_hinas.py` and `train_hinas_res.py` support the following parameters:
+
+- **random_flip_left_right**: Random flip image horizontally. (Default: True)
+- **random_flip_up_down**: Randomly flip image vertically. (Default: False)
+- **cutout**: Add cutout action to image. (Default: True)
+- **standardize_image**: Image standardize. (Default: True)
+- **pad_and_cut_image**: Random padding image and then crop back to the original size. (Default: True)
+- **shuffle_image**: Shuffle the order of the input images during training. (Default: True)
+- **lr_max**: Learning rate at the begin of training. (Default: 0.1)
+- **lr_min**: Learning rate at the end of training. (Default: 0.0001)
+- **batch_size**: Training batch size (Default: 128)
+- **num_epochs**: Total training epoch (Default: 200)
+- **weight_decay**: L2 Regularization value (Default: 0.0004)
+- **momentum**: The momentum parameter in momentum optimizer (Default: 0.9)
+- **dropout_rate**: Dropout rate of the dropout layer (Default: 0.5)
+- **bn_decay**: The decay/momentum parameter (or called moving average decay) in batch norm layer (Default: 0.9)
+
+
+## Model performances
+
+Train all six models using same hyperparameters:
+
+- learning rate: 0.1 -> 0.0001 with cosine annealing
+- total epoch: 200
+- batch size: 128
+- L2 decay: 0.000400
+- optimizer: momentum optimizer with m=0.9 and use nesterov
+- preprocess: random horizontal flip + image standardization + cutout
+
+And below is the accuracy on CIFAR-10 dataset：
+
+| model    | round 1 | round 2 | round 3 | max    | avg    |
+|----------|---------|---------|---------|--------|--------|
+| HiNAS-0  | 0.9548  | 0.9520  | 0.9513  | 0.9548 | 0.9527 |
+| HiNAS-1  | 0.9452  | 0.9462  | 0.9420  | 0.9462 | 0.9445 |
+| HiNAS-2  | 0.9508  | 0.9506  | 0.9483  | 0.9508 | 0.9499 |
+| HiNAS-3  | 0.9607  | 0.9623  | 0.9601  | 0.9623 | 0.9611 |
+| HiNAS-4  | 0.9611  | 0.9584  | 0.9586  | 0.9611 | 0.9594 |
+| HiNAS-5  | 0.9578  | 0.9588  | 0.9594  | 0.9594 | 0.9586 |

AutoDL/HiNAS_models/README_cn.md

+# Image Classification Models
+本目录下包含6个图像分类模型，都是百度大数据实验室 Hierarchical Neural Architecture Search (HiNAS) 项目通过机器自动发现的模型，在CIFAR-10数据集上达到96.1%的准确率。这6个模型分为两类，前3个没有skip link，分别命名为 HiNAS 0-2号，后三个网络带有skip link，功能类似于Resnet中的shortcut connection，分别命名 HiNAS 3-5号。
+
+---
+## Table of Contents
+- [Installation](#installation)
+- [Data preparation](#data-preparation)
+- [Training a model](#training-a-model)
+- [Model performances](#model-performances)
+
+## Installation
+最低环境要求:
+
+- PadddlePaddle Fluid >= v0.15.0
+- Cudnn >=6.0
+
+如果您的运行环境无法满足要求，可以参考此文档升级PaddlePaddle：[installation document](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html)
+
+## Data preparation
+
+第一次训练模型的时候，Trainer会自动下载CIFAR-10数据集，请确保您的环境有互联网连接。
+
+数据集会被下载到Trainer同目录下的`dataset/cifar/cifar-10-python.tar.gz`，如果自动下载失败，您可以自行从 https://www.cs.toronto.edu/~kriz/cifar.html 下载cifar-10-python.tar.gz，然后放到上述位置。
+
+
+## Training a model
+准备好环境后，可以训练模型，训练有2个入口，`train_hinas.py`和`train_hinas_res.py`，前者用来训练0-2号不含skip link的模型，后者用来训练3-5号包含skip link的模型。
+
+训练0~2号不含skip link的模型：
+```
+python train_hinas.py --model=m_id       # m_id can be 0, 1 or 2.
+```
+训练3~5号包含skip link的模型：
+```
+python train_hinas_res.py --model=m_id    # m_id can be 0, 1 or 2.
+```
+
+此外，`train_hinas.py`和`train_hinas_res.py` 都支持以下参数：
+
+初始化部分：
+
+- random_flip_left_right：图片随机水平翻转（Default：True）
+- random_flip_up_down：图片随机垂直翻转（Default：False）
+- cutout：图片随机遮挡（Default：True）
+- standardize_image：对图片每个像素做 standardize（Default：True）
+- pad_and_cut_image：图片随机padding，并裁剪回原大小（Default：True）
+- shuffle_image：训练时对输入图片的顺序做shuffle（Default：True）
+- lr_max：训练开始时的learning rate（Default：0.1）
+- lr_min：训练结束时的learning rate（Default：0.0001）
+- batch_size：训练的batch size（Default：128）
+- num_epochs：训练总的epoch（Default：200）
+- weight_decay：训练时L2 Regularization大小（Default：0.0004）
+- momentum：momentum优化器中的momentum系数（Default：0.9）
+- dropout_rate：dropout层的dropout_rate（Default：0.5）
+- bn_decay：batch norm层的decay/momentum系数（即moving average decay）大小（Default：0.9）
+
+
+
+## Model performances
+6个模型使用相同的参数训练：
+
+- learning rate: 0.1 -> 0.0001 with cosine annealing
+- total epoch: 200
+- batch size: 128
+- L2 decay: 0.000400
+- optimizer: momentum optimizer with m=0.9 and use nesterov
+- preprocess: random horizontal flip + image standardization + cutout
+
+以下是6个模型在CIFAR-10数据集上的准确率：
+
+| model    | round 1 | round 2 | round 3 | max    | avg    |
+|----------|---------|---------|---------|--------|--------|
+| HiNAS-0  | 0.9548  | 0.9520  | 0.9513  | 0.9548 | 0.9527 |
+| HiNAS-1  | 0.9452  | 0.9462  | 0.9420  | 0.9462 | 0.9445 |
+| HiNAS-2  | 0.9508  | 0.9506  | 0.9483  | 0.9508 | 0.9499 |
+| HiNAS-3  | 0.9607  | 0.9623  | 0.9601  | 0.9623 | 0.9611 |
+| HiNAS-4  | 0.9611  | 0.9584  | 0.9586  | 0.9611 | 0.9594 |
+| HiNAS-5  | 0.9578  | 0.9588  | 0.9594  | 0.9594 | 0.9586 |

AutoDL/LRC/README.md

+# LRC Local Rademachar Complexity Regularization
+Regularization of Deep Neural Networks(DNNs) for the sake of improving their generalization capability is important and chllenging. This directory contains image classification model based on a novel regularizer rooted in Local Rademacher Complexity (LRC). We appreciate the contribution by [DARTS](https://arxiv.org/abs/1806.09055) for our research. The regularization by LRC and DARTS are combined in this model on CIFAR-10 dataset. Code accompanying the paper
+> [An Empirical Study on Regularization of Deep Neural Networks by Local Rademacher Complexity](https://arxiv.org/abs/1902.00873)\
+> Yingzhen Yang, Xingjian Li, Jun Huan.\
+> _arXiv:1902.00873_.
+
+---
+# Table of Contents
+
+- [Installation](#installation)
+- [Data preparation](#data-preparation)
+- [Training](#training)
+
+## Installation
+
+Running sample code in this directory requires PaddelPaddle Fluid v.1.2.0 and later. If the PaddlePaddle on your device is lower than this version, please follow the instructions in [installation document](http://www.paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html#paddlepaddle) and make an update.
+
+## Data preparation
+
+When you want to use the cifar-10 dataset for the first time, you can download the dataset as:
+
+    sh ./dataset/download.sh
+
+Please make sure your environment has an internet connection.
+
+The dataset will be downloaded to `dataset/cifar/cifar-10-batches-py` in the same directory as the `train.py`. If automatic download fails, you can download cifar-10-python.tar.gz from https://www.cs.toronto.edu/~kriz/cifar.html and decompress it to the location mentioned above.
+
+
+## Training
+
+After data preparation, one can start the training step by:
+
+    python -u train_mixup.py \
+        --batch_size=80 \
+        --auxiliary \
+        --weight_decay=0.0003 \
+        --learning_rate=0.025 \
+        --lrc_loss_lambda=0.7 \
+        --cutout
+- Set ```export CUDA_VISIBLE_DEVICES=0``` to specifiy one GPU to train.
+- For more help on arguments:
+
+    python train_mixup.py --help
+
+**data reader introduction:**
+
+* Data reader is defined in `reader.py`.
+* Reshape the images to 32 * 32.
+* In training stage, images are padding to 40 * 40 and cropped randomly to the original size.
+* In training stage, images are horizontally random flipped.
+* Images are standardized to (0, 1).
+* In training stage, cutout images randomly.
+* Shuffle the order of the input images during training.
+
+**model configuration:**
+
+* Use auxiliary loss and auxiliary\_weight=0.4.
+* Use dropout and drop\_path\_prob=0.2.
+* Set lrc\_loss\_lambda=0.7.
+
+**training strategy:**
+
+*  Use momentum optimizer with momentum=0.9.
+*  Weight decay is 0.0003.
+*  Use cosine decay with init\_lr=0.025.
+*  Total epoch is 600.
+*  Use Xaiver initalizer to weight in conv2d, Constant initalizer to weight in batch norm and Normal initalizer to weight in fc.
+*  Initalize bias in batch norm and fc to zero constant and do not add bias to conv2d.
+
+
+## Reference
+
+  - DARTS: Differentiable Architecture Search [`paper`](https://arxiv.org/abs/1806.09055)
+  - Differentiable architecture search in PyTorch [`code`](https://github.com/quark0/darts)

AutoDL/LRC/README_cn.md

+# LRC 局部Rademachar复杂度正则化
+为了在深度神经网络中提升泛化能力，正则化的选择十分重要也具有挑战性。本目录包括了一种基于局部rademacher复杂度的新型正则（LRC）的图像分类模型。十分感谢[DARTS](https://arxiv.org/abs/1806.09055)模型对本研究提供的帮助。该模型将LRC正则和DARTS网络相结合，在CIFAR-10数据集中得到了很出色的效果。代码和文章一同发布
+> [An Empirical Study on Regularization of Deep Neural Networks by Local Rademacher Complexity](https://arxiv.org/abs/1902.00873)\
+> Yingzhen Yang, Xingjian Li, Jun Huan.\
+> _arXiv:1902.00873_.
+
+---
+# 内容
+
+- [安装](#安装)
+- [数据准备](#数据准备)
+- [模型训练](#模型训练)
+
+## 安装
+
+在当前目录下运行样例代码需要PadddlePaddle Fluid的v.1.2.0或以上的版本。如果你的运行环境中的PaddlePaddle低于此版本，请根据[安装文档](http://www.paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html#paddlepaddle)中的说明来更新PaddlePaddle。
+
+## 数据准备
+
+第一次使用CIFAR-10数据集时，您可以通过如果命令下载：
+
+    sh ./dataset/download.sh
+
+请确保您的环境有互联网连接。数据会下载到`train.py`同目录下的`dataset/cifar/cifar-10-batches-py`。如果下载失败，您可以自行从https://www.cs.toronto.edu/~kriz/cifar.html上下载cifar-10-python.tar.gz并解压到上述位置。
+
+## 模型训练
+
+数据准备好后，可以通过如下命令开始训练：
+
+    python -u train_mixup.py \
+        --batch_size=80 \
+        --auxiliary \
+        --weight_decay=0.0003 \
+        --learning_rate=0.025 \
+        --lrc_loss_lambda=0.7 \
+        --cutout
+- 通过设置 ```export CUDA_VISIBLE_DEVICES=0```指定单张GPU训练。
+- 可选参数见：
+
+    python train_mixup.py --help
+
+**数据读取器说明：**
+
+* 数据读取器定义在`reader.py`中
+* 输入图像尺寸统一变换为32 * 32
+* 训练时将图像填充为40 * 40然后随机剪裁为原输入图像大小
+* 训练时图像随机水平翻转
+* 对图像每个像素做归一化处理
+* 训练时对图像做随机遮挡
+* 训练时对输入图像做随机洗牌
+
+**模型配置：**
+
+* 使用辅助损失，辅助损失权重为0.4
+* 使用dropout，随机丢弃率为0.2
+* 设置lrc\_loss\_lambda为0.7
+
+**训练策略：**
+
+* 采用momentum优化算法训练，momentum=0.9
+* 权重衰减系数为0.0001
+* 采用正弦学习率衰减，初始学习率为0.025
+* 总共训练600轮
+* 对卷积权重采用Xaiver初始化，对batch norm权重采用固定初始化，对全连接层权重采用高斯初始化
+* 对batch norm和全连接层偏差采用固定初始化，不对卷积设置偏差
+
+
+## 引用
+
+  - DARTS: Differentiable Architecture Search [`论文`](https://arxiv.org/abs/1806.09055)
+  - Differentiable Architecture Search in PyTorch [`代码`](https://github.com/quark0/darts)

fluid/README.md → PaddleCV/README.md

-Fluid 模型库
-============
+PaddleCV
+========
 
 图像分类
 --------
 
 图像分类是根据图像的语义信息对不同类别图像进行区分，是计算机视觉中重要的基础问题，是物体检测、图像分割、物体跟踪、行为分析、人脸识别等其他高层视觉任务的基础，在许多领域都有着广泛的应用。如：安防领域的人脸识别和智能视频分析等，交通领域的交通场景识别，互联网领域基于内容的图像检索和相册自动归类，医学领域的图像识别等。
 
-在深度学习时代，图像分类的准确率大幅度提升，在图像分类任务中，我们向大家介绍了如何在经典的数据集ImageNet上，训练常用的模型，包括AlexNet、VGG、GoogLeNet、ResNet、Inception-v4、MobileNet、DPN(Dual Path Network)、SE-ResNeXt模型，也开源了[训练的模型](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleCV/image_classification/README_cn.md#已有模型及其性能) 方便用户下载使用。同时提供了能够将Caffe模型转换为PaddlePaddle
+在深度学习时代，图像分类的准确率大幅度提升，在图像分类任务中，我们向大家介绍了如何在经典的数据集ImageNet上，训练常用的模型，包括AlexNet、VGG、GoogLeNet、ResNet、Inception-v4、MobileNet、DPN(Dual Path Network)、SE-ResNeXt模型，也开源了[训练的模型](https://github.com/PaddlePaddle/models/blob/develop/PaddleCV/image_classification/README_cn.md#已有模型及其性能) 方便用户下载使用。同时提供了能够将Caffe模型转换为PaddlePaddle
 Fluid模型配置和参数文件的工具。
 
--  [AlexNet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)
--  [VGG](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)
--  [GoogleNet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)
--  [Residual Network](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)
--  [Inception-v4](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)
--  [MobileNet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)
--  [Dual Path Network](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)
--  [SE-ResNeXt](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)
--  [Caffe模型转换为Paddle Fluid配置和模型文件工具](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/caffe2fluid)
+-  [AlexNet](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification/models)
+-  [VGG](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification/models)
+-  [GoogleNet](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification/models)
+-  [Residual Network](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification/models)
+-  [Inception-v4](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification/models)
+-  [MobileNet](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification/models)
+-  [Dual Path Network](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification/models)
+-  [SE-ResNeXt](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification/models)
+-  [Caffe模型转换为Paddle Fluid配置和模型文件工具](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/caffe2fluid)
 
 目标检测
 --------
@@ -32,10 +32,10 @@ Faster RCNN模型是典型的两阶段目标检测器，相较于传统提取区
 
 Mask RCNN模型是基于Faster RCNN模型的经典实例分割模型，在原有Faster RCNN模型基础上添加分割分支，得到掩码结果，实现了掩码和类别预测关系的解藕。
 
--  [Single Shot MultiBox Detector](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleCV/object_detection/README_cn.md)
--  [Face Detector: PyramidBox](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/face_detection/README_cn.md)
--  [Faster RCNN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/rcnn/README_cn.md)
--  [Mask RCNN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/rcnn/README_cn.md)
+-  [Single Shot MultiBox Detector](https://github.com/PaddlePaddle/models/blob/develop/PaddleCV/object_detection/README_cn.md)
+-  [Face Detector: PyramidBox](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/face_detection/README_cn.md)
+-  [Faster RCNN](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/rcnn/README_cn.md)
+-  [Mask RCNN](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/rcnn/README_cn.md)
 
 图像语义分割
 ------------
@@ -45,7 +45,7 @@ Mask RCNN模型是基于Faster RCNN模型的经典实例分割模型，在原有
 在图像语义分割任务中，我们介绍如何基于图像级联网络(Image Cascade
 Network,ICNet)进行语义分割，相比其他分割算法，ICNet兼顾了准确率和速度。
 
--  [ICNet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/icnet)
+-  [ICNet](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/icnet)
 
 图像生成
 -----------
@@ -55,8 +55,8 @@ Network,ICNet)进行语义分割，相比其他分割算法，ICNet兼顾了准
 
 在图像生成任务中，我们介绍了如何使用DCGAN和ConditioanlGAN来进行手写数字的生成，另外还介绍了用于风格迁移的CycleGAN.
 
-- [DCGAN & ConditionalGAN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/gan/c_gan)
-- [CycleGAN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/gan/cycle_gan)
+- [DCGAN & ConditionalGAN](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/gan/c_gan)
+- [CycleGAN](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/gan/cycle_gan)
 
 场景文字识别
 ------------
@@ -65,8 +65,8 @@ Network,ICNet)进行语义分割，相比其他分割算法，ICNet兼顾了准
 
 在场景文字识别任务中，我们介绍如何将基于CNN的图像特征提取和基于RNN的序列翻译技术结合，免除人工定义特征，避免字符分割，使用自动学习到的图像特征，完成字符识别。当前，介绍了CRNN-CTC模型和基于注意力机制的序列到序列模型。
 
--  [CRNN-CTC模型](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/ocr_recognition)
--  [Attention模型](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/ocr_recognition)
+-  [CRNN-CTC模型](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/ocr_recognition)
+-  [Attention模型](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/ocr_recognition)
 
 
 度量学习
@@ -75,7 +75,7 @@ Network,ICNet)进行语义分割，相比其他分割算法，ICNet兼顾了准
 
 度量学习也称作距离度量学习、相似度学习，通过学习对象之间的距离，度量学习能够用于分析对象时间的关联、比较关系，在实际问题中应用较为广泛，可应用于辅助分类、聚类问题，也广泛用于图像检索、人脸识别等领域。以往，针对不同的任务，需要选择合适的特征并手动构建距离函数，而度量学习可根据不同的任务来自主学习出针对特定任务的度量距离函数。度量学习和深度学习的结合，在人脸识别/验证、行人再识别(human Re-ID)、图像检索等领域均取得较好的性能，在这个任务中我们主要介绍了基于Fluid的深度度量学习模型，包含了三元组、四元组等损失函数。
 
-- [Metric Learning](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/metric_learning)
+- [Metric Learning](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/metric_learning)
 
 
 视频分类
@@ -84,89 +84,4 @@ Network,ICNet)进行语义分割，相比其他分割算法，ICNet兼顾了准
 视频分类是视频理解任务的基础，与图像分类不同的是，分类的对象不再是静止的图像，而是一个由多帧图像构成的、包含语音数据、包含运动信息等的视频对象，因此理解视频需要获得更多的上下文信息，不仅要理解每帧图像是什么、包含什么，还需要结合不同帧，知道上下文的关联信息。视频分类方法主要包含基于卷积神经网络、基于循环神经网络、或将这两者结合的方法。该任务中我们介绍基于Fluid的视频分类模型，目前包含Temporal Segment Network(TSN)模型，后续会持续增加更多模型。
 
 
-- [TSN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/video_classification)
-
-
-语音识别
---------
-
-自动语音识别（Automatic Speech Recognition, ASR）是将人类声音中的词汇内容转录成计算机可输入的文字的技术。语音识别的相关研究经历了漫长的探索过程，在HMM/GMM模型之后其发展一直较为缓慢，随着深度学习的兴起，其迎来了春天。在多种语言识别任务中，将深度神经网络(DNN)作为声学模型，取得了比GMM更好的性能，使得 ASR 成为深度学习应用最为成功的领域之一。而由于识别准确率的不断提高，有越来越多的语言技术产品得以落地，例如语言输入法、以智能音箱为代表的智能家居设备等 — 基于语言的交互方式正在深刻的改变人类的生活。
-
-与 [DeepSpeech](https://github.com/PaddlePaddle/DeepSpeech) 中深度学习模型端到端直接预测字词的分布不同，本实例更接近传统的语言识别流程，以音素为建模单元，关注语言识别中声学模型的训练，利用[kaldi](http://www.kaldi-asr.org) 进行音频数据的特征提取和标签对齐，并集成 kaldi 的解码器完成解码。
-
--  [DeepASR](https://github.com/PaddlePaddle/models/blob/develop/fluid/DeepASR/README_cn.md)
-
-机器翻译
---------
-
-机器翻译（Machine Translation）将一种自然语言(源语言)转换成一种自然语言（目标语言），是自然语言处理中非常基础和重要的研究方向。在全球化的浪潮中，机器翻译在促进跨语言文明的交流中所起的重要作用是不言而喻的。其发展经历了统计机器翻译和基于神经网络的神经机器翻译(Nueural
-Machine Translation, NMT)等阶段。在 NMT 成熟后，机器翻译才真正得以大规模应用。而早阶段的 NMT 主要是基于循环神经网络 RNN 的，其训练过程中当前时间步依赖于前一个时间步的计算，时间步之间难以并行化以提高训练速度。因此，非 RNN 结构的 NMT 得以应运而生，例如基 卷积神经网络 CNN 的结构和基于自注意力机制（Self-Attention）的结构。
-
-本实例所实现的 Transformer 就是一个基于自注意力机制的机器翻译模型，其中不再有RNN或CNN结构，而是完全利用 Attention 学习语言中的上下文依赖。相较于RNN/CNN, 这种结构在单层内计算复杂度更低、易于并行化、对长程依赖更易建模，最终在多种语言之间取得了最好的翻译效果。
-
--  [Transformer](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleNLP/neural_machine_translation/transformer/README_cn.md)
-
-强化学习
---------
-
-强化学习是近年来一个愈发重要的机器学习方向，特别是与深度学习相结合而形成的深度强化学习(Deep Reinforcement Learning, DRL)，取得了很多令人惊异的成就。人们所熟知的战胜人类顶级围棋职业选手的 AlphaGo 就是 DRL 应用的一个典型例子，除游戏领域外，其它的应用还包括机器人、自然语言处理等。
-
-深度强化学习的开山之作是在Atari视频游戏中的成功应用， 其可直接接受视频帧这种高维输入并根据图像内容端到端地预测下一步的动作，所用到的模型被称为深度Q网络(Deep Q-Network, DQN)。本实例就是利用PaddlePaddle Fluid这个灵活的框架，实现了 DQN 及其变体，并测试了它们在 Atari 游戏中的表现。
-
--  [DeepQNetwork](https://github.com/PaddlePaddle/models/blob/develop/fluid/DeepQNetwork/README_cn.md)
-
-中文词法分析
-------------
-
-中文分词(Word Segmentation)是将连续的自然语言文本，切分出具有语义合理性和完整性的词汇序列的过程。因为在汉语中，词是承担语义的最基本单位，切词是文本分类、情感分析、信息检索等众多自然语言处理任务的基础。 词性标注（Part-of-speech Tagging）是为自然语言文本中的每一个词汇赋予一个词性的过程，这里的词性包括名词、动词、形容词、副词等等。 命名实体识别（Named Entity Recognition，NER）又称作“专名识别”，是指识别自然语言文本中具有特定意义的实体，主要包括人名、地名、机构名、专有名词等。 我们将这三个任务统一成一个联合任务，称为词法分析任务，基于深度神经网络，利用海量标注语料进行训练，提供了一个端到端的解决方案。
-
-我们把这个联合的中文词法分析解决方案命名为LAC。LAC既可以认为是Lexical Analysis of Chinese的首字母缩写，也可以认为是LAC Analyzes Chinese的递归缩写。
-
-- [LAC](https://github.com/baidu/lac/blob/master/README.md)
-
-情感倾向分析
-------------
-
-情感倾向分析针对带有主观描述的中文文本，可自动判断该文本的情感极性类别并给出相应的置信度。情感类型分为积极、消极、中性。情感倾向分析能够帮助企业理解用户消费习惯、分析热点话题和危机舆情监控，为企业提供有力的决策支持。本次我们开放 AI 开放平台中情感倾向分析采用的[模型](http://ai.baidu.com/tech/nlp/sentiment_classify)，提供给用户使用。
-
-- [Senta](https://github.com/baidu/Senta/blob/master/README.md)
-
-语义匹配
---------
-
-在自然语言处理很多场景中，需要度量两个文本在语义上的相似度，这类任务通常被称为语义匹配。例如在搜索中根据查询与候选文档的相似度对搜索结果进行排序，文本去重中文本与文本相似度的计算，自动问答中候选答案与问题的匹配等。
-
-本例所开放的DAM (Deep Attention Matching Network)为百度自然语言处理部发表于ACL-2018的工作，用于检索式聊天机器人多轮对话中应答的选择。DAM受Transformer的启发，其网络结构完全基于注意力(attention)机制，利用栈式的self-attention结构分别学习不同粒度下应答和语境的语义表示，然后利用cross-attention获取应答与语境之间的相关性，在两个大规模多轮对话数据集上的表现均好于其它模型。
-
-- [Deep Attention Matching Network](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleNLP/deep_attention_matching_net)
-
-AnyQ
-----
-
-[AnyQ](https://github.com/baidu/AnyQ)(ANswer Your Questions) 开源项目主要包含面向FAQ集合的问答系统框架、文本语义匹配工具SimNet。 问答系统框架采用了配置化、插件化的设计，各功能均通过插件形式加入，当前共开放了20+种插件。开发者可以使用AnyQ系统快速构建和定制适用于特定业务场景的FAQ问答系统，并加速迭代和升级。
-
-SimNet是百度自然语言处理部于2013年自主研发的语义匹配框架，该框架在百度各产品上广泛应用，主要包括BOW、CNN、RNN、MM-DNN等核心网络结构形式，同时基于该框架也集成了学术界主流的语义匹配模型，如MatchPyramid、MV-LSTM、K-NRM等模型。使用SimNet构建出的模型可以便捷的加入AnyQ系统中，增强AnyQ系统的语义匹配能力。
-
--  [SimNet in PaddlePaddle Fluid](https://github.com/baidu/AnyQ/blob/master/tools/simnet/train/paddle/README.md)
-
-机器阅读理解
-----------
-
-机器阅读理解(MRC)是自然语言处理(NLP)中的核心任务之一，最终目标是让机器像人类一样阅读文本，提炼文本信息并回答相关问题。深度学习近年来在NLP中得到广泛使用，也使得机器阅读理解能力在近年有了大幅提高，但是目前研究的机器阅读理解都采用人工构造的数据集，以及回答一些相对简单的问题，和人类处理的数据还有明显差距，因此亟需大规模真实训练数据推动MRC的进一步发展。
-
-百度阅读理解数据集是由百度自然语言处理部开源的一个真实世界数据集，所有的问题、原文都来源于实际数据(百度搜索引擎数据和百度知道问答社区)，答案是由人类回答的。每个问题都对应多个答案，数据集包含200k问题、1000k原文和420k答案，是目前最大的中文MRC数据集。百度同时开源了对应的阅读理解模型，称为DuReader，采用当前通用的网络分层结构，通过双向attention机制捕捉问题和原文之间的交互关系，生成query-aware的原文表示，最终基于query-aware的原文表示通过point network预测答案范围。
-
--  [DuReader in PaddlePaddle Fluid](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleNLP/machine_reading_comprehension/README.md)
-
-个性化推荐
--------
-
-推荐系统在当前的互联网服务中正在发挥越来越大的作用，目前大部分电子商务系统、社交网络，广告推荐，搜索引擎，都不同程度的使用了各种形式的个性化推荐技术，帮助用户快速找到他们想要的信息。
-
-在工业可用的推荐系统中，推荐策略一般会被划分为多个模块串联执行。以新闻推荐系统为例，存在多个可以使用深度学习技术的环节，例如新闻的自动化标注，个性化新闻召回，个性化匹配与排序等。PaddlePaddle对推荐算法的训练提供了完整的支持，并提供了多种模型配置供用户选择。
-
-- [TagSpace](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleRec/tagspace)
-- [GRU4Rec](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleRec/gru4rec)
-- [SequenceSemanticRetrieval](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleRec/ssr)
-- [DeepCTR](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleRec/ctr/README.cn.md)
-- [Multiview-Simnet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleRec/multiview_simnet)
+- [TSN](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/video_classification)

PaddleCV/adversarial/README.md

+The minimum PaddlePaddle version needed for the code sample in this directory is the lastest develop branch. If you are on a version of PaddlePaddle earlier than this, [please update your installation](http://www.paddlepaddle.org/docs/develop/documentation/en/build_and_install/pip_install_en.html).
+
+---
+
+# Advbox
+
+Advbox is a toolbox to generate adversarial examples that fool neural networks and Advbox can benchmark the robustness of machine learning models.
+
+The Advbox is based on [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) Fluid and is under continual development, always welcoming contributions of the latest method of adversarial attacks and defenses.
+
+
+## Overview
+[Szegedy et al.](https://arxiv.org/abs/1312.6199) discovered an intriguing properties of deep neural networks in the context of image classification for the first time. They showed that despite the state-of-the-art deep networks are surprisingly susceptible to adversarial attacks in the form of small perturbations to images that remain (almost) imperceptible to human vision system. These perturbations are found by optimizing the input to maximize the prediction error and the images modified by these perturbations are called as `adversarial examples`. The profound implications of these results triggered a wide interest of researchers in adversarial attacks and their defenses for deep learning in general.
+
+Advbox is similar to [Foolbox](https://github.com/bethgelab/foolbox) and [CleverHans](https://github.com/tensorflow/cleverhans). CleverHans only supports TensorFlow framework while foolbox interfaces with many popular machine learning frameworks such as PyTorch, Keras, TensorFlow, Theano, Lasagne and MXNet. However, these two great libraries don't support PaddlePaddle, an easy-to-use, efficient, flexible and scalable deep learning platform which is originally developed by Baidu scientists and engineers for the purpose of applying deep learning to many products at Baidu.
+
+## Usage
+Advbox provides many stable reference implementations of modern methods to generate adversarial examples such as FGSM, DeepFool, JSMA. When you want to benchmark the robustness of your neural networks , you can use the advbox to generate some adversarial examples and benchmark the networks. Some tips of using Advbox:
+
+1. Train a model and save the parameters.
+2. Load the parameters which has been trained，then reconstruct the model.
+3. Use advbox to generate the adversarial samples.
+
+
+#### Dependencies
+* PaddlePaddle: [the lastest develop branch](http://www.paddlepaddle.org/docs/develop/documentation/en/build_and_install/pip_install_en.html)
+* Python 2.x
+
+#### Structure
+
+Network models, attack method's implements and the criterion that defines adversarial examples are three essential elements to generate adversarial examples. Misclassification is adopted as the adversarial criterion for briefness in Advbox.
+
+The structure of Advbox module are as follows:
+
+    .
+    ├── advbox
+    |   ├── __init__.py
+    |   ├── attack
+    |        ├── __init__.py
+    |        ├── base.py
+    |        ├── deepfool.py
+    |        ├── gradient_method.py
+    |        ├── lbfgs.py
+    |        └── saliency.py
+    |   ├── models
+    |        ├── __init__.py
+    |        ├── base.py
+    |        └── paddle.py
+    |   └── adversary.py
+    ├── tutorials
+    |   ├── __init__.py
+    |   ├── mnist_model.py
+    |   ├── mnist_tutorial_lbfgs.py
+    |   ├── mnist_tutorial_fgsm.py
+    |   ├── mnist_tutorial_bim.py
+    |   ├── mnist_tutorial_ilcm.py
+    |   ├── mnist_tutorial_mifgsm.py
+    |   ├── mnist_tutorial_jsma.py
+    |   └── mnist_tutorial_deepfool.py
+    └── README.md
+
+**advbox.attack**
+
+Advbox implements several popular adversarial attacks which search adversarial examples. Each attack method uses a distance measure(L1, L2, etc.) to quantify the size of adversarial perturbations. Advbox is easy to craft adversarial example as some attack methods could perform internal hyperparameter tuning to find the minimum perturbation.
+
+**advbox.model**
+
+Advbox implements interfaces to PaddlePaddle. Additionally, other deep learning framworks such as TensorFlow can also be defined and employed. The module is use to compute predictions and gradients for given inputs in a specific framework.
+
+**advbox.adversary**
+
+Adversary contains the original object, the target and the adversarial examples. It provides the misclassification as the criterion to accept a adversarial example.
+
+## Tutorials
+The `./tutorials/` folder provides some tutorials to generate adversarial examples on the MNIST dataset. You can slightly modify the code to apply to other dataset. These attack methods are supported in Advbox:
+
+* [L-BFGS](https://arxiv.org/abs/1312.6199)
+* [FGSM](https://arxiv.org/abs/1412.6572)
+* [BIM](https://arxiv.org/abs/1607.02533)
+* [ILCM](https://arxiv.org/abs/1607.02533)
+* [MI-FGSM](https://arxiv.org/pdf/1710.06081.pdf)
+* [JSMA](https://arxiv.org/pdf/1511.07528)
+* [DeepFool](https://arxiv.org/abs/1511.04599)
+
+## Testing
+Benchmarks on a vanilla CNN model.
+
+> MNIST
+
+|  adversarial attacks  |  fooling rate (non-targeted)  | fooling rate (targeted) | max_epsilon | iterations | Strength |
+|:-----:| :----: | :---: | :----: | :----: | :----: |
+|L-BFGS| --- | 89.2% | --- | One shot | *** |
+|FGSM| 57.8% | 26.55% | 0.3 | One shot| *** |
+|BIM| 97.4% | --- | 0.1 | 100 | **** |
+|ILCM| ---  | 100.0% | 0.1 | 100 | **** |
+|MI-FGSM| 94.4% | 100.0% | 0.1 | 100 | **** |
+|JSMA| 96.8% | 90.4%| 0.1 | 2000 | *** |
+|DeepFool| 97.7% | 51.3% | --- | 100 | **** |
+
+* The strength (higher for more asterisks) is based on the impression from the reviewed literature.
+
+---
+## References
+* [Intriguing properties of neural networks](https://arxiv.org/abs/1312.6199), C. Szegedy et al., arxiv 2014
+* [Explaining and Harnessing Adversarial Examples](https://arxiv.org/abs/1412.6572), I. Goodfellow et al., ICLR 2015
+* [Adversarial Examples In The Physical World](https://arxiv.org/pdf/1607.02533v3.pdf), A. Kurakin et al., ICLR workshop 2017
+* [Boosting Adversarial Attacks with Momentum](https://arxiv.org/abs/1710.06081), Yinpeng Dong et al., arxiv 2018
+* [The Limitations of Deep Learning in Adversarial Settings](https://arxiv.org/abs/1511.07528), N. Papernot et al., ESSP 2016
+* [DeepFool: a simple and accurate method to fool deep neural networks](https://arxiv.org/abs/1511.04599), S. Moosavi-Dezfooli et al., CVPR 2016
+* [Foolbox: A Python toolbox to benchmark the robustness of machine learning models](https://arxiv.org/abs/1707.04131), Jonas Rauber et al., arxiv 2018
+* [CleverHans: An adversarial example library for constructing attacks, building defenses, and benchmarking both](https://github.com/tensorflow/cleverhans#setting-up-cleverhans)
+* [Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey](https://arxiv.org/abs/1801.00553), Naveed Akhtar, Ajmal Mian, arxiv 2018

PaddleCV/caffe2fluid/README.md

+### Caffe2Fluid
+This tool is used to convert a Caffe model to a Fluid model
+
+### Key Features
+1. Convert caffe model to fluid model with codes of defining a network(useful for re-training)
+
+2. Pycaffe is not necessary when just want convert model without do caffe-inference
+
+3. Caffe's customized layers convertion also be supported by extending this tool
+
+4. A bunch of tools in `examples/imagenet/tools` are provided to compare the difference
+
+### HowTo
+1. Prepare `caffepb.py` in `./proto` if your python has no `pycaffe` module, two options provided here:
+    - Generate pycaffe from caffe.proto
+        ```
+        bash ./proto/compile.sh
+        ```
+
+    - Download one from github directly
+        ```
+        cd proto/ && wget https://raw.githubusercontent.com/ethereon/caffe-tensorflow/master/kaffe/caffe/caffepb.py
+        ```
+
+2. Convert the Caffe model to Fluid model
+   - Generate fluid code and weight file
+       ```
+       python convert.py alexnet.prototxt \
+               --caffemodel alexnet.caffemodel \
+               --data-output-path alexnet.npy \
+               --code-output-path alexnet.py
+       ```
+
+   - Save weights as fluid model file
+       ```
+       # only infer the last layer's result
+       python alexnet.py alexnet.npy ./fluid
+       # infer these 2 layer's result
+       python alexnet.py alexnet.npy ./fluid fc8,prob
+       ```
+
+3. Use the converted model to infer
+    - See more details in `examples/imagenet/tools/run.sh`
+
+4. Compare the inference results with caffe
+    - See more details in `examples/imagenet/tools/diff.sh`
+
+### How to convert custom layer
+1. Implement your custom layer in a file under `kaffe/custom_layers`, eg: mylayer.py
+    - Implement ```shape_func(input_shape, [other_caffe_params])``` to calculate the output shape
+    - Implement ```layer_func(inputs, name, [other_caffe_params])``` to construct a fluid layer
+    - Register these two functions ```register(kind='MyType', shape=shape_func, layer=layer_func)```
+    - Notes: more examples can be found in `kaffe/custom_layers`
+
+2. Add ```import mylayer``` to  `kaffe/custom_layers/\_\_init__.py`
+
+3. Prepare your pycaffe as your customized version(same as previous env prepare)
+    - (option1) replace `proto/caffe.proto` with your own caffe.proto and compile it
+    - (option2) change your `pycaffe` to the customized version
+
+4. Convert the Caffe model to Fluid model
+
+5. Set env $CAFFE2FLUID_CUSTOM_LAYERS to the parent directory of 'custom_layers'
+   ```
+   export CAFFE2FLUID_CUSTOM_LAYERS=/path/to/caffe2fluid/kaffe
+   ```
+
+6. Use the converted model when loading model in `xxxnet.py` and `xxxnet.npy`(no need if model is already in `fluid/model` and `fluid/params`)
+
+### Tested models
+- Lenet:
+[model addr](https://github.com/ethereon/caffe-tensorflow/blob/master/examples/mnist)
+
+- ResNets:(ResNet-50, ResNet-101, ResNet-152)
+[model addr](https://onedrive.live.com/?authkey=%21AAFW2-FVoxeVRck&id=4006CBB8476FF777%2117887&cid=4006CBB8476FF777)
+
+- GoogleNet:
+[model addr](https://gist.github.com/jimmie33/7ea9f8ac0da259866b854460f4526034)
+
+- VGG:
+[model addr](https://gist.github.com/ksimonyan/211839e770f7b538e2d8)
+
+- AlexNet:
+[model addr](https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet)
+
+### Notes
+Some of this code come from here: [caffe-tensorflow](https://github.com/ethereon/caffe-tensorflow)

PaddleCV/deeplabv3+/README.md

+DeepLab运行本目录下的程序示例需要使用PaddlePaddle Fluid v1.3.0版本或以上。如果您的PaddlePaddle安装版本低于此要求，请按照安装文档中的说明更新PaddlePaddle安装版本，如果使用GPU，该程序需要使用cuDNN v7版本。
+
+
+## 代码结构
+```
+├── models.py # 网络结构定义脚本
+├── train.py   # 训练任务脚本
+├── eval.py    # 评估脚本
+└── reader.py    # 定义通用的函数以及数据预处理脚本
+```
+
+## 简介
+
+DeepLabv3+ 是DeepLab语义分割系列网络的最新作，其前作有 DeepLabv1，DeepLabv2, DeepLabv3,
+在最新作中，DeepLab的作者通过encoder-decoder进行多尺度信息的融合，同时保留了原来的空洞卷积和ASSP层，
+其骨干网络使用了Xception模型，提高了语义分割的健壮性和运行速率，在 PASCAL VOC 2012 dataset取得新的state-of-art performance，89.0mIOU。
+
+![](./imgs/model.png)
+
+
+## 数据准备
+
+
+
+本文采用Cityscape数据集，请前往[Cityscape官网](https://www.cityscapes-dataset.com)注册下载。
+下载以后的数据目录结构如下
+```
+data/cityscape/
+|-- gtFine
+|   |-- test
+|   |-- train
+|   `-- val
+|-- leftImg8bit
+    |-- test
+    |-- train
+    `-- val
+```
+
+# 预训练模型准备
+
+我们为了节约更多的显存，在这里我们使用Group Norm作为我们的归一化手段。
+如果需要从头开始训练模型，用户需要下载我们的初始化模型
+```
+wget https://paddle-deeplab.bj.bcebos.com/deeplabv3plus_gn_init.tgz
+tar -xf deeplabv3plus_gn_init.tgz && rm deeplabv3plus_gn_init.tgz
+```
+如果需要最终训练模型进行fine tune或者直接用于预测，请下载我们的最终模型
+```
+wget https://paddle-deeplab.bj.bcebos.com/deeplabv3plus_gn.tgz
+tar -xf deeplabv3plus_gn.tgz && rm deeplabv3plus_gn.tgz
+```
+
+
+## 模型训练与预测
+
+### 训练
+执行以下命令进行训练，同时指定weights的保存路径，初始化路径，以及数据存放位置:
+```
+python ./train.py \
+    --batch_size=1 \
+    --train_crop_size=769 \
+    --total_step=50 \
+    --norm_type=gn \
+    --init_weights_path=$INIT_WEIGHTS_PATH \
+    --save_weights_path=$SAVE_WEIGHTS_PATH \
+    --dataset_path=$DATASET_PATH
+```
+使用以下命令获得更多使用说明：
+```
+python train.py --help
+```
+以上命令用于测试训练过程是否正常，仅仅迭代了50次并且使用了1的batch size，如果需要复现
+原论文的实验，请使用以下设置：
+```
+CUDA_VISIBLE_DEVICES=0 \
+python ./train.py \
+    --batch_size=4 \
+    --parallel=True \
+    --norm_type=gn \
+    --train_crop_size=769 \
+    --total_step=500000 \
+    --base_lr=0.001 \
+    --init_weights_path=deeplabv3plus_gn_init \
+    --save_weights_path=output \
+    --dataset_path=$DATASET_PATH
+```
+如果您的显存不足，可以尝试减小`batch_size`，同时等比例放大`total_step`, 缩小`base_lr`, 保证相乘的值不变，这得益于Group Norm的特性，改变 `batch_size` 并不会显著影响结果，而且能够节约更多显存, 比如您可以设置`--batch_size=2 --total_step=1000000 --base_lr=0.0005`。
+
+### 测试
+执行以下命令在`Cityscape`测试数据集上进行测试：
+```
+python ./eval.py \
+    --init_weights_path=deeplabv3plus_gn \
+    --norm_type=gn \
+    --dataset_path=$DATASET_PATH
+```
+需要通过选项`--init_weights_path`指定模型文件。测试脚本的输出的评估指标为mean IoU。
+
+
+## 实验结果
+训练完成以后，使用`eval.py`在验证集上进行测试，得到以下结果：
+```
+load from: ../models/deeplabv3plus_gn
+total number 500
+step: 500, mIoU: 0.7881
+```
+
+## 其他信息
+
+|数据集 | norm type | pretrained model | trained model | mean IoU
+|---|---|---|---|---|
+|CityScape | group norm | [deeplabv3plus_gn_init.tgz](https://paddle-deeplab.bj.bcebos.com/deeplabv3plus_gn_init.tgz) | [deeplabv3plus_gn.tgz](https://paddle-deeplab.bj.bcebos.com/deeplabv3plus_gn.tgz) | 0.7881 |
+
+## 参考
+
+- [Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation](https://arxiv.org/abs/1802.02611)