diff --git a/CMakeLists.txt b/CMakeLists.txt index 68447727118a91a2a8c0d06404353c7ccb734c6d..9cbc6e874d0b98b54abcc966b6c59dab9a69a566 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -204,6 +204,10 @@ include(external/snappy) # download snappy include(external/snappystream) include(external/threadpool) +include(cudnn) # set cudnn libraries, must before configure +include(cupti) +include(configure) # add paddle env configuration + if(WITH_GPU) include(cuda) include(tensorrt) @@ -212,9 +216,6 @@ elseif() set(WITH_ANAKIN OFF CACHE STRING "Anakin is used in GPU only now." FORCE) endif() -include(cudnn) # set cudnn libraries, must before configure -include(cupti) -include(configure) # add paddle env configuration include(generic) # simplify cmake module include(package) # set paddle packages include(ccache) # set ccache for compilation diff --git a/cmake/configure.cmake b/cmake/configure.cmake index 53454d79f7205415d0634c9350520eee7106b0b5..959b1381b7ac54cebe4891fb7a464aeec0199621 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -110,10 +110,11 @@ if(WITH_GPU) message(WARNING "Anakin needs CUDNN >= 7.0 to compile. Force WITH_ANAKIN=OFF") set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when CUDNN >= 7.0." FORCE) endif() - endif() - if(WITH_ANAKIN) - set(ENV{CUDNN_INCLUDE_DIR} ${CUDNN_INCLUDE_DIR}) - set(ENV{CUDNN_LIBRARY} ${CUDNN_LIBRARY}) + set(ENV{CUDNN_INCLUDE_DIR} "${CUDNN_INCLUDE_DIR}/") + get_filename_component(CUDNN_LIBRARY_DIR ${CUDNN_LIBRARY} DIRECTORY) + set(ENV{CUDNN_LIBRARY} ${CUDNN_LIBRARY_DIR}) + message(STATUS "cudnn include header is ${CUDNN_INCLUDE_DIR}/cudnn.h") + message(STATUS "cudnn library is ${CUDNN_LIBRARY_DIR}") endif() elseif(WITH_AMD_GPU) add_definitions(-DPADDLE_WITH_HIP) diff --git a/cmake/external/anakin.cmake b/cmake/external/anakin.cmake index 855897394a602736f3a4fa1132c634f8cb537468..5d11d238cd73653b60623bf7e32b94f5f1ac9ecc 100644 --- a/cmake/external/anakin.cmake +++ b/cmake/external/anakin.cmake @@ -19,17 +19,17 @@ execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-cer include_directories(${ANAKIN_INCLUDE}) include_directories(${ANAKIN_INCLUDE}/saber/) -set(ANAKIN_COMPILE_EXTRA_FLAGS +set(ANAKIN_COMPILE_EXTRA_FLAGS -Wno-error=unused-but-set-variable -Wno-unused-but-set-variable - -Wno-error=unused-variable -Wno-unused-variable + -Wno-error=unused-variable -Wno-unused-variable -Wno-error=format-extra-args -Wno-format-extra-args - -Wno-error=comment -Wno-comment - -Wno-error=format -Wno-format + -Wno-error=comment -Wno-comment + -Wno-error=format -Wno-format -Wno-error=switch -Wno-switch - -Wno-error=return-type -Wno-return-type + -Wno-error=return-type -Wno-return-type -Wno-error=non-virtual-dtor -Wno-non-virtual-dtor -Wno-sign-compare - -Wno-reorder + -Wno-reorder -Wno-error=cpp) ExternalProject_Add( @@ -47,6 +47,7 @@ ExternalProject_Add( -DPROTOBUF_ROOT=${THIRD_PARTY_PATH}/install/protobuf -DMKLML_ROOT=${THIRD_PARTY_PATH}/install/mklml -DCUDNN_ROOT=${CUDNN_ROOT} + -DCUDNN_INCLUDE_DIR=${CUDNN_INCLUDE_DIR} ${EXTERNAL_OPTIONAL_ARGS} CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ANAKIN_INSTALL_DIR} ) diff --git a/doc/fluid/new_docs/advanced_usage/deploy/index_anakin.rst b/doc/fluid/new_docs/advanced_usage/deploy/index_anakin.rst index b782242a6632a5d42a512cf3b830d6e047c064ab..e4682ccb94e6fc60e184632dff9ee16a6bf16ec0 100644 --- a/doc/fluid/new_docs/advanced_usage/deploy/index_anakin.rst +++ b/doc/fluid/new_docs/advanced_usage/deploy/index_anakin.rst @@ -1,5 +1,5 @@ -服务器端部署 - Anakin -##################### +Anakin - 服务器端加速引擎 +####################### 使用文档 diff --git a/doc/fluid/new_docs/advanced_usage/deploy/index_native.rst b/doc/fluid/new_docs/advanced_usage/deploy/index_native.rst deleted file mode 100644 index a5209e8560b31e9f0f776fba9a2b8c5bc150165c..0000000000000000000000000000000000000000 --- a/doc/fluid/new_docs/advanced_usage/deploy/index_native.rst +++ /dev/null @@ -1,8 +0,0 @@ -服务器端部署 - 原生引擎 -####################### - -.. toctree:: - :maxdepth: 2 - - build_and_install_lib_cn.rst - native_infer.rst diff --git a/doc/fluid/new_docs/advanced_usage/index.rst b/doc/fluid/new_docs/advanced_usage/index.rst index dea7c236619a0bdbf402f371571d947d1cdbba65..89166573eebca045e948046c69f3b7a3e0031d58 100644 --- a/doc/fluid/new_docs/advanced_usage/index.rst +++ b/doc/fluid/new_docs/advanced_usage/index.rst @@ -10,7 +10,6 @@ .. toctree:: :maxdepth: 2 - deploy/index_native.rst deploy/index_anakin.rst deploy/index_mobile.rst development/contribute_to_paddle.md diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/index.md b/doc/fluid/new_docs/beginners_guide/basics/image_classification/README.cn.md similarity index 69% rename from doc/fluid/new_docs/beginners_guide/basics/image_classification/index.md rename to doc/fluid/new_docs/beginners_guide/basics/image_classification/README.cn.md index ce0d2bb1dc0cf73151ee9aceea7e4d7b24af1926..88cc17475eaae946ef260dabd2e123276b55cdd2 100644 --- a/doc/fluid/new_docs/beginners_guide/basics/image_classification/index.md +++ b/doc/fluid/new_docs/beginners_guide/basics/image_classification/README.cn.md @@ -1,7 +1,7 @@ # 图像分类 -本教程源代码目录在[book/image_classification](https://github.com/PaddlePaddle/book/tree/develop/03.image_classification), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 +本教程源代码目录在[book/image_classification](https://github.com/PaddlePaddle/book/tree/develop/03.image_classification), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书),更多内容请参考本教程的[视频课堂](http://bit.baidu.com/course/detail/id/168.html)。 ## 背景介绍 @@ -20,24 +20,25 @@ 图像分类包括通用图像分类、细粒度图像分类等。图1展示了通用图像分类效果,即模型可以正确识别图像上的主要物体。 -![dogCatClassification](./image/dog_cat.png)

+
图1. 通用图像分类展示

图2展示了细粒度图像分类-花卉识别的效果,要求模型可以正确识别花的类别。 -![flowersClassification](./image/flowers.png) +

+
图2. 细粒度图像分类展示

一个好的模型既要对不同类别识别正确,同时也应该能够对不同视角、光照、背景、变形或部分遮挡的图像正确识别(这里我们统一称作图像扰动)。图3展示了一些图像的扰动,较好的模型会像聪明的人类一样能够正确识别。 -![imageVariations](https://raw.githubusercontent.com/PaddlePaddle/book/develop/03.image_classification/image/variations.png)

+
图3. 扰动图片展示[22]

@@ -46,17 +47,21 @@ 图像识别领域大量的研究成果都是建立在[PASCAL VOC](http://host.robots.ox.ac.uk/pascal/VOC/)、[ImageNet](http://image-net.org/)等公开的数据集上,很多图像识别算法通常在这些数据集上进行测试和比较。PASCAL VOC是2005年发起的一个视觉挑战赛,ImageNet是2010年发起的大规模视觉识别竞赛(ILSVRC)的数据集,在本章中我们基于这些竞赛的一些论文介绍图像分类模型。 在2012年之前的传统图像分类方法可以用背景描述中提到的三步完成,但通常完整建立图像识别模型一般包括底层特征学习、特征编码、空间约束、分类器设计、模型融合等几个阶段。 -1). **底层特征提取**: 通常从图像中按照固定步长、尺度提取大量局部特征描述。常用的局部特征包括SIFT(Scale-Invariant Feature Transform, 尺度不变特征转换) \[[1](#参考文献)\]、HOG(Histogram of Oriented Gradient, 方向梯度直方图) \[[2](#参考文献)\]、LBP(Local Bianray Pattern, 局部二值模式) \[[3](#参考文献)\] 等,一般也采用多种特征描述子,防止丢失过多的有用信息。 -2). **特征编码**: 底层特征中包含了大量冗余与噪声,为了提高特征表达的鲁棒性,需要使用一种特征变换算法对底层特征进行编码,称作特征编码。常用的特征编码包括向量量化编码 \[[4](#参考文献)\]、稀疏编码 \[[5](#参考文献)\]、局部线性约束编码 \[[6](#参考文献)\]、Fisher向量编码 \[[7](#参考文献)\] 等。 -3). **空间特征约束**: 特征编码之后一般会经过空间特征约束,也称作**特征汇聚**。特征汇聚是指在一个空间范围内,对每一维特征取最大值或者平均值,可以获得一定特征不变形的特征表达。金字塔特征匹配是一种常用的特征聚会方法,这种方法提出将图像均匀分块,在分块内做特征汇聚。 -4). **通过分类器分类**: 经过前面步骤之后一张图像可以用一个固定维度的向量进行描述,接下来就是经过分类器对图像进行分类。通常使用的分类器包括SVM(Support Vector Machine, 支持向量机)、随机森林等。而使用核方法的SVM是最为广泛的分类器,在传统图像分类任务上性能很好。 + + 1). **底层特征提取**: 通常从图像中按照固定步长、尺度提取大量局部特征描述。常用的局部特征包括SIFT(Scale-Invariant Feature Transform, 尺度不变特征转换) \[[1](#参考文献)\]、HOG(Histogram of Oriented Gradient, 方向梯度直方图) \[[2](#参考文献)\]、LBP(Local Bianray Pattern, 局部二值模式) \[[3](#参考文献)\] 等,一般也采用多种特征描述子,防止丢失过多的有用信息。 + + 2). **特征编码**: 底层特征中包含了大量冗余与噪声,为了提高特征表达的鲁棒性,需要使用一种特征变换算法对底层特征进行编码,称作特征编码。常用的特征编码包括向量量化编码 \[[4](#参考文献)\]、稀疏编码 \[[5](#参考文献)\]、局部线性约束编码 \[[6](#参考文献)\]、Fisher向量编码 \[[7](#参考文献)\] 等。 + + 3). **空间特征约束**: 特征编码之后一般会经过空间特征约束,也称作**特征汇聚**。特征汇聚是指在一个空间范围内,对每一维特征取最大值或者平均值,可以获得一定特征不变形的特征表达。金字塔特征匹配是一种常用的特征聚会方法,这种方法提出将图像均匀分块,在分块内做特征汇聚。 + + 4). **通过分类器分类**: 经过前面步骤之后一张图像可以用一个固定维度的向量进行描述,接下来就是经过分类器对图像进行分类。通常使用的分类器包括SVM(Support Vector Machine, 支持向量机)、随机森林等。而使用核方法的SVM是最为广泛的分类器,在传统图像分类任务上性能很好。 这种方法在PASCAL VOC竞赛中的图像分类算法中被广泛使用 \[[18](#参考文献)\]。[NEC实验室](http://www.nec-labs.com/)在ILSVRC2010中采用SIFT和LBP特征,两个非线性编码器以及SVM分类器获得图像分类的冠军 \[[8](#参考文献)\]。 Alex Krizhevsky在2012年ILSVRC提出的CNN模型 \[[9](#参考文献)\] 取得了历史性的突破,效果大幅度超越传统方法,获得了ILSVRC2012冠军,该模型被称作AlexNet。这也是首次将深度学习用于大规模图像分类中。从AlexNet之后,涌现了一系列CNN模型,不断地在ImageNet上刷新成绩,如图4展示。随着模型变得越来越深以及精妙的结构设计,Top-5的错误率也越来越低,降到了3.5%附近。而在同样的ImageNet数据集上,人眼的辨识错误率大概在5.1%,也就是目前的深度学习模型的识别能力已经超过了人眼。 -![ilsvrc](./image/ilsvrc.png)

+
图4. ILSVRC图像分类Top-5错误率

@@ -64,8 +69,8 @@ Alex Krizhevsky在2012年ILSVRC提出的CNN模型 \[[9](#参考文献)\] 取得 传统CNN包含卷积层、全连接层等组件,并采用softmax多类别分类器和多类交叉熵损失函数,一个典型的卷积神经网络如图5所示,我们先介绍用来构造CNN的常见组件。 -![cnnStructure](./image/lenet.png)

+
图5. CNN网络示例[20]

@@ -83,8 +88,8 @@ Alex Krizhevsky在2012年ILSVRC提出的CNN模型 \[[9](#参考文献)\] 取得 牛津大学VGG(Visual Geometry Group)组在2014年ILSVRC提出的模型被称作VGG模型 \[[11](#参考文献)\] 。该模型相比以往模型进一步加宽和加深了网络结构,它的核心是五组卷积操作,每两组之间做Max-Pooling空间降维。同一组内采用多次连续的3X3卷积,卷积核的数目由较浅组的64增多到最深组的512,同一组内的卷积核数目是一样的。卷积之后接两层全连接层,之后是分类层。由于每组内卷积层的不同,有11、13、16、19层这几种模型,下图展示一个16层的网络结构。VGG模型结构相对简洁,提出之后也有很多文章基于此模型进行研究,如在ImageNet上首次公开超过人眼识别的模型\[[19](#参考文献)\]就是借鉴VGG模型的结构。 -![vgg16](./image/vgg16.png)

+
图6. 基于ImageNet的VGG16模型

@@ -92,12 +97,16 @@ Alex Krizhevsky在2012年ILSVRC提出的CNN模型 \[[9](#参考文献)\] 取得 GoogleNet \[[12](#参考文献)\] 在2014年ILSVRC的获得了冠军,在介绍该模型之前我们先来了解NIN(Network in Network)模型 \[[13](#参考文献)\] 和Inception模块,因为GoogleNet模型由多组Inception模块组成,模型设计借鉴了NIN的一些思想。 -NIN模型主要有两个特点:1) 引入了多层感知卷积网络(Multi-Layer Perceptron Convolution, MLPconv)代替一层线性卷积网络。MLPconv是一个微小的多层卷积网络,即在线性卷积后面增加若干层1x1的卷积,这样可以提取出高度非线性特征。2) 传统的CNN最后几层一般都是全连接层,参数较多。而NIN模型设计最后一层卷积层包含类别维度大小的特征图,然后采用全局均值池化(Avg-Pooling)替代全连接层,得到类别维度大小的向量,再进行分类。这种替代全连接层的方式有利于减少参数。 +NIN模型主要有两个特点: + +1) 引入了多层感知卷积网络(Multi-Layer Perceptron Convolution, MLPconv)代替一层线性卷积网络。MLPconv是一个微小的多层卷积网络,即在线性卷积后面增加若干层1x1的卷积,这样可以提取出高度非线性特征。 + +2) 传统的CNN最后几层一般都是全连接层,参数较多。而NIN模型设计最后一层卷积层包含类别维度大小的特征图,然后采用全局均值池化(Avg-Pooling)替代全连接层,得到类别维度大小的向量,再进行分类。这种替代全连接层的方式有利于减少参数。 Inception模块如下图7所示,图(a)是最简单的设计,输出是3个卷积层和一个池化层的特征拼接。这种设计的缺点是池化层不会改变特征通道数,拼接后会导致特征的通道数较大,经过几层这样的模块堆积后,通道数会越来越大,导致参数和计算量也随之增大。为了改善这个缺点,图(b)引入3个1x1卷积层进行降维,所谓的降维就是减少通道数,同时如NIN模型中提到的1x1卷积也可以修正线性特征。 -![inception](./image/inception.png)

+
图7. Inception模块

@@ -105,8 +114,8 @@ GoogleNet由多组Inception模块堆积而成。另外,在网络最后也没 GoogleNet整体网络结构如图8所示,总共22层网络:开始由3层普通的卷积组成;接下来由三组子网络组成,第一组子网络包含2个Inception模块,第二组包含5个Inception模块,第三组包含2个Inception模块;然后接均值池化层、全连接层。 -![googleNet](./image/googlenet.jpeg)

+
图8. GoogleNet[12]

@@ -120,15 +129,15 @@ ResNet(Residual Network) \[[15](#参考文献)\] 是2015年ImageNet图像分类 残差模块如图9所示,左边是基本模块连接方式,由两个输出通道数相同的3x3卷积组成。右边是瓶颈模块(Bottleneck)连接方式,之所以称为瓶颈,是因为上面的1x1卷积用来降维(图示例即256->64),下面的1x1卷积用来升维(图示例即64->256),这样中间3x3卷积的输入和输出通道数都较小(图示例即64->64)。 -![ResNetBlock](./image/resnet_block.jpg)

+
图9. 残差模块

图10展示了50、101、152层网络连接示意图,使用的是瓶颈模块。这三个模型的区别在于每组中残差模块的重复次数不同(见图右上角)。ResNet训练收敛较快,成功的训练了上百乃至近千层的卷积神经网络。 -![ResNet](./image/resnet.png)

+
图10. 基于ImageNet的ResNet模型

@@ -139,8 +148,8 @@ ResNet(Residual Network) \[[15](#参考文献)\] 是2015年ImageNet图像分类 由于ImageNet数据集较大,下载和训练较慢,为了方便大家学习,我们使用[CIFAR10]()数据集。CIFAR10数据集包含60,000张32x32的彩色图片,10个类别,每个类包含6,000张。其中50,000张图片作为训练集,10000张作为测试集。图11从每个类别中随机抽取了10张图片,展示了所有的类别。 -![CIFAR](https://raw.githubusercontent.com/PaddlePaddle/book/develop/03.image_classification/image/cifar.png)

+
图11. CIFAR10数据集[21]

@@ -159,6 +168,7 @@ import paddle import paddle.fluid as fluid import numpy import sys +from __future__ import print_function ``` 本教程中我们提供了VGG和ResNet两个模型的配置。 @@ -170,33 +180,34 @@ VGG核心模块的输入是数据层,`vgg_bn_drop` 定义了16层VGG结构, ```python def vgg_bn_drop(input): -def conv_block(ipt, num_filter, groups, dropouts): -return fluid.nets.img_conv_group( -input=ipt, -pool_size=2, -pool_stride=2, -conv_num_filter=[num_filter] * groups, -conv_filter_size=3, -conv_act='relu', -conv_with_batchnorm=True, -conv_batchnorm_drop_rate=dropouts, -pool_type='max') - -conv1 = conv_block(input, 64, 2, [0.3, 0]) -conv2 = conv_block(conv1, 128, 2, [0.4, 0]) -conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0]) -conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0]) -conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0]) - -drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5) -fc1 = fluid.layers.fc(input=drop, size=512, act=None) -bn = fluid.layers.batch_norm(input=fc1, act='relu') -drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) -fc2 = fluid.layers.fc(input=drop2, size=512, act=None) -predict = fluid.layers.fc(input=fc2, size=10, act='softmax') -return predict + def conv_block(ipt, num_filter, groups, dropouts): + return fluid.nets.img_conv_group( + input=ipt, + pool_size=2, + pool_stride=2, + conv_num_filter=[num_filter] * groups, + conv_filter_size=3, + conv_act='relu', + conv_with_batchnorm=True, + conv_batchnorm_drop_rate=dropouts, + pool_type='max') + + conv1 = conv_block(input, 64, 2, [0.3, 0]) + conv2 = conv_block(conv1, 128, 2, [0.4, 0]) + conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0]) + conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0]) + conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0]) + + drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5) + fc1 = fluid.layers.fc(input=drop, size=512, act=None) + bn = fluid.layers.batch_norm(input=fc1, act='relu') + drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) + fc2 = fluid.layers.fc(input=drop2, size=512, act=None) + predict = fluid.layers.fc(input=fc2, size=10, act='softmax') + return predict ``` + 1. 首先定义了一组卷积网络,即conv_block。卷积核大小为3x3,池化窗口大小为2x2,窗口滑动大小为2,groups决定每组VGG模块是几次连续的卷积操作,dropouts指定Dropout操作的概率。所使用的`img_conv_group`是在`paddle.networks`中预定义的模块,由若干组 Conv->BN->ReLu->Dropout 和 一组 Pooling 组成。 2. 五组卷积操作,即 5个conv_block。 第一、二组采用两次连续的卷积操作。第三、四、五组采用三次连续的卷积操作。每组最后一个卷积后面Dropout概率为0,即不使用Dropout操作。 @@ -211,74 +222,76 @@ ResNet模型的第1、3、4步和VGG模型相同,这里不再介绍。主要 先介绍`resnet_cifar10`中的一些基本函数,再介绍网络连接过程。 -- `conv_bn_layer` : 带BN的卷积层。 -- `shortcut` : 残差模块的"直连"路径,"直连"实际分两种形式:残差模块输入和输出特征通道数不等时,采用1x1卷积的升维操作;残差模块输入和输出通道相等时,采用直连操作。 -- `basicblock` : 一个基础残差模块,即图9左边所示,由两组3x3卷积组成的路径和一条"直连"路径组成。 -- `bottleneck` : 一个瓶颈残差模块,即图9右边所示,由上下1x1卷积和中间3x3卷积组成的路径和一条"直连"路径组成。 -- `layer_warp` : 一组残差模块,由若干个残差模块堆积而成。每组中第一个残差模块滑动窗口大小与其他可以不同,以用来减少特征图在垂直和水平方向的大小。 + - `conv_bn_layer` : 带BN的卷积层。 + - `shortcut` : 残差模块的"直连"路径,"直连"实际分两种形式:残差模块输入和输出特征通道数不等时,采用1x1卷积的升维操作;残差模块输入和输出通道相等时,采用直连操作。 + - `basicblock` : 一个基础残差模块,即图9左边所示,由两组3x3卷积组成的路径和一条"直连"路径组成。 + - `bottleneck` : 一个瓶颈残差模块,即图9右边所示,由上下1x1卷积和中间3x3卷积组成的路径和一条"直连"路径组成。 + - `layer_warp` : 一组残差模块,由若干个残差模块堆积而成。每组中第一个残差模块滑动窗口大小与其他可以不同,以用来减少特征图在垂直和水平方向的大小。 ```python def conv_bn_layer(input, -ch_out, -filter_size, -stride, -padding, -act='relu', -bias_attr=False): -tmp = fluid.layers.conv2d( -input=input, -filter_size=filter_size, -num_filters=ch_out, -stride=stride, -padding=padding, -act=None, -bias_attr=bias_attr) -return fluid.layers.batch_norm(input=tmp, act=act) + ch_out, + filter_size, + stride, + padding, + act='relu', + bias_attr=False): + tmp = fluid.layers.conv2d( + input=input, + filter_size=filter_size, + num_filters=ch_out, + stride=stride, + padding=padding, + act=None, + bias_attr=bias_attr) + return fluid.layers.batch_norm(input=tmp, act=act) def shortcut(input, ch_in, ch_out, stride): -if ch_in != ch_out: -return conv_bn_layer(input, ch_out, 1, stride, 0, None) -else: -return input + if ch_in != ch_out: + return conv_bn_layer(input, ch_out, 1, stride, 0, None) + else: + return input def basicblock(input, ch_in, ch_out, stride): -tmp = conv_bn_layer(input, ch_out, 3, stride, 1) -tmp = conv_bn_layer(tmp, ch_out, 3, 1, 1, act=None, bias_attr=True) -short = shortcut(input, ch_in, ch_out, stride) -return fluid.layers.elementwise_add(x=tmp, y=short, act='relu') + tmp = conv_bn_layer(input, ch_out, 3, stride, 1) + tmp = conv_bn_layer(tmp, ch_out, 3, 1, 1, act=None, bias_attr=True) + short = shortcut(input, ch_in, ch_out, stride) + return fluid.layers.elementwise_add(x=tmp, y=short, act='relu') def layer_warp(block_func, input, ch_in, ch_out, count, stride): -tmp = block_func(input, ch_in, ch_out, stride) -for i in range(1, count): -tmp = block_func(tmp, ch_out, ch_out, 1) -return tmp + tmp = block_func(input, ch_in, ch_out, stride) + for i in range(1, count): + tmp = block_func(tmp, ch_out, ch_out, 1) + return tmp ``` `resnet_cifar10` 的连接结构主要有以下几个过程。 1. 底层输入连接一层 `conv_bn_layer`,即带BN的卷积层。 + 2. 然后连接3组残差模块即下面配置3组 `layer_warp` ,每组采用图 10 左边残差模块组成。 + 3. 最后对网络做均值池化并返回该层。 -注意:除过第一层卷积层和最后一层全连接层之外,要求三组 `layer_warp` 总的含参层数能够被6整除,即 `resnet_cifar10` 的 depth 要满足 `$(depth - 2) % 6 == 0$` 。 +注意:除过第一层卷积层和最后一层全连接层之外,要求三组 `layer_warp` 总的含参层数能够被6整除,即 `resnet_cifar10` 的 depth 要满足 $(depth - 2) % 6 == 0$ 。 ```python def resnet_cifar10(ipt, depth=32): -# depth should be one of 20, 32, 44, 56, 110, 1202 -assert (depth - 2) % 6 == 0 -n = (depth - 2) / 6 -nStages = {16, 64, 128} -conv1 = conv_bn_layer(ipt, ch_out=16, filter_size=3, stride=1, padding=1) -res1 = layer_warp(basicblock, conv1, 16, 16, n, 1) -res2 = layer_warp(basicblock, res1, 16, 32, n, 2) -res3 = layer_warp(basicblock, res2, 32, 64, n, 2) -pool = fluid.layers.pool2d( -input=res3, pool_size=8, pool_type='avg', pool_stride=1) -predict = fluid.layers.fc(input=pool, size=10, act='softmax') -return predict + # depth should be one of 20, 32, 44, 56, 110, 1202 + assert (depth - 2) % 6 == 0 + n = (depth - 2) / 6 + nStages = {16, 64, 128} + conv1 = conv_bn_layer(ipt, ch_out=16, filter_size=3, stride=1, padding=1) + res1 = layer_warp(basicblock, conv1, 16, 16, n, 1) + res2 = layer_warp(basicblock, res1, 16, 32, n, 2) + res3 = layer_warp(basicblock, res2, 32, 64, n, 2) + pool = fluid.layers.pool2d( + input=res3, pool_size=8, pool_type='avg', pool_stride=1) + predict = fluid.layers.fc(input=pool, size=10, act='softmax') + return predict ``` ## Infererence Program 配置 @@ -287,13 +300,13 @@ return predict ```python def inference_program(): -# The image is 32 * 32 with RGB representation. -data_shape = [3, 32, 32] -images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') + # The image is 32 * 32 with RGB representation. + data_shape = [3, 32, 32] + images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') -predict = resnet_cifar10(images, 32) -# predict = vgg_bn_drop(images) # un-comment to use vgg net -return predict + predict = resnet_cifar10(images, 32) + # predict = vgg_bn_drop(images) # un-comment to use vgg net + return predict ``` ## Train Program 配置 @@ -306,13 +319,13 @@ return predict ```python def train_program(): -predict = inference_program() + predict = inference_program() -label = fluid.layers.data(name='label', shape=[1], dtype='int64') -cost = fluid.layers.cross_entropy(input=predict, label=label) -avg_cost = fluid.layers.mean(cost) -accuracy = fluid.layers.accuracy(input=predict, label=label) -return [avg_cost, accuracy] + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(cost) + accuracy = fluid.layers.accuracy(input=predict, label=label) + return [avg_cost, accuracy] ``` ## Optimizer Function 配置 @@ -321,7 +334,7 @@ return [avg_cost, accuracy] ```python def optimizer_program(): -return fluid.optimizer.Adam(learning_rate=0.001) + return fluid.optimizer.Adam(learning_rate=0.001) ``` ## 训练模型 @@ -334,9 +347,9 @@ return fluid.optimizer.Adam(learning_rate=0.001) use_cuda = False place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( -train_func=train_program, -optimizer_func=optimizer_program, -place=place) + train_func=train_program, + optimizer_func=optimizer_program, + place=place) ``` ### Data Feeders 配置 @@ -349,12 +362,12 @@ BATCH_SIZE = 128 # Reader for training train_reader = paddle.batch( -paddle.reader.shuffle(paddle.dataset.cifar.train10(), buf_size=50000), -batch_size=BATCH_SIZE) + paddle.reader.shuffle(paddle.dataset.cifar.train10(), buf_size=50000), + batch_size=BATCH_SIZE) # Reader for testing. A separated data set for testing. test_reader = paddle.batch( -paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) + paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) ``` ### Event Handler @@ -363,7 +376,11 @@ paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) `event_handler_plot`可以用来利用回调数据来打点画图: -![png](./image/train_and_test.png) +

+
+图12. 训练结果 +

+ ```python params_dirname = "image_classification_resnet.inference.model" @@ -376,21 +393,21 @@ cost_ploter = Ploter(train_title, test_title) step = 0 def event_handler_plot(event): -global step -if isinstance(event, fluid.EndStepEvent): -if step % 1 == 0: -cost_ploter.append(train_title, step, event.metrics[0]) -cost_ploter.plot() -step += 1 -if isinstance(event, fluid.EndEpochEvent): -avg_cost, accuracy = trainer.test( -reader=test_reader, -feed_order=['pixel', 'label']) -cost_ploter.append(test_title, step, avg_cost) - -# save parameters -if params_dirname is not None: -trainer.save_params(params_dirname) + global step + if isinstance(event, fluid.EndStepEvent): + if step % 1 == 0: + cost_ploter.append(train_title, step, event.metrics[0]) + cost_ploter.plot() + step += 1 + if isinstance(event, fluid.EndEpochEvent): + avg_cost, accuracy = trainer.test( + reader=test_reader, + feed_order=['pixel', 'label']) + cost_ploter.append(test_title, step, avg_cost) + + # save parameters + if params_dirname is not None: + trainer.save_params(params_dirname) ``` `event_handler` 用来在训练过程中输出文本日志 @@ -400,39 +417,39 @@ params_dirname = "image_classification_resnet.inference.model" # event handler to track training and testing process def event_handler(event): -if isinstance(event, fluid.EndStepEvent): -if event.step % 100 == 0: -print("\nPass %d, Batch %d, Cost %f, Acc %f" % -(event.step, event.epoch, event.metrics[0], -event.metrics[1])) -else: -sys.stdout.write('.') -sys.stdout.flush() - -if isinstance(event, fluid.EndEpochEvent): -# Test against with the test dataset to get accuracy. -avg_cost, accuracy = trainer.test( -reader=test_reader, feed_order=['pixel', 'label']) - -print('\nTest with Pass {0}, Loss {1:2.2}, Acc {2:2.2}'.format(event.epoch, avg_cost, accuracy)) - -# save parameters -if params_dirname is not None: -trainer.save_params(params_dirname) + if isinstance(event, fluid.EndStepEvent): + if event.step % 100 == 0: + print("\nPass %d, Batch %d, Cost %f, Acc %f" % + (event.step, event.epoch, event.metrics[0], + event.metrics[1])) + else: + sys.stdout.write('.') + sys.stdout.flush() + + if isinstance(event, fluid.EndEpochEvent): + # Test against with the test dataset to get accuracy. + avg_cost, accuracy = trainer.test( + reader=test_reader, feed_order=['pixel', 'label']) + + print('\nTest with Pass {0}, Loss {1:2.2}, Acc {2:2.2}'.format(event.epoch, avg_cost, accuracy)) + + # save parameters + if params_dirname is not None: + trainer.save_params(params_dirname) ``` ### 训练 通过`trainer.train`函数训练: -**注意:** CPU,每个 Epoch 将花费大约15~20分钟。这部分可能需要一段时间。请随意修改代码,在GPU上运行测试,以提高培训速度。 +**注意:** CPU,每个 Epoch 将花费大约15~20分钟。这部分可能需要一段时间。请随意修改代码,在GPU上运行测试,以提高训练速度。 ```python trainer.train( -reader=train_reader, -num_epochs=2, -event_handler=event_handler, -feed_order=['pixel', 'label']) + reader=train_reader, + num_epochs=2, + event_handler=event_handler, + feed_order=['pixel', 'label']) ``` 一轮训练log示例如下所示,经过1个pass, 训练集上平均 Accuracy 为0.59 ,测试集上平均 Accuracy 为0.6 。 @@ -449,11 +466,11 @@ Pass 300, Batch 0, Cost 1.223424, Acc 0.593750 Test with Pass 0, Loss 1.1, Acc 0.6 ``` -图12是训练的分类错误率曲线图,运行到第200个pass后基本收敛,最终得到测试集上分类错误率为8.54%。 +图13是训练的分类错误率曲线图,运行到第200个pass后基本收敛,最终得到测试集上分类错误率为8.54%。 -![CIFARErrorRate](./image/plot.png)

-图12. CIFAR10数据集上VGG模型的分类错误率 +
+图13. CIFAR10数据集上VGG模型的分类错误率

## 应用模型 @@ -471,19 +488,19 @@ import numpy as np import os def load_image(file): -im = Image.open(file) -im = im.resize((32, 32), Image.ANTIALIAS) + im = Image.open(file) + im = im.resize((32, 32), Image.ANTIALIAS) -im = np.array(im).astype(np.float32) -# The storage order of the loaded image is W(width), -# H(height), C(channel). PaddlePaddle requires -# the CHW order, so transpose them. -im = im.transpose((2, 0, 1)) # CHW -im = im / 255.0 + im = np.array(im).astype(np.float32) + # The storage order of the loaded image is W(width), + # H(height), C(channel). PaddlePaddle requires + # the CHW order, so transpose them. + im = im.transpose((2, 0, 1)) # CHW + im = im / 255.0 -# Add one dimension to mimic the list format. -im = numpy.expand_dims(im, axis=0) -return im + # Add one dimension to mimic the list format. + im = numpy.expand_dims(im, axis=0) + return im cur_dir = os.getcwd() img = load_image(cur_dir + '/image/dog.png') @@ -497,11 +514,11 @@ img = load_image(cur_dir + '/image/dog.png') ```python inferencer = fluid.Inferencer( -infer_func=inference_program, param_path=params_dirname, place=place) - + infer_func=inference_program, param_path=params_dirname, place=place) +label_list = ["airplane", "automobile", "bird", "cat", "deer", "dog", "frog", "horse", "ship", "truck"] # inference results = inferencer.infer({'pixel': img}) -print("infer results: ", results) +print("infer results: %s" % label_list[np.argmax(results[0])]) ``` ## 总结 diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog.png deleted file mode 100644 index ca8f858a902ea723d886d2b88c2c0a1005301c50..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog_cat.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog_cat.png deleted file mode 100644 index 38b21f21604b1bb84fc3f6aa96bd5fce45d15a55..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog_cat.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/fea_conv0.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/fea_conv0.png deleted file mode 100644 index 647c822e52cd55d50e5f207978f5e6ada86cf34c..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/fea_conv0.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/flowers.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/flowers.png deleted file mode 100644 index 04245cef60fe7126ae4c92ba8085273965078bee..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/flowers.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/googlenet.jpeg b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/googlenet.jpeg deleted file mode 100644 index 249dbf96df61c3352ea5bd80470f6c4a1e03ff10..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/googlenet.jpeg and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/ilsvrc.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/ilsvrc.png deleted file mode 100644 index 4660ac122e9d533023a21154d35eee29e3b08d27..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/ilsvrc.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/inception.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/inception.png deleted file mode 100644 index 9591a0c1e8c0165c40ca560be35a7b9a91cd5027..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/inception.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/lenet.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/lenet.png deleted file mode 100644 index 77f785e03bacd38c4c64a817874a58ff3298d2f3..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/lenet.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/plot.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/plot.png deleted file mode 100644 index 57e45cc0c27dd99b9918de2ff1228bc6b65f7424..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/plot.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet.png deleted file mode 100644 index 0aeb4f254639fdbf18e916dc219ca61602596d85..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet_block.jpg b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet_block.jpg deleted file mode 100644 index c500eb01a90190ff66150871fe83ec275e2de8d7..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet_block.jpg and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/train_and_test.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/train_and_test.png deleted file mode 100644 index c6336a9a69b95dc978719ce68896e3e752e67fed..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/train_and_test.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/vgg16.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/vgg16.png deleted file mode 100644 index 6270eefcfd7071bc1643ee06567e5b81aaf4c177..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/vgg16.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/index.rst b/doc/fluid/new_docs/beginners_guide/basics/index.rst index d16f8b947253a535567ddc8d7b227dd153d9b154..0fcb008e0a7773e81e5124da09fe07366130b924 100644 --- a/doc/fluid/new_docs/beginners_guide/basics/index.rst +++ b/doc/fluid/new_docs/beginners_guide/basics/index.rst @@ -10,9 +10,9 @@ .. toctree:: :maxdepth: 2 - image_classification/index.md - word2vec/index.md - recommender_system/index.md - understand_sentiment/index.md - label_semantic_roles/index.md - machine_translation/index.md + image_classification/README.cn.md + word2vec/README.cn.md + recommender_system/README.cn.md + understand_sentiment/README.cn.md + label_semantic_roles/README.cn.md + machine_translation/README.cn.md diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/index.md b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/README.cn.md similarity index 54% rename from doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/index.md rename to doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/README.cn.md index 828ca738317992270487647e66b08b6d2f80e209..545f6002f23c1c698bd0d77fb39a8797ff7f5bde 100644 --- a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/index.md +++ b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/README.cn.md @@ -1,6 +1,6 @@ # 语义角色标注 -本教程源代码目录在[book/label_semantic_roles](https://github.com/PaddlePaddle/book/tree/develop/07.label_semantic_roles), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 +本教程源代码目录在[book/label_semantic_roles](https://github.com/PaddlePaddle/book/tree/develop/07.label_semantic_roles), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书),更多内容请参考本教程的[视频课堂](http://bit.baidu.com/course/detail/id/178.html)。 ## 背景介绍 @@ -8,7 +8,7 @@ 请看下面的例子,“遇到” 是谓词(Predicate,通常简写为“Pred”),“小明”是施事者(Agent),“小红”是受事者(Patient),“昨天” 是事件发生的时间(Time),“公园”是事情发生的地点(Location)。 -$$\mbox{[小明]}_{\mbox{Agent}}\mbox{[昨天]}_{\mbox{Time}}\mbox{[晚上]}_{\mbox{Time}}\mbox{在[公园]}_{\mbox{Location}}\mbox{[遇到]}_{\mbox{Predicate}}\mbox{了[小红]}_{\mbox{Patient}}\mbox{。}$$ +$$\mbox{[小明]}_{\mbox{Agent}}\mbox{[昨天]}_{\mbox{Time}}\mbox{[晚上]}_\mbox{Time}\mbox{在[公园]}_{\mbox{Location}}\mbox{[遇到]}_{\mbox{Predicate}}\mbox{了[小红]}_{\mbox{Patient}}\mbox{。}$$ 语义角色标注(Semantic Role Labeling,SRL)以句子的谓词为中心,不对句子所包含的语义信息进行深入分析,只分析句子中各成分与谓词之间的关系,即句子的谓词(Predicate)- 论元(Argument)结构,并用语义角色来描述这些结构关系,是许多自然语言理解任务(如信息抽取,篇章分析,深度问答等)的一个重要中间步骤。在研究中一般都假定谓词是给定的,所要做的就是找出给定谓词的各个论元和它们的语义角色。 @@ -20,17 +20,17 @@ $$\mbox{[小明]}_{\mbox{Agent}}\mbox{[昨天]}_{\mbox{Time}}\mbox{[晚上]}_{\m 4. 论元识别:这个过程是从上一步剪除之后的候选中判断哪些是真正的论元,通常当做一个二分类问题来解决。 5. 对第4步的结果,通过多分类得到论元的语义角色标签。可以看到,句法分析是基础,并且后续步骤常常会构造的一些人工特征,这些特征往往也来自句法分析。 -![dependencyParsing](./image/dependency_parsing.png)
+
图1. 依存句法分析句法树示例
-然而,完全句法分析需要确定句子所包含的全部句法信息,并确定句子各成分之间的关系,是一个非常困难的任务,目前技术下的句法分析准确率并不高,句法分析的细微错误都会导致SRL的错误。为了降低问题的复杂度,同时获得一定的句法结构信息,“浅层句法分析”的思想应运而生。浅层句法分析也称为部分句法分析(partial parsing)或语块划分(chunking)。和完全句法分析得到一颗完整的句法树不同,浅层句法分析只需要识别句子中某些结构相对简单的独立成分,例如:动词短语,这些被识别出来的结构称为语块。为了回避 “无法获得准确率较高的句法树” 所带来的困难,一些研究\[[1](#参考文献)\]也提出了基于语块(chunk)的SRL方法。基于语块的SRL方法将SRL作为一个序列标注问题来解决。序列标注任务一般都会采用BIO表示方式来定义序列标注的标签集,我们先来介绍这种表示方法。在BIO表示法中,B代表语块的开始,I代表语块的中间,O代表语块结束。通过B、I、O 三种标记将不同的语块赋予不同的标签,例如:对于一个角色为A的论元,将它所包含的第一个语块赋予标签B-A,将它所包含的其它语块赋予标签I-A,不属于任何论元的语块赋予标签O。 +然而,完全句法分析需要确定句子所包含的全部句法信息,并确定句子各成分之间的关系,是一个非常困难的任务,目前技术下的句法分析准确率并不高,句法分析的细微错误都会导致SRL的错误。为了降低问题的复杂度,同时获得一定的句法结构信息,“浅层句法分析”的思想应运而生。浅层句法分析也称为部分句法分析(partial parsing)或语块划分(chunking)。和完全句法分析得到一颗完整的句法树不同,浅层句法分析只需要识别句子中某些结构相对简单的独立成分,例如:动词短语,这些被识别出来的结构称为语块。为了回避 “无法获得准确率较高的句法树” 所带来的困难,一些研究\[[1](#参考文献)\]也提出了基于语块(chunk)的SRL方法。基于语块的SRL方法将SRL作为一个序列标注问题来解决。序列标注任务一般都会采用BIO表示方式来定义序列标注的标签集,我们先来介绍这种表示方法。在BIO表示法中,B代表语块的开始,I代表语块的中间,O代表语块结束。通过B、I、O 三种标记将不同的语块赋予不同的标签,例如:对于一个由角色A拓展得到的语块组,将它所包含的第一个语块赋予标签B-A,将它所包含的其它语块赋予标签I-A,不属于任何论元的语块赋予标签O。 我们继续以上面的这句话为例,图1展示了BIO表示方法。 -![bioExample](./image/bio_example.png)
+
图2. BIO标注方法示例
@@ -44,27 +44,27 @@ $$\mbox{[小明]}_{\mbox{Agent}}\mbox{[昨天]}_{\mbox{Time}}\mbox{[晚上]}_{\m ### 栈式循环神经网络(Stacked Recurrent Neural Network) -深层网络有助于形成层次化特征,网络上层在下层已经学习到的初级特征基础上,形成更复杂的高级特征。尽管LSTM沿时间轴展开后等价于一个非常“深”的前馈网络,但由于LSTM各个时间步参数共享,`$t-1$`时刻状态到`$t$`时刻的映射,始终只经过了一次非线性映射,也就是说单层LSTM对状态转移的建模是 “浅” 的。堆叠多个LSTM单元,令前一个LSTM`$t$`时刻的输出,成为下一个LSTM单元`$t$`时刻的输入,帮助我们构建起一个深层网络,我们把它称为第一个版本的栈式循环神经网络。深层网络提高了模型拟合复杂模式的能力,能够更好地建模跨不同时间步的模式\[[2](#参考文献)\]。 +深层网络有助于形成层次化特征,网络上层在下层已经学习到的初级特征基础上,形成更复杂的高级特征。尽管LSTM沿时间轴展开后等价于一个非常“深”的前馈网络,但由于LSTM各个时间步参数共享,$t-1$时刻状态到$t$时刻的映射,始终只经过了一次非线性映射,也就是说单层LSTM对状态转移的建模是 “浅” 的。堆叠多个LSTM单元,令前一个LSTM$t$时刻的输出,成为下一个LSTM单元$t$时刻的输入,帮助我们构建起一个深层网络,我们把它称为第一个版本的栈式循环神经网络。深层网络提高了模型拟合复杂模式的能力,能够更好地建模跨不同时间步的模式\[[2](#参考文献)\]。 然而,训练一个深层LSTM网络并非易事。纵向堆叠多个LSTM单元可能遇到梯度在纵向深度上传播受阻的问题。通常,堆叠4层LSTM单元可以正常训练,当层数达到4~8层时,会出现性能衰减,这时必须考虑一些新的结构以保证梯度纵向顺畅传播,这是训练深层LSTM网络必须解决的问题。我们可以借鉴LSTM解决 “梯度消失梯度爆炸” 问题的智慧之一:在记忆单元(Memory Cell)这条信息传播的路线上没有非线性映射,当梯度反向传播时既不会衰减、也不会爆炸。因此,深层LSTM模型也可以在纵向上添加一条保证梯度顺畅传播的路径。 -一个LSTM单元完成的运算可以被分为三部分:(1)输入到隐层的映射(input-to-hidden) :每个时间步输入信息`$x$`会首先经过一个矩阵映射,再作为遗忘门,输入门,记忆单元,输出门的输入,注意,这一次映射没有引入非线性激活;(2)隐层到隐层的映射(hidden-to-hidden):这一步是LSTM计算的主体,包括遗忘门,输入门,记忆单元更新,输出门的计算;(3)隐层到输出的映射(hidden-to-output):通常是简单的对隐层向量进行激活。我们在第一个版本的栈式网络的基础上,加入一条新的路径:除上一层LSTM输出之外,将前层LSTM的输入到隐层的映射作为的一个新的输入,同时加入一个线性映射去学习一个新的变换。 +一个LSTM单元完成的运算可以被分为三部分:(1)输入到隐层的映射(input-to-hidden) :每个时间步输入信息$x$会首先经过一个矩阵映射,再作为遗忘门,输入门,记忆单元,输出门的输入,注意,这一次映射没有引入非线性激活;(2)隐层到隐层的映射(hidden-to-hidden):这一步是LSTM计算的主体,包括遗忘门,输入门,记忆单元更新,输出门的计算;(3)隐层到输出的映射(hidden-to-output):通常是简单的对隐层向量进行激活。我们在第一个版本的栈式网络的基础上,加入一条新的路径:除上一层LSTM输出之外,将前层LSTM的输入到隐层的映射作为的一个新的输入,同时加入一个线性映射去学习一个新的变换。 图3是最终得到的栈式循环神经网络结构示意图。 -![lstmStructure](./image/stacked_lstm.png) -

+

+
图3. 基于LSTM的栈式循环神经网络结构示意图

### 双向循环神经网络(Bidirectional Recurrent Neural Network) -在LSTM中,`$t$`时刻的隐藏层向量编码了到`$t$`时刻为止所有输入的信息,但`$t$`时刻的LSTM可以看到历史,却无法看到未来。在绝大多数自然语言处理任务中,我们几乎总是能拿到整个句子。这种情况下,如果能够像获取历史信息一样,得到未来的信息,对序列学习任务会有很大的帮助。 +在LSTM中,$t$时刻的隐藏层向量编码了到$t$时刻为止所有输入的信息,但$t$时刻的LSTM可以看到历史,却无法看到未来。在绝大多数自然语言处理任务中,我们几乎总是能拿到整个句子。这种情况下,如果能够像获取历史信息一样,得到未来的信息,对序列学习任务会有很大的帮助。 -为了克服这一缺陷,我们可以设计一种双向循环网络单元,它的思想简单且直接:对上一节的栈式循环神经网络进行一个小小的修改,堆叠多个LSTM单元,让每一层LSTM单元分别以:正向、反向、正向 …… 的顺序学习上一层的输出序列。于是,从第2层开始,`$t$`时刻我们的LSTM单元便总是可以看到历史和未来的信息。图4是基于LSTM的双向循环神经网络结构示意图。 +为了克服这一缺陷,我们可以设计一种双向循环网络单元,它的思想简单且直接:对上一节的栈式循环神经网络进行一个小小的修改,堆叠多个LSTM单元,让每一层LSTM单元分别以:正向、反向、正向 …… 的顺序学习上一层的输出序列。于是,从第2层开始,$t$时刻我们的LSTM单元便总是可以看到历史和未来的信息。图4是基于LSTM的双向循环神经网络结构示意图。 -![lstmStructure](./image/bidirectional_stacked_lstm.png) -

+

+
图4. 基于LSTM的双向循环神经网络结构示意图

@@ -74,56 +74,56 @@ $$\mbox{[小明]}_{\mbox{Agent}}\mbox{[昨天]}_{\mbox{Time}}\mbox{[晚上]}_{\m 使用神经网络模型解决问题的思路通常是:前层网络学习输入的特征表示,网络的最后一层在特征基础上完成最终的任务。在SRL任务中,深层LSTM网络学习输入的特征表示,条件随机场(Conditional Random Filed, CRF)在特征的基础上完成序列标注,处于整个网络的末端。 -CRF是一种概率化结构模型,可以看作是一个概率无向图模型,结点表示随机变量,边表示随机变量之间的概率依赖关系。简单来讲,CRF学习条件概率`$P(X|Y)$`,其中 `$X = (x_1, x_2, ... , x_n)$` 是输入序列,`$Y = (y_1, y_2, ... , y_n)$` 是标记序列;解码过程是给定 `$X$`序列求解令`$P(Y|X)$`最大的`$Y$`序列,即`$Y^* = \mbox{arg max}_{Y} P(Y | X)$`。 +CRF是一种概率化结构模型,可以看作是一个概率无向图模型,结点表示随机变量,边表示随机变量之间的概率依赖关系。简单来讲,CRF学习条件概率$P(X|Y)$,其中 $X = (x_1, x_2, ... , x_n)$ 是输入序列,$Y = (y_1, y_2, ... , y_n)$ 是标记序列;解码过程是给定 $X$序列求解令$P(Y|X)$最大的$Y$序列,即$Y^* = \mbox{arg max}_{Y} P(Y | X)$。 序列标注任务只需要考虑输入和输出都是一个线性序列,并且由于我们只是将输入序列作为条件,不做任何条件独立假设,因此输入序列的元素之间并不存在图结构。综上,在序列标注任务中使用的是如图5所示的定义在链式图上的CRF,称之为线性链条件随机场(Linear Chain Conditional Random Field)。 -![linear_chain_crf](./image/linear_chain_crf.png) -

+

+
图5. 序列标注任务中使用的线性链条件随机场

-根据线性链条件随机场上的因子分解定理\[[5](#参考文献)\],在给定观测序列`$X$`时,一个特定标记序列`$Y$`的概率可以定义为: +根据线性链条件随机场上的因子分解定理\[[5](#参考文献)\],在给定观测序列$X$时,一个特定标记序列$Y$的概率可以定义为: $$p(Y | X) = \frac{1}{Z(X)} \text{exp}\left(\sum_{i=1}^{n}\left(\sum_{j}\lambda_{j}t_{j} (y_{i - 1}, y_{i}, X, i) + \sum_{k} \mu_k s_k (y_i, X, i)\right)\right)$$ -其中`$Z(X)$`是归一化因子,`$t_j$` 是定义在边上的特征函数,依赖于当前和前一个位置,称为转移特征,表示对于输入序列`$X$`及其标注序列在 `$i$`及`$i - 1$`位置上标记的转移概率。`$s_k$`是定义在结点上的特征函数,称为状态特征,依赖于当前位置,表示对于观察序列`$X$`及其`$i$`位置的标记概率。`$\lambda_j$` 和 `$\mu_k$` 分别是转移特征函数和状态特征函数对应的权值。实际上,`$t$`和`$s$`可以用相同的数学形式表示,再对转移特征和状态特在各个位置`$i$`求和有:`$f_{k}(Y, X) = \sum_{i=1}^{n}f_k({y_{i - 1}, y_i, X, i})$`,把`$f$`统称为特征函数,于是`$P(Y|X)$`可表示为: +其中$Z(X)$是归一化因子,$t_j$ 是定义在边上的特征函数,依赖于当前和前一个位置,称为转移特征,表示对于输入序列$X$及其标注序列在 $i$及$i - 1$位置上标记的转移概率。$s_k$是定义在结点上的特征函数,称为状态特征,依赖于当前位置,表示对于观察序列$X$及其$i$位置的标记概率。$\lambda_j$ 和 $\mu_k$ 分别是转移特征函数和状态特征函数对应的权值。实际上,$t$和$s$可以用相同的数学形式表示,再对转移特征和状态特在各个位置$i$求和有:$f_{k}(Y, X) = \sum_{i=1}^{n}f_k({y_{i - 1}, y_i, X, i})$,把$f$统称为特征函数,于是$P(Y|X)$可表示为: $$p(Y|X, W) = \frac{1}{Z(X)}\text{exp}\sum_{k}\omega_{k}f_{k}(Y, X)$$ -`$\omega$`是特征函数对应的权值,是CRF模型要学习的参数。训练时,对于给定的输入序列和对应的标记序列集合`$D = \left[(X_1, Y_1), (X_2 , Y_2) , ... , (X_N, Y_N)\right]$` ,通过正则化的极大似然估计,求解如下优化目标: +$\omega$是特征函数对应的权值,是CRF模型要学习的参数。训练时,对于给定的输入序列和对应的标记序列集合$D = \left[(X_1, Y_1), (X_2 , Y_2) , ... , (X_N, Y_N)\right]$ ,通过正则化的极大似然估计,求解如下优化目标: $$\DeclareMathOperator*{\argmax}{arg\,max} L(\lambda, D) = - \text{log}\left(\prod_{m=1}^{N}p(Y_m|X_m, W)\right) + C \frac{1}{2}\lVert W\rVert^{2}$$ -这个优化目标可以通过反向传播算法和整个神经网络一起求解。解码时,对于给定的输入序列`$X$`,通过解码算法(通常有:维特比算法、Beam Search)求令出条件概率`$\bar{P}(Y|X)$`最大的输出序列 `$\bar{Y}$`。 +这个优化目标可以通过反向传播算法和整个神经网络一起求解。解码时,对于给定的输入序列$X$,通过解码算法(通常有:维特比算法、Beam Search)求令出条件概率$\bar{P}(Y|X)$最大的输出序列 $\bar{Y}$。 ### 深度双向LSTM(DB-LSTM)SRL模型 -在SRL任务中,输入是 “谓词” 和 “一句话”,目标是从这句话中找到谓词的论元,并标注论元的语义角色。如果一个句子含有`$n$`个谓词,这个句子会被处理`$n$`次。一个最为直接的模型是下面这样: +在SRL任务中,输入是 “谓词” 和 “一句话”,目标是从这句话中找到谓词的论元,并标注论元的语义角色。如果一个句子含有$n$个谓词,这个句子会被处理$n$次。一个最为直接的模型是下面这样: 1. 构造输入; -- 输入1是谓词,输入2是句子 -- 将输入1扩展成和输入2一样长的序列,用one-hot方式表示; + - 输入1是谓词,输入2是句子 + - 将输入1扩展成和输入2一样长的序列,用one-hot方式表示; 2. one-hot方式的谓词序列和句子序列通过词表,转换为实向量表示的词向量序列; 3. 将步骤2中的2个词向量序列作为双向LSTM的输入,学习输入序列的特征表示; 4. CRF以步骤3中模型学习到的特征为输入,以标记序列为监督信号,实现序列标注; 大家可以尝试上面这种方法。这里,我们提出一些改进,引入两个简单但对提高系统性能非常有效的特征: -- 谓词上下文:上面的方法中,只用到了谓词的词向量表达谓词相关的所有信息,这种方法始终是非常弱的,特别是如果谓词在句子中出现多次,有可能引起一定的歧义。从经验出发,谓词前后若干个词的一个小片段,能够提供更丰富的信息,帮助消解歧义。于是,我们把这样的经验也添加到模型中,为每个谓词同时抽取一个“谓词上下文” 片段,也就是从这个谓词前后各取`$n$`个词构成的一个窗口片段; +- 谓词上下文:上面的方法中,只用到了谓词的词向量表达谓词相关的所有信息,这种方法始终是非常弱的,特别是如果谓词在句子中出现多次,有可能引起一定的歧义。从经验出发,谓词前后若干个词的一个小片段,能够提供更丰富的信息,帮助消解歧义。于是,我们把这样的经验也添加到模型中,为每个谓词同时抽取一个“谓词上下文” 片段,也就是从这个谓词前后各取$n$个词构成的一个窗口片段; - 谓词上下文区域标记:为句子中的每一个词引入一个0-1二值变量,表示它们是否在“谓词上下文”片段中; 修改后的模型如下(图6是一个深度为4的模型结构示意图): 1. 构造输入 -- 输入1是句子序列,输入2是谓词序列,输入3是谓词上下文,从句子中抽取这个谓词前后各`$n$`个词,构成谓词上下文,用one-hot方式表示,输入4是谓词上下文区域标记,标记了句子中每一个词是否在谓词上下文中; -- 将输入2~3均扩展为和输入1一样长的序列; + - 输入1是句子序列,输入2是谓词序列,输入3是谓词上下文,从句子中抽取这个谓词前后各$n$个词,构成谓词上下文,用one-hot方式表示,输入4是谓词上下文区域标记,标记了句子中每一个词是否在谓词上下文中; + - 将输入2~3均扩展为和输入1一样长的序列; 2. 输入1~4均通过词表取词向量转换为实向量表示的词向量序列;其中输入1、3共享同一个词表,输入2和4各自独有词表; 3. 第2步的4个词向量序列作为双向LSTM模型的输入;LSTM模型学习输入序列的特征表示,得到新的特性表示序列; 4. CRF以第3步中LSTM学习到的特征为输入,以标记序列为监督信号,完成序列标注; -![db_lstm_network](./image/db_lstm_network.png) -
+
+
图6. SRL任务上的深层双向LSTM模型
@@ -137,8 +137,8 @@ $$\DeclareMathOperator*{\argmax}{arg\,max} L(\lambda, D) = - \text{log}\left(\pr ```text conll05st-release/ └── test.wsj -├── props # 标注结果 -└── words # 输入文本序列 + ├── props # 标注结果 + └── words # 输入文本序列 ``` 标注信息源自Penn TreeBank\[[7](#参考文献)\]和PropBank\[[8](#参考文献)\]的标注结果。PropBank标注结果的标签和我们在文章一开始示例中使用的标注结果标签不同,但原理是相同的,关于标注结果标签含义的说明,请参考论文\[[9](#参考文献)\]。 @@ -146,19 +146,11 @@ conll05st-release/ 原始数据需要进行数据预处理才能被PaddlePaddle处理,预处理包括下面几个步骤: 1. 将文本序列和标记序列其合并到一条记录中; -2. 一个句子如果含有`$n$`个谓词,这个句子会被处理`$n$`次,变成`$n$`条独立的训练样本,每个样本一个不同的谓词; +2. 一个句子如果含有$n$个谓词,这个句子会被处理$n$次,变成$n$条独立的训练样本,每个样本一个不同的谓词; 3. 抽取谓词上下文和构造谓词上下文区域标记; 4. 构造以BIO法表示的标记; 5. 依据词典获取词对应的整数索引。 - -```python -# import paddle.v2.dataset.conll05 as conll05 -# conll05.corpus_reader函数完成上面第1步和第2步. -# conll05.reader_creator函数完成上面第3步到第5步. -# conll05.test函数可以获取处理之后的每条样本来供PaddlePaddle训练. -``` - 预处理完成之后一条训练样本包含9个特征,分别是:句子序列、谓词、谓词上下文(占 5 列)、谓词上下区域标志、标注序列。下表是一条训练样本的示例。 | 句子序列 | 谓词 | 谓词上下文(窗口 = 5) | 谓词上下文区域标记 | 标注序列 | @@ -187,6 +179,8 @@ conll05st-release/ 获取词典,打印词典大小: ```python +from __future__ import print_function + import math, os import numpy as np import paddle @@ -201,9 +195,9 @@ word_dict_len = len(word_dict) label_dict_len = len(label_dict) pred_dict_len = len(verb_dict) -print word_dict_len -print label_dict_len -print pred_dict_len +print('word_dict_len: ', word_dict_len) +print('label_dict_len: ', label_dict_len) +print('pred_dict_len: ', pred_dict_len) ``` ## 模型配置说明 @@ -232,96 +226,96 @@ embedding_name = 'emb' ```python # 这里加载PaddlePaddle上版保存的二进制模型 def load_parameter(file_name, h, w): -with open(file_name, 'rb') as f: -f.read(16) # skip header. -return np.fromfile(f, dtype=np.float32).reshape(h, w) + with open(file_name, 'rb') as f: + f.read(16) # skip header. + return np.fromfile(f, dtype=np.float32).reshape(h, w) ``` - 8个LSTM单元以“正向/反向”的顺序对所有输入序列进行学习。 -```python +```python def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, -**ignored): -# 8 features -predicate_embedding = fluid.layers.embedding( -input=predicate, -size=[pred_dict_len, word_dim], -dtype='float32', -is_sparse=IS_SPARSE, -param_attr='vemb') - -mark_embedding = fluid.layers.embedding( -input=mark, -size=[mark_dict_len, mark_dim], -dtype='float32', -is_sparse=IS_SPARSE) - -word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2] -# Since word vector lookup table is pre-trained, we won't update it this time. -# trainable being False prevents updating the lookup table during training. -emb_layers = [ -fluid.layers.embedding( -size=[word_dict_len, word_dim], -input=x, -param_attr=fluid.ParamAttr( -name=embedding_name, trainable=False)) for x in word_input -] -emb_layers.append(predicate_embedding) -emb_layers.append(mark_embedding) - -# 8 LSTM units are trained through alternating left-to-right / right-to-left order -# denoted by the variable `reverse`. -hidden_0_layers = [ -fluid.layers.fc(input=emb, size=hidden_dim, act='tanh') -for emb in emb_layers -] - -hidden_0 = fluid.layers.sums(input=hidden_0_layers) - -lstm_0 = fluid.layers.dynamic_lstm( -input=hidden_0, -size=hidden_dim, -candidate_activation='relu', -gate_activation='sigmoid', -cell_activation='sigmoid') - -# stack L-LSTM and R-LSTM with direct edges -input_tmp = [hidden_0, lstm_0] - -# In PaddlePaddle, state features and transition features of a CRF are implemented -# by a fully connected layer and a CRF layer seperately. The fully connected layer -# with linear activation learns the state features, here we use fluid.layers.sums -# (fluid.layers.fc can be uesed as well), and the CRF layer in PaddlePaddle: -# fluid.layers.linear_chain_crf only -# learns the transition features, which is a cost layer and is the last layer of the network. -# fluid.layers.linear_chain_crf outputs the log probability of true tag sequence -# as the cost by given the input sequence and it requires the true tag sequence -# as target in the learning process. - -for i in range(1, depth): -mix_hidden = fluid.layers.sums(input=[ -fluid.layers.fc(input=input_tmp[0], size=hidden_dim, act='tanh'), -fluid.layers.fc(input=input_tmp[1], size=hidden_dim, act='tanh') -]) - -lstm = fluid.layers.dynamic_lstm( -input=mix_hidden, -size=hidden_dim, -candidate_activation='relu', -gate_activation='sigmoid', -cell_activation='sigmoid', -is_reverse=((i % 2) == 1)) - -input_tmp = [mix_hidden, lstm] - -# 取最后一个栈式LSTM的输出和这个LSTM单元的输入到隐层映射, -# 经过一个全连接层映射到标记字典的维度,来学习 CRF 的状态特征 -feature_out = fluid.layers.sums(input=[ -fluid.layers.fc(input=input_tmp[0], size=label_dict_len, act='tanh'), -fluid.layers.fc(input=input_tmp[1], size=label_dict_len, act='tanh') -]) - -return feature_out + **ignored): + # 8 features + predicate_embedding = fluid.layers.embedding( + input=predicate, + size=[pred_dict_len, word_dim], + dtype='float32', + is_sparse=IS_SPARSE, + param_attr='vemb') + + mark_embedding = fluid.layers.embedding( + input=mark, + size=[mark_dict_len, mark_dim], + dtype='float32', + is_sparse=IS_SPARSE) + + word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2] + # Since word vector lookup table is pre-trained, we won't update it this time. + # trainable being False prevents updating the lookup table during training. + emb_layers = [ + fluid.layers.embedding( + size=[word_dict_len, word_dim], + input=x, + param_attr=fluid.ParamAttr( + name=embedding_name, trainable=False)) for x in word_input + ] + emb_layers.append(predicate_embedding) + emb_layers.append(mark_embedding) + + # 8 LSTM units are trained through alternating left-to-right / right-to-left order + # denoted by the variable `reverse`. + hidden_0_layers = [ + fluid.layers.fc(input=emb, size=hidden_dim, act='tanh') + for emb in emb_layers + ] + + hidden_0 = fluid.layers.sums(input=hidden_0_layers) + + lstm_0 = fluid.layers.dynamic_lstm( + input=hidden_0, + size=hidden_dim, + candidate_activation='relu', + gate_activation='sigmoid', + cell_activation='sigmoid') + + # stack L-LSTM and R-LSTM with direct edges + input_tmp = [hidden_0, lstm_0] + + # In PaddlePaddle, state features and transition features of a CRF are implemented + # by a fully connected layer and a CRF layer seperately. The fully connected layer + # with linear activation learns the state features, here we use fluid.layers.sums + # (fluid.layers.fc can be uesed as well), and the CRF layer in PaddlePaddle: + # fluid.layers.linear_chain_crf only + # learns the transition features, which is a cost layer and is the last layer of the network. + # fluid.layers.linear_chain_crf outputs the log probability of true tag sequence + # as the cost by given the input sequence and it requires the true tag sequence + # as target in the learning process. + + for i in range(1, depth): + mix_hidden = fluid.layers.sums(input=[ + fluid.layers.fc(input=input_tmp[0], size=hidden_dim, act='tanh'), + fluid.layers.fc(input=input_tmp[1], size=hidden_dim, act='tanh') + ]) + + lstm = fluid.layers.dynamic_lstm( + input=mix_hidden, + size=hidden_dim, + candidate_activation='relu', + gate_activation='sigmoid', + cell_activation='sigmoid', + is_reverse=((i % 2) == 1)) + + input_tmp = [mix_hidden, lstm] + + # 取最后一个栈式LSTM的输出和这个LSTM单元的输入到隐层映射, + # 经过一个全连接层映射到标记字典的维度,来学习 CRF 的状态特征 + feature_out = fluid.layers.sums(input=[ + fluid.layers.fc(input=input_tmp[0], size=label_dict_len, act='tanh'), + fluid.layers.fc(input=input_tmp[1], size=label_dict_len, act='tanh') + ]) + + return feature_out ``` ## 训练模型 @@ -338,116 +332,116 @@ return feature_out ```python def train(use_cuda, save_dirname=None, is_local=True): -# define network topology - -# 句子序列 -word = fluid.layers.data( -name='word_data', shape=[1], dtype='int64', lod_level=1) - -# 谓词 -predicate = fluid.layers.data( -name='verb_data', shape=[1], dtype='int64', lod_level=1) - -# 谓词上下文5个特征 -ctx_n2 = fluid.layers.data( -name='ctx_n2_data', shape=[1], dtype='int64', lod_level=1) -ctx_n1 = fluid.layers.data( -name='ctx_n1_data', shape=[1], dtype='int64', lod_level=1) -ctx_0 = fluid.layers.data( -name='ctx_0_data', shape=[1], dtype='int64', lod_level=1) -ctx_p1 = fluid.layers.data( -name='ctx_p1_data', shape=[1], dtype='int64', lod_level=1) -ctx_p2 = fluid.layers.data( -name='ctx_p2_data', shape=[1], dtype='int64', lod_level=1) - -# 谓词上下区域标志 -mark = fluid.layers.data( -name='mark_data', shape=[1], dtype='int64', lod_level=1) - -# define network topology -feature_out = db_lstm(**locals()) - -# 标注序列 -target = fluid.layers.data( -name='target', shape=[1], dtype='int64', lod_level=1) - -# 学习 CRF 的转移特征 -crf_cost = fluid.layers.linear_chain_crf( -input=feature_out, -label=target, -param_attr=fluid.ParamAttr( -name='crfw', learning_rate=mix_hidden_lr)) - -avg_cost = fluid.layers.mean(crf_cost) - -sgd_optimizer = fluid.optimizer.SGD( -learning_rate=fluid.layers.exponential_decay( -learning_rate=0.01, -decay_steps=100000, -decay_rate=0.5, -staircase=True)) - -sgd_optimizer.minimize(avg_cost) - -# The CRF decoding layer is used for evaluation and inference. -# It shares weights with CRF layer. The sharing of parameters among multiple layers -# is specified by using the same parameter name in these layers. If true tag sequence -# is provided in training process, `fluid.layers.crf_decoding` calculates labelling error -# for each input token and sums the error over the entire sequence. -# Otherwise, `fluid.layers.crf_decoding` generates the labelling tags. -crf_decode = fluid.layers.crf_decoding( -input=feature_out, param_attr=fluid.ParamAttr(name='crfw')) - -train_data = paddle.batch( -paddle.reader.shuffle( -paddle.dataset.conll05.test(), buf_size=8192), -batch_size=BATCH_SIZE) - -place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - - -feeder = fluid.DataFeeder( -feed_list=[ -word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, predicate, mark, target -], -place=place) -exe = fluid.Executor(place) - -def train_loop(main_program): -exe.run(fluid.default_startup_program()) -embedding_param = fluid.global_scope().find_var( -embedding_name).get_tensor() -embedding_param.set( -load_parameter(conll05.get_embedding(), word_dict_len, word_dim), -place) - -start_time = time.time() -batch_id = 0 -for pass_id in xrange(PASS_NUM): -for data in train_data(): -cost = exe.run(main_program, -feed=feeder.feed(data), -fetch_list=[avg_cost]) -cost = cost[0] - -if batch_id % 10 == 0: -print("avg_cost:" + str(cost)) -if batch_id != 0: -print("second per batch: " + str((time.time( -) - start_time) / batch_id)) -# Set the threshold low to speed up the CI test -if float(cost) < 60.0: -if save_dirname is not None: -fluid.io.save_inference_model(save_dirname, [ -'word_data', 'verb_data', 'ctx_n2_data', -'ctx_n1_data', 'ctx_0_data', 'ctx_p1_data', -'ctx_p2_data', 'mark_data' -], [feature_out], exe) -return - -batch_id = batch_id + 1 - -train_loop(fluid.default_main_program()) + # define network topology + + # 句子序列 + word = fluid.layers.data( + name='word_data', shape=[1], dtype='int64', lod_level=1) + + # 谓词 + predicate = fluid.layers.data( + name='verb_data', shape=[1], dtype='int64', lod_level=1) + + # 谓词上下文5个特征 + ctx_n2 = fluid.layers.data( + name='ctx_n2_data', shape=[1], dtype='int64', lod_level=1) + ctx_n1 = fluid.layers.data( + name='ctx_n1_data', shape=[1], dtype='int64', lod_level=1) + ctx_0 = fluid.layers.data( + name='ctx_0_data', shape=[1], dtype='int64', lod_level=1) + ctx_p1 = fluid.layers.data( + name='ctx_p1_data', shape=[1], dtype='int64', lod_level=1) + ctx_p2 = fluid.layers.data( + name='ctx_p2_data', shape=[1], dtype='int64', lod_level=1) + + # 谓词上下区域标志 + mark = fluid.layers.data( + name='mark_data', shape=[1], dtype='int64', lod_level=1) + + # define network topology + feature_out = db_lstm(**locals()) + + # 标注序列 + target = fluid.layers.data( + name='target', shape=[1], dtype='int64', lod_level=1) + + # 学习 CRF 的转移特征 + crf_cost = fluid.layers.linear_chain_crf( + input=feature_out, + label=target, + param_attr=fluid.ParamAttr( + name='crfw', learning_rate=mix_hidden_lr)) + + avg_cost = fluid.layers.mean(crf_cost) + + sgd_optimizer = fluid.optimizer.SGD( + learning_rate=fluid.layers.exponential_decay( + learning_rate=0.01, + decay_steps=100000, + decay_rate=0.5, + staircase=True)) + + sgd_optimizer.minimize(avg_cost) + + # The CRF decoding layer is used for evaluation and inference. + # It shares weights with CRF layer. The sharing of parameters among multiple layers + # is specified by using the same parameter name in these layers. If true tag sequence + # is provided in training process, `fluid.layers.crf_decoding` calculates labelling error + # for each input token and sums the error over the entire sequence. + # Otherwise, `fluid.layers.crf_decoding` generates the labelling tags. + crf_decode = fluid.layers.crf_decoding( + input=feature_out, param_attr=fluid.ParamAttr(name='crfw')) + + train_data = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.conll05.test(), buf_size=8192), + batch_size=BATCH_SIZE) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + + feeder = fluid.DataFeeder( + feed_list=[ + word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, predicate, mark, target + ], + place=place) + exe = fluid.Executor(place) + + def train_loop(main_program): + exe.run(fluid.default_startup_program()) + embedding_param = fluid.global_scope().find_var( + embedding_name).get_tensor() + embedding_param.set( + load_parameter(conll05.get_embedding(), word_dict_len, word_dim), + place) + + start_time = time.time() + batch_id = 0 + for pass_id in xrange(PASS_NUM): + for data in train_data(): + cost = exe.run(main_program, + feed=feeder.feed(data), + fetch_list=[avg_cost]) + cost = cost[0] + + if batch_id % 10 == 0: + print("avg_cost: " + str(cost)) + if batch_id != 0: + print("second per batch: " + str((time.time( + ) - start_time) / batch_id)) + # Set the threshold low to speed up the CI test + if float(cost) < 60.0: + if save_dirname is not None: + fluid.io.save_inference_model(save_dirname, [ + 'word_data', 'verb_data', 'ctx_n2_data', + 'ctx_n1_data', 'ctx_0_data', 'ctx_p1_data', + 'ctx_p2_data', 'mark_data' + ], [feature_out], exe) + return + + batch_id = batch_id + 1 + + train_loop(fluid.default_main_program()) ``` @@ -457,92 +451,92 @@ train_loop(fluid.default_main_program()) ```python def infer(use_cuda, save_dirname=None): -if save_dirname is None: -return - -place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() -exe = fluid.Executor(place) - -inference_scope = fluid.core.Scope() -with fluid.scope_guard(inference_scope): -# Use fluid.io.load_inference_model to obtain the inference program desc, -# the feed_target_names (the names of variables that will be fed -# data using feed operators), and the fetch_targets (variables that -# we want to obtain data from using fetch operators). -[inference_program, feed_target_names, -fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) - -# Setup inputs by creating LoDTensors to represent sequences of words. -# Here each word is the basic element of these LoDTensors and the shape of -# each word (base_shape) should be [1] since it is simply an index to -# look up for the corresponding word vector. -# Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], -# which has only one lod level. Then the created LoDTensors will have only -# one higher level structure (sequence of words, or sentence) than the basic -# element (word). Hence the LoDTensor will hold data for three sentences of -# length 3, 4 and 2, respectively. -# Note that lod info should be a list of lists. -lod = [[3, 4, 2]] -base_shape = [1] -# The range of random integers is [low, high] -word = fluid.create_random_int_lodtensor( -lod, base_shape, place, low=0, high=word_dict_len - 1) -pred = fluid.create_random_int_lodtensor( -lod, base_shape, place, low=0, high=pred_dict_len - 1) -ctx_n2 = fluid.create_random_int_lodtensor( -lod, base_shape, place, low=0, high=word_dict_len - 1) -ctx_n1 = fluid.create_random_int_lodtensor( -lod, base_shape, place, low=0, high=word_dict_len - 1) -ctx_0 = fluid.create_random_int_lodtensor( -lod, base_shape, place, low=0, high=word_dict_len - 1) -ctx_p1 = fluid.create_random_int_lodtensor( -lod, base_shape, place, low=0, high=word_dict_len - 1) -ctx_p2 = fluid.create_random_int_lodtensor( -lod, base_shape, place, low=0, high=word_dict_len - 1) -mark = fluid.create_random_int_lodtensor( -lod, base_shape, place, low=0, high=mark_dict_len - 1) - -# Construct feed as a dictionary of {feed_target_name: feed_target_data} -# and results will contain a list of data corresponding to fetch_targets. -assert feed_target_names[0] == 'word_data' -assert feed_target_names[1] == 'verb_data' -assert feed_target_names[2] == 'ctx_n2_data' -assert feed_target_names[3] == 'ctx_n1_data' -assert feed_target_names[4] == 'ctx_0_data' -assert feed_target_names[5] == 'ctx_p1_data' -assert feed_target_names[6] == 'ctx_p2_data' -assert feed_target_names[7] == 'mark_data' - -results = exe.run(inference_program, -feed={ -feed_target_names[0]: word, -feed_target_names[1]: pred, -feed_target_names[2]: ctx_n2, -feed_target_names[3]: ctx_n1, -feed_target_names[4]: ctx_0, -feed_target_names[5]: ctx_p1, -feed_target_names[6]: ctx_p2, -feed_target_names[7]: mark -}, -fetch_list=fetch_targets, -return_numpy=False) -print(results[0].lod()) -np_data = np.array(results[0]) -print("Inference Shape: ", np_data.shape) + if save_dirname is None: + return + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place) + + inference_scope = fluid.core.Scope() + with fluid.scope_guard(inference_scope): + # Use fluid.io.load_inference_model to obtain the inference program desc, + # the feed_target_names (the names of variables that will be fed + # data using feed operators), and the fetch_targets (variables that + # we want to obtain data from using fetch operators). + [inference_program, feed_target_names, + fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) + + # Setup inputs by creating LoDTensors to represent sequences of words. + # Here each word is the basic element of these LoDTensors and the shape of + # each word (base_shape) should be [1] since it is simply an index to + # look up for the corresponding word vector. + # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], + # which has only one lod level. Then the created LoDTensors will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that lod info should be a list of lists. + lod = [[3, 4, 2]] + base_shape = [1] + # The range of random integers is [low, high] + word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + pred = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=pred_dict_len - 1) + ctx_n2 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + ctx_n1 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + ctx_0 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + ctx_p1 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + ctx_p2 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + mark = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=mark_dict_len - 1) + + # Construct feed as a dictionary of {feed_target_name: feed_target_data} + # and results will contain a list of data corresponding to fetch_targets. + assert feed_target_names[0] == 'word_data' + assert feed_target_names[1] == 'verb_data' + assert feed_target_names[2] == 'ctx_n2_data' + assert feed_target_names[3] == 'ctx_n1_data' + assert feed_target_names[4] == 'ctx_0_data' + assert feed_target_names[5] == 'ctx_p1_data' + assert feed_target_names[6] == 'ctx_p2_data' + assert feed_target_names[7] == 'mark_data' + + results = exe.run(inference_program, + feed={ + feed_target_names[0]: word, + feed_target_names[1]: pred, + feed_target_names[2]: ctx_n2, + feed_target_names[3]: ctx_n1, + feed_target_names[4]: ctx_0, + feed_target_names[5]: ctx_p1, + feed_target_names[6]: ctx_p2, + feed_target_names[7]: mark + }, + fetch_list=fetch_targets, + return_numpy=False) + print(results[0].lod()) + np_data = np.array(results[0]) + print("Inference Shape: ", np_data.shape) ``` 整个程序的入口如下: ```python def main(use_cuda, is_local=True): -if use_cuda and not fluid.core.is_compiled_with_cuda(): -return + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return -# Directory for saving the trained model -save_dirname = "label_semantic_roles.inference.model" + # Directory for saving the trained model + save_dirname = "label_semantic_roles.inference.model" -train(use_cuda, save_dirname, is_local) -infer(use_cuda, save_dirname) + train(use_cuda, save_dirname, is_local) + infer(use_cuda, save_dirname) main(use_cuda=False) diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm.png deleted file mode 100644 index e63f5ebd6d00f2e4ecf97b9ab2027e74683013f2..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm_en.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm_en.png deleted file mode 100755 index f0a195c24d9ee493f96bb93c28a99e70566be7a4..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example.png deleted file mode 100755 index e5f7151c9fcc50a7cf7af485cbbc7e4fccab0c20..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example_en.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example_en.png deleted file mode 100755 index 93b44dd4874402ef29ad7bd7d94147609b92e309..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network.png deleted file mode 100644 index 592f7ee23bdc88a9a35059612e5ab880bbc9d34b..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network_en.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network_en.png deleted file mode 100755 index c3646312e48db977402fb353dc0c9b4d02269bf4..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing.png deleted file mode 100755 index 9265b671735940ed6549e2980064d2ce08baae64..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing_en.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing_en.png deleted file mode 100755 index 23f4f45b603e3d60702af2b2464d10fc8deed061..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/linear_chain_crf.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/linear_chain_crf.png deleted file mode 100644 index 0778fda74b2ad22ce4b631791a7b028cdef780a5..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/linear_chain_crf.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm.png deleted file mode 100644 index 3d2914c726b5f4c46e66dfa85d4e88649fede6b3..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm_en.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm_en.png deleted file mode 100755 index 0b944ef91e8b5ba4b14d2a35bd8879f261cf8f61..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/index.md b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/README.cn.md similarity index 70% rename from doc/fluid/new_docs/beginners_guide/basics/machine_translation/index.md rename to doc/fluid/new_docs/beginners_guide/basics/machine_translation/README.cn.md index fc161aaae9c37b0e1a596204e7138025a98adb1d..3f6efec8841a608638389331152a720cbb3d9735 100644 --- a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/index.md +++ b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/README.cn.md @@ -11,10 +11,10 @@ 为解决以上问题,统计机器翻译(Statistical Machine Translation, SMT)技术应运而生。在统计机器翻译技术中,转化规则是由机器自动从大规模的语料中学习得到的,而非我们人主动提供规则。因此,它克服了基于规则的翻译系统所面临的知识获取瓶颈的问题,但仍然存在许多挑战:1)人为设计许多特征(feature),但永远无法覆盖所有的语言现象;2)难以利用全局的特征;3)依赖于许多预处理环节,如词语对齐、分词或符号化(tokenization)、规则抽取、句法分析等,而每个环节的错误会逐步累积,对翻译的影响也越来越大。 近年来,深度学习技术的发展为解决上述挑战提供了新的思路。将深度学习应用于机器翻译任务的方法大致分为两类:1)仍以统计机器翻译系统为框架,只是利用神经网络来改进其中的关键模块,如语言模型、调序模型等(见图1的左半部分);2)不再以统计机器翻译系统为框架,而是直接用神经网络将源语言映射到目标语言,即端到端的神经网络机器翻译(End-to-End Neural Machine Translation, End-to-End NMT)(见图1的右半部分),简称为NMT模型。 -![nmt](./image/nmt.png) -

+

+
图1. 基于神经网络的机器翻译系统 -

+
本教程主要介绍NMT模型,以及如何用PaddlePaddle来训练一个NMT模型。 @@ -30,7 +30,9 @@ 1 -6.23177 These are the light of hope and relief . 2 -7.7914 These are the light of hope and the relief of hope . ``` + - 左起第一列是生成句子的序号;左起第二列是该条句子的得分(从大到小),分值越高越好;左起第三列是生成的英语句子。 + - 另外有两个特殊标志:``表示句子的结尾,``表示未登录词(unknown word),即未在训练字典中出现的词。 ## 模型概览 @@ -43,18 +45,20 @@ 具体来说,该双向循环神经网络分别在时间维以顺序和逆序——即前向(forward)和后向(backward)——依次处理输入序列,并将每个时间步RNN的输出拼接成为最终的输出层。这样每个时间步的输出节点,都包含了输入序列中当前时刻完整的过去和未来的上下文信息。下图展示的是一个按时间步展开的双向循环神经网络。该网络包含一个前向和一个后向RNN,其中有六个权重矩阵:输入到前向隐层和后向隐层的权重矩阵(`$W_1, W_3$`),隐层到隐层自己的权重矩阵(`$W_2,W_5$`),前向隐层和后向隐层到输出层的权重矩阵(`$W_4, W_6$`)。注意,该网络的前向隐层和后向隐层之间没有连接。 -![bi_rnn](./image/bi_rnn.png) -

-图3. 按时间步展开的双向循环神经网络 -

+ +
+
+图2. 按时间步展开的双向循环神经网络 +
### 编码器-解码器框架 编码器-解码器(Encoder-Decoder)\[[2](#参考文献)\]框架用于解决由一个任意长度的源序列到另一个任意长度的目标序列的变换问题。即编码阶段将整个源序列编码成一个向量,解码阶段通过最大化预测序列概率,从中解码出整个目标序列。编码和解码的过程通常都使用RNN实现。 ![encoder_decoder](./image/encoder_decoder.png) -

-图4. 编码器-解码器框架 -

+
+
+图3. 编码器-解码器框架 +
#### 编码器 @@ -67,29 +71,24 @@ 3. 用RNN编码源语言词序列:这一过程的计算公式为`$h_i=\varnothing _\theta \left ( h_{i-1}, s_i \right )$`,其中`$h_0$`是一个全零的向量,`$\varnothing _\theta$`是一个非线性激活函数,最后得到的`$\mathbf{h}=\left \{ h_1,..., h_T \right \}$`就是RNN依次读入源语言`$T$`个词的状态编码序列。整句话的向量表示可以采用`$\mathbf{h}$`在最后一个时间步`$T$`的状态编码,或使用时间维上的池化(pooling)结果。 第3步也可以使用双向循环神经网络实现更复杂的句编码表示,具体可以用双向GRU实现。前向GRU按照词序列`$(x_1,x_2,...,x_T)$`的顺序依次编码源语言端词,并得到一系列隐层状态`$(\overrightarrow{h_1},\overrightarrow{h_2},...,\overrightarrow{h_T})$`。类似的,后向GRU按照`$(x_T,x_{T-1},...,x_1)$`的顺序依次编码源语言端词,得到`$(\overleftarrow{h_1},\overleftarrow{h_2},...,\overleftarrow{h_T})$`。最后对于词`$x_i$`,通过拼接两个GRU的结果得到它的隐层状态,即`$h_i=\left [ \overrightarrow{h_i^T},\overleftarrow{h_i^T} \right ]^{T}$`。 - -![encoder_attention](./image/encoder_attention.png) -

-图5. 使用双向GRU的编码器 -

+
+
+图4. 使用双向GRU的编码器 +
#### 解码器 机器翻译任务的训练过程中,解码阶段的目标是最大化下一个正确的目标语言词的概率。思路是: - 1. 每一个时刻,根据源语言句子的编码信息(又叫上下文向量,context vector)`$c$`、真实目标语言序列的第`$i$`个词`$u_i$`和`$i$`时刻RNN的隐层状态`$z_i$`,计算出下一个隐层状态`$z_{i+1}$`。计算公式如下: - -$$z_{i+1}=\phi _{\theta '}\left ( c,u_i,z_i \right )$$ - +$$z_{i+1}=\phi_{\theta '} \left ( c,u_i,z_i \right )$$ 其中`$\phi _{\theta '}$`是一个非线性激活函数;`$c=q\mathbf{h}$`是源语言句子的上下文向量,在不使用[注意力机制](#注意力机制)时,如果[编码器](#编码器)的输出是源语言句子编码后的最后一个元素,则可以定义`$c=h_T$`;`$u_i$`是目标语言序列的第`$i$`个单词,`$u_0$`是目标语言序列的开始标记``,表示解码开始;`$z_i$`是`$i$`时刻解码RNN的隐层状态,`$z_0$`是一个全零的向量。 2. 将`$z_{i+1}$`通过`softmax`归一化,得到目标语言序列的第`$i+1$`个单词的概率分布`$p_{i+1}$`。概率分布公式如下: - $$p\left ( u_{i+1}|u_{<i+1},\mathbf{x} \right )=softmax(W_sz_{i+1}+b_z)$$ - 其中`$W_sz_{i+1}+b_z$`是对每个可能的输出单词进行打分,再用softmax归一化就可以得到第`$i+1$`个词的概率`$p_{i+1}$`。 3. 根据`$p_{i+1}$`和`$u_{i+1}$`计算代价。 + 4. 重复步骤1~3,直到目标语言序列中的所有词处理完毕。 机器翻译任务的生成过程,通俗来讲就是根据预先训练的模型来翻译源语言句子。生成过程中的解码阶段和上述训练过程的有所差异,具体介绍请见[柱搜索算法](#柱搜索算法)。 @@ -101,10 +100,12 @@ $$p\left ( u_{i+1}|u_{<i+1},\mathbf{x} \right )=softmax(W_sz_{i+1}+b_z)$$ 柱搜索算法使用广度优先策略建立搜索树,在树的每一层,按照启发代价(heuristic cost)(本教程中,为生成词的log概率之和)对节点进行排序,然后仅留下预先确定的个数(文献中通常称为beam width、beam size、柱宽度等)的节点。只有这些节点会在下一层继续扩展,其他节点就被剪掉了,也就是说保留了质量较高的节点,剪枝了质量较差的节点。因此,搜索所占用的空间和时间大幅减少,但缺点是无法保证一定获得最优解。 使用柱搜索算法的解码阶段,目标是最大化生成序列的概率。思路是: - 1. 每一个时刻,根据源语言句子的编码信息`$c$`、生成的第`$i$`个目标语言序列单词`$u_i$`和`$i$`时刻RNN的隐层状态`$z_i$`,计算出下一个隐层状态`$z_{i+1}$`。 + 2. 将`$z_{i+1}$`通过`softmax`归一化,得到目标语言序列的第`$i+1$`个单词的概率分布`$p_{i+1}$`。 + 3. 根据`$p_{i+1}$`采样出单词`$u_{i+1}$`。 + 4. 重复步骤1~3,直到获得句子结束标记``或超过句子的最大生成长度为止。 注意:`$z_{i+1}$`和`$p_{i+1}$`的计算公式同[解码器](#解码器)中的一样。且由于生成时的每一步都是通过贪心法实现的,因此并不能保证得到全局最优解。 @@ -116,9 +117,13 @@ $$p\left ( u_{i+1}|u_{<i+1},\mathbf{x} \right )=softmax(W_sz_{i+1}+b_z)$$ ### 数据预处理 我们的预处理流程包括两步: + - 将每个源语言到目标语言的平行语料库文件合并为一个文件: + - 合并每个`XXX.src`和`XXX.trg`文件为`XXX`。 + - `XXX`中的第`$i$`行内容为`XXX.src`中的第`$i$`行和`XXX.trg`中的第`$i$`行连接,用'\t'分隔。 + - 创建训练数据的“源字典”和“目标字典”。每个字典都有**DICTSIZE**个单词,包括:语料中词频最高的(DICTSIZE - 3)个单词,和3个特殊符号``(序列的开始)、``(序列的结束)和``(未登录词)。 ### 示例数据 @@ -132,6 +137,7 @@ $$p\left ( u_{i+1}|u_{<i+1},\mathbf{x} \right )=softmax(W_sz_{i+1}+b_z)$$ 下面我们开始根据输入数据的形式配置模型。首先引入所需的库函数以及定义全局变量。 ```python +from __future__ import print_function import contextlib import numpy as np @@ -157,139 +163,152 @@ decoder_size = hidden_dim 然后如下实现编码器框架: -```python -def encoder(is_sparse): -src_word_id = pd.data( -name="src_word_id", shape=[1], dtype='int64', lod_level=1) -src_embedding = pd.embedding( -input=src_word_id, -size=[dict_size, word_dim], -dtype='float32', -is_sparse=is_sparse, -param_attr=fluid.ParamAttr(name='vemb')) - -fc1 = pd.fc(input=src_embedding, size=hidden_dim * 4, act='tanh') -lstm_hidden0, lstm_0 = pd.dynamic_lstm(input=fc1, size=hidden_dim * 4) -encoder_out = pd.sequence_last_step(input=lstm_hidden0) -return encoder_out -``` + ```python + def encoder(is_sparse): + src_word_id = pd.data( + name="src_word_id", shape=[1], dtype='int64', lod_level=1) + src_embedding = pd.embedding( + input=src_word_id, + size=[dict_size, word_dim], + dtype='float32', + is_sparse=is_sparse, + param_attr=fluid.ParamAttr(name='vemb')) + + fc1 = pd.fc(input=src_embedding, size=hidden_dim * 4, act='tanh') + lstm_hidden0, lstm_0 = pd.dynamic_lstm(input=fc1, size=hidden_dim * 4) + encoder_out = pd.sequence_last_step(input=lstm_hidden0) + return encoder_out + ``` 再实现训练模式下的解码器: ```python -def train_decoder(context, is_sparse): -trg_language_word = pd.data( -name="target_language_word", shape=[1], dtype='int64', lod_level=1) -trg_embedding = pd.embedding( -input=trg_language_word, -size=[dict_size, word_dim], -dtype='float32', -is_sparse=is_sparse, -param_attr=fluid.ParamAttr(name='vemb')) - -rnn = pd.DynamicRNN() -with rnn.block(): -current_word = rnn.step_input(trg_embedding) -pre_state = rnn.memory(init=context) -current_state = pd.fc(input=[current_word, pre_state], -size=decoder_size, -act='tanh') - -current_score = pd.fc(input=current_state, -size=target_dict_dim, -act='softmax') -rnn.update_memory(pre_state, current_state) -rnn.output(current_score) - -return rnn() + def train_decoder(context, is_sparse): + trg_language_word = pd.data( + name="target_language_word", shape=[1], dtype='int64', lod_level=1) + trg_embedding = pd.embedding( + input=trg_language_word, + size=[dict_size, word_dim], + dtype='float32', + is_sparse=is_sparse, + param_attr=fluid.ParamAttr(name='vemb')) + + rnn = pd.DynamicRNN() + with rnn.block(): + current_word = rnn.step_input(trg_embedding) + pre_state = rnn.memory(init=context) + current_state = pd.fc(input=[current_word, pre_state], + size=decoder_size, + act='tanh') + + current_score = pd.fc(input=current_state, + size=target_dict_dim, + act='softmax') + rnn.update_memory(pre_state, current_state) + rnn.output(current_score) + + return rnn() ``` 实现推测模式下的解码器: ```python def decode(context, is_sparse): -init_state = context -array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length) -counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True) - -# fill the first element with init_state -state_array = pd.create_array('float32') -pd.array_write(init_state, array=state_array, i=counter) - -# ids, scores as memory -ids_array = pd.create_array('int64') -scores_array = pd.create_array('float32') - -init_ids = pd.data(name="init_ids", shape=[1], dtype="int64", lod_level=2) -init_scores = pd.data( -name="init_scores", shape=[1], dtype="float32", lod_level=2) - -pd.array_write(init_ids, array=ids_array, i=counter) -pd.array_write(init_scores, array=scores_array, i=counter) - -cond = pd.less_than(x=counter, y=array_len) - -while_op = pd.While(cond=cond) -with while_op.block(): -pre_ids = pd.array_read(array=ids_array, i=counter) -pre_state = pd.array_read(array=state_array, i=counter) -pre_score = pd.array_read(array=scores_array, i=counter) - -# expand the lod of pre_state to be the same with pre_score -pre_state_expanded = pd.sequence_expand(pre_state, pre_score) - -pre_ids_emb = pd.embedding( -input=pre_ids, -size=[dict_size, word_dim], -dtype='float32', -is_sparse=is_sparse) - -# use rnn unit to update rnn -current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb], -size=decoder_size, -act='tanh') -current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score) -# use score to do beam search -current_score = pd.fc(input=current_state_with_lod, -size=target_dict_dim, -act='softmax') -topk_scores, topk_indices = pd.topk(current_score, k=topk_size) -selected_ids, selected_scores = pd.beam_search( -pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0) - -pd.increment(x=counter, value=1, in_place=True) - -# update the memories -pd.array_write(current_state, array=state_array, i=counter) -pd.array_write(selected_ids, array=ids_array, i=counter) -pd.array_write(selected_scores, array=scores_array, i=counter) - -pd.less_than(x=counter, y=array_len, cond=cond) - -translation_ids, translation_scores = pd.beam_search_decode( -ids=ids_array, scores=scores_array) - -return translation_ids, translation_scores + init_state = context + array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length) + counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True) + + # fill the first element with init_state + state_array = pd.create_array('float32') + pd.array_write(init_state, array=state_array, i=counter) + + # ids, scores as memory + ids_array = pd.create_array('int64') + scores_array = pd.create_array('float32') + + init_ids = pd.data(name="init_ids", shape=[1], dtype="int64", lod_level=2) + init_scores = pd.data( + name="init_scores", shape=[1], dtype="float32", lod_level=2) + + pd.array_write(init_ids, array=ids_array, i=counter) + pd.array_write(init_scores, array=scores_array, i=counter) + + cond = pd.less_than(x=counter, y=array_len) + + while_op = pd.While(cond=cond) + with while_op.block(): + pre_ids = pd.array_read(array=ids_array, i=counter) + pre_state = pd.array_read(array=state_array, i=counter) + pre_score = pd.array_read(array=scores_array, i=counter) + + # expand the lod of pre_state to be the same with pre_score + pre_state_expanded = pd.sequence_expand(pre_state, pre_score) + + pre_ids_emb = pd.embedding( + input=pre_ids, + size=[dict_size, word_dim], + dtype='float32', + is_sparse=is_sparse) + + # use rnn unit to update rnn + current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb], + size=decoder_size, + act='tanh') + current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score) + # use score to do beam search + current_score = pd.fc(input=current_state_with_lod, + size=target_dict_dim, + act='softmax') + topk_scores, topk_indices = pd.topk(current_score, k=beam_size) + # calculate accumulated scores after topk to reduce computation cost + accu_scores = pd.elementwise_add( + x=pd.log(topk_scores), y=pd.reshape(pre_score, shape=[-1]), axis=0) + selected_ids, selected_scores = pd.beam_search( + pre_ids, + pre_score, + topk_indices, + accu_scores, + beam_size, + end_id=10, + level=0) + + pd.increment(x=counter, value=1, in_place=True) + + # update the memories + pd.array_write(current_state, array=state_array, i=counter) + pd.array_write(selected_ids, array=ids_array, i=counter) + pd.array_write(selected_scores, array=scores_array, i=counter) + + # update the break condition: up to the max length or all candidates of + # source sentences have ended. + length_cond = pd.less_than(x=counter, y=array_len) + finish_cond = pd.logical_not(pd.is_empty(x=selected_ids)) + pd.logical_and(x=length_cond, y=finish_cond, out=cond) + + translation_ids, translation_scores = pd.beam_search_decode( + ids=ids_array, scores=scores_array, beam_size=beam_size, end_id=10) + + return translation_ids, translation_scores ``` 进而,我们定义一个`train_program`来使用`inference_program`计算出的结果,在标记数据的帮助下来计算误差。我们还定义了一个`optimizer_func`来定义优化器。 ```python def train_program(is_sparse): -context = encoder(is_sparse) -rnn_out = train_decoder(context, is_sparse) -label = pd.data( -name="target_language_next_word", shape=[1], dtype='int64', lod_level=1) -cost = pd.cross_entropy(input=rnn_out, label=label) -avg_cost = pd.mean(cost) -return avg_cost + context = encoder(is_sparse) + rnn_out = train_decoder(context, is_sparse) + label = pd.data( + name="target_language_next_word", shape=[1], dtype='int64', lod_level=1) + cost = pd.cross_entropy(input=rnn_out, label=label) + avg_cost = pd.mean(cost) + return avg_cost def optimizer_func(): -return fluid.optimizer.Adagrad( -learning_rate=1e-4, -regularization=fluid.regularizer.L2DecayRegularizer( -regularization_coeff=0.1)) + return fluid.optimizer.Adagrad( + learning_rate=1e-4, + regularization=fluid.regularizer.L2DecayRegularizer( + regularization_coeff=0.1)) ``` ## 训练模型 @@ -307,9 +326,9 @@ place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() ```python train_reader = paddle.batch( -paddle.reader.shuffle( -paddle.dataset.wmt14.train(dict_size), buf_size=1000), -batch_size=batch_size) + paddle.reader.shuffle( + paddle.dataset.wmt14.train(dict_size), buf_size=1000), + batch_size=batch_size) ``` ### 构造训练器(trainer) @@ -318,9 +337,9 @@ batch_size=batch_size) ```python is_sparse = False trainer = fluid.Trainer( -train_func=partial(train_program, is_sparse), -place=place, -optimizer_func=optimizer_func) + train_func=partial(train_program, is_sparse), + place=place, + optimizer_func=optimizer_func) ``` ### 提供数据 @@ -329,8 +348,8 @@ optimizer_func=optimizer_func) ```python feed_order = [ -'src_word_id', 'target_language_word', 'target_language_next_word' -] + 'src_word_id', 'target_language_word', 'target_language_next_word' + ] ``` ### 事件处理器 @@ -338,12 +357,12 @@ feed_order = [ ```python def event_handler(event): -if isinstance(event, fluid.EndStepEvent): -if event.step % 10 == 0: -print('pass_id=' + str(event.epoch) + ' batch=' + str(event.step)) + if isinstance(event, fluid.EndStepEvent): + if event.step % 10 == 0: + print('pass_id=' + str(event.epoch) + ' batch=' + str(event.step)) -if event.step == 20: -trainer.stop() + if event.step == 20: + trainer.stop() ``` ### 开始训练 @@ -353,10 +372,10 @@ trainer.stop() EPOCH_NUM = 1 trainer.train( -reader=train_reader, -num_epochs=EPOCH_NUM, -event_handler=event_handler, -feed_order=feed_order) + reader=train_reader, + num_epochs=EPOCH_NUM, + event_handler=event_handler, + feed_order=feed_order) ``` ## 应用模型 @@ -377,7 +396,7 @@ translation_ids, translation_scores = decode(context, is_sparse) ```python init_ids_data = np.array([1 for _ in range(batch_size)], dtype='int64') init_scores_data = np.array( -[1. for _ in range(batch_size)], dtype='float32') + [1. for _ in range(batch_size)], dtype='float32') init_ids_data = init_ids_data.reshape((batch_size, 1)) init_scores_data = init_scores_data.reshape((batch_size, 1)) init_lod = [1] * batch_size @@ -387,14 +406,14 @@ init_ids = fluid.create_lod_tensor(init_ids_data, init_lod, place) init_scores = fluid.create_lod_tensor(init_scores_data, init_lod, place) test_data = paddle.batch( -paddle.reader.shuffle( -paddle.dataset.wmt14.test(dict_size), buf_size=1000), -batch_size=batch_size) + paddle.reader.shuffle( + paddle.dataset.wmt14.test(dict_size), buf_size=1000), + batch_size=batch_size) feed_order = ['src_word_id'] feed_list = [ -framework.default_main_program().global_block().var(var_name) -for var_name in feed_order + framework.default_main_program().global_block().var(var_name) + for var_name in feed_order ] feeder = fluid.DataFeeder(feed_list, place) @@ -409,27 +428,30 @@ exe = Executor(place) exe.run(framework.default_startup_program()) for data in test_data(): -feed_data = map(lambda x: [x[0]], data) -feed_dict = feeder.feed(feed_data) -feed_dict['init_ids'] = init_ids -feed_dict['init_scores'] = init_scores - -results = exe.run( -framework.default_main_program(), -feed=feed_dict, -fetch_list=[translation_ids, translation_scores], -return_numpy=False) - -result_ids = np.array(results[0]) -result_scores = np.array(results[1]) - -print("Original sentence:") -print(" ".join([src_dict[w] for w in feed_data[0][0]])) -print("Translated sentence:") -print(" ".join([trg_dict[w] for w in result_ids])) -print("Corresponding score: ", result_scores) - -break + feed_data = map(lambda x: [x[0]], data) + feed_dict = feeder.feed(feed_data) + feed_dict['init_ids'] = init_ids + feed_dict['init_scores'] = init_scores + + results = exe.run( + framework.default_main_program(), + feed=feed_dict, + fetch_list=[translation_ids, translation_scores], + return_numpy=False) + + result_ids = np.array(results[0]) + result_scores = np.array(results[1]) + + print("Original sentence:") + print(" ".join([src_dict[w] for w in feed_data[0][0][1:-1]])) + print("Translated score and sentence:") + for i in xrange(beam_size): + start_pos = result_ids_lod[1][i] + 1 + end_pos = result_ids_lod[1][i+1] + print("%d\t%.4f\t%s\n" % (i+1, result_scores[end_pos-1], + " ".join([trg_dict[w] for w in result_ids[start_pos:end_pos]]))) + + break ``` ## 总结 diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn.png deleted file mode 100644 index 9d8efd50a49d0305586f550344472ab94c93bed3..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn_en.png deleted file mode 100755 index 4b35c88fc8ea2c503473c0c15711744e784d6af6..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention.png deleted file mode 100644 index 1b355e7786d25487a3f564af758c2c52c43b4690..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention_en.png deleted file mode 100755 index 3728f782ee09d9308d02b42305027b2735467ead..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention.png deleted file mode 100644 index 28d7a15a3bd65262bde22a3f41b5aa78b46b368a..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention_en.png deleted file mode 100755 index ea8585565da1ecaf241654c278c6f9b15e283286..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder.png deleted file mode 100755 index 60aee0017de73f462e35708b1055aff8992c03e1..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder_en.png deleted file mode 100755 index 6b73798fe632e0873b35c117b86f347c8cf3116a..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru.png deleted file mode 100644 index 0cde685b84106650a4df18ce335a23e6338d3d11..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru_en.png deleted file mode 100755 index a6af429f23f0f7e82650139bbd8dcbef27a34abe..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt.png deleted file mode 100644 index bf56d73ebf297fadf522389c7b6836dd379aa097..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt_en.png deleted file mode 100755 index 557310e044b2b6687e5ea6895417ed946ac7bc11..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/index.md b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/README.cn.md similarity index 67% rename from doc/fluid/new_docs/beginners_guide/basics/recommender_system/index.md rename to doc/fluid/new_docs/beginners_guide/basics/recommender_system/README.cn.md index 09a07f3dc30abc57ab3731af054dd83491acc9a6..3174a8c6d70166619306c784db9126f20a85f4c8 100644 --- a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/index.md +++ b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/README.cn.md @@ -1,6 +1,6 @@ # 个性化推荐 -本教程源代码目录在[book/recommender_system](https://github.com/PaddlePaddle/book/tree/develop/05.recommender_system), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 +本教程源代码目录在[book/recommender_system](https://github.com/PaddlePaddle/book/tree/develop/05.recommender_system), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书),更多内容请参考本教程的[视频课堂](http://bit.baidu.com/course/detail/id/176.html)。 ## 背景介绍 @@ -36,8 +36,8 @@ Prediction Score is 4.25 YouTube是世界上最大的视频上传、分享和发现网站,YouTube推荐系统为超过10亿用户从不断增长的视频库中推荐个性化的内容。整个系统由两个神经网络组成:候选生成网络和排序网络。候选生成网络从百万量级的视频库中生成上百个候选,排序网络对候选进行打分排序,输出排名最高的数十个结果。系统结构如图1所示: -![YouTube_Overview](./image/YouTube_Overview.png)

+
图1. YouTube 推荐系统结构

@@ -45,20 +45,20 @@ YouTube是世界上最大的视频上传、分享和发现网站,YouTube推荐 候选生成网络将推荐问题建模为一个类别数极大的多类分类问题:对于一个Youtube用户,使用其观看历史(视频ID)、搜索词记录(search tokens)、人口学信息(如地理位置、用户登录设备)、二值特征(如性别,是否登录)和连续特征(如用户年龄)等,对视频库中所有视频进行多分类,得到每一类别的分类结果(即每一个视频的推荐概率),最终输出概率较高的几百个视频。 -首先,将观看历史及搜索词记录这类历史信息,映射为向量后取平均值得到定长表示;同时,输入人口学特征以优化新用户的推荐效果,并将二值特征和连续特征归一化处理到[0, 1]范围。接下来,将所有特征表示拼接为一个向量,并输入给非线形多层感知器(MLP,详见[识别数字](https://github.com/PaddlePaddle/book/blob/develop/02.recognize_digits/README.cn.md)教程)处理。最后,训练时将MLP的输出给softmax做分类,预测时计算用户的综合特征(MLP的输出)与所有视频的相似度,取得分最高的`$k$`个作为候选生成网络的筛选结果。图2显示了候选生成网络结构。 +首先,将观看历史及搜索词记录这类历史信息,映射为向量后取平均值得到定长表示;同时,输入人口学特征以优化新用户的推荐效果,并将二值特征和连续特征归一化处理到[0, 1]范围。接下来,将所有特征表示拼接为一个向量,并输入给非线形多层感知器(MLP,详见[识别数字](https://github.com/PaddlePaddle/book/blob/develop/02.recognize_digits/README.cn.md)教程)处理。最后,训练时将MLP的输出给softmax做分类,预测时计算用户的综合特征(MLP的输出)与所有视频的相似度,取得分最高的$k$个作为候选生成网络的筛选结果。图2显示了候选生成网络结构。 -![Deep_candidate_generation_model_architecture](./image/Deep_candidate_generation_model_architecture.png)

+
图2. 候选生成网络结构

-对于一个用户`$U$`,预测此刻用户要观看的视频`$\omega$`为视频`$i$`的概率公式为: +对于一个用户$U$,预测此刻用户要观看的视频$\omega$为视频$i$的概率公式为: $$P(\omega=i|u)=\frac{e^{v_{i}u}}{\sum_{j \in V}e^{v_{j}u}}$$ -其中`$u$`为用户`$U$`的特征表示,`$V$`为视频库集合,`$v_i$`为视频库中第`$i$`个视频的特征表示。`$u$`和`$v_i$`为长度相等的向量,两者点积可以通过全连接层实现。 +其中$u$为用户$U$的特征表示,$V$为视频库集合,$v_i$为视频库中第$i$个视频的特征表示。$u$和$v_i$为长度相等的向量,两者点积可以通过全连接层实现。 -考虑到softmax分类的类别数非常多,为了保证一定的计算效率:1)训练阶段,使用负样本类别采样将实际计算的类别数缩小至数千;2)推荐(预测)阶段,忽略softmax的归一化计算(不影响结果),将类别打分问题简化为点积(dot product)空间中的最近邻(nearest neighbor)搜索问题,取与`$u$`最近的`$k$`个视频作为生成的候选。 +考虑到softmax分类的类别数非常多,为了保证一定的计算效率:1)训练阶段,使用负样本类别采样将实际计算的类别数缩小至数千;2)推荐(预测)阶段,忽略softmax的归一化计算(不影响结果),将类别打分问题简化为点积(dot product)空间中的最近邻(nearest neighbor)搜索问题,取与$u$最近的$k$个视频作为生成的候选。 #### 排序网络(Ranking Network) 排序网络的结构类似于候选生成网络,但是它的目标是对候选进行更细致的打分排序。和传统广告排序中的特征抽取方法类似,这里也构造了大量的用于视频排序的相关特征(如视频 ID、上次观看时间等)。这些特征的处理方式和候选生成网络类似,不同之处是排序网络的顶部是一个加权逻辑回归(weighted logistic regression),它对所有候选视频进行打分,从高到底排序后将分数较高的一些视频返回给用户。 @@ -72,20 +72,20 @@ $$P(\omega=i|u)=\frac{e^{v_{i}u}}{\sum_{j \in V}e^{v_{j}u}}$$ 卷积神经网络主要由卷积(convolution)和池化(pooling)操作构成,其应用及组合方式灵活多变,种类繁多。本小结我们以如图3所示的网络进行讲解: -![text_cnn](./image/text_cnn.png)

+
图3. 卷积神经网络文本分类模型

-假设待处理句子的长度为`$n$`,其中第`$i$`个词的词向量(word embedding)为`$x_i\in\mathbb{R}^k$`,`$k$`为维度大小。 +假设待处理句子的长度为$n$,其中第$i$个词的词向量(word embedding)为$x_i\in\mathbb{R}^k$,$k$为维度大小。 -首先,进行词向量的拼接操作:将每`$h$`个词拼接起来形成一个大小为`$h$`的词窗口,记为`$x_{i:i+h-1}$`,它表示词序列`$x_{i},x_{i+1},\ldots,x_{i+h-1}$`的拼接,其中,`$i$`表示词窗口中第一个词在整个句子中的位置,取值范围从`$1$`到`$n-h+1$`,`$x_{i:i+h-1}\in\mathbb{R}^{hk}$`。 +首先,进行词向量的拼接操作:将每$h$个词拼接起来形成一个大小为$h$的词窗口,记为$x_{i:i+h-1}$,它表示词序列$x_{i},x_{i+1},\ldots,x_{i+h-1}$的拼接,其中,$i$表示词窗口中第一个词在整个句子中的位置,取值范围从$1$到$n-h+1$,$x_{i:i+h-1}\in\mathbb{R}^{hk}$。 -其次,进行卷积操作:把卷积核(kernel)`$w\in\mathbb{R}^{hk}$`应用于包含`$h$`个词的窗口`$x_{i:i+h-1}$`,得到特征`$c_i=f(w\cdot x_{i:i+h-1}+b)$`,其中`$b\in\mathbb{R}$`为偏置项(bias),`$f$`为非线性激活函数,如`$sigmoid$`。将卷积核应用于句子中所有的词窗口`${x_{1:h},x_{2:h+1},\ldots,x_{n-h+1:n}}$`,产生一个特征图(feature map): +其次,进行卷积操作:把卷积核(kernel)$w\in\mathbb{R}^{hk}$应用于包含$h$个词的窗口$x_{i:i+h-1}$,得到特征$c_i=f(w\cdot x_{i:i+h-1}+b)$,其中$b\in\mathbb{R}$为偏置项(bias),$f$为非线性激活函数,如$sigmoid$。将卷积核应用于句子中所有的词窗口${x_{1:h},x_{2:h+1},\ldots,x_{n-h+1:n}}$,产生一个特征图(feature map): $$c=[c_1,c_2,\ldots,c_{n-h+1}], c \in \mathbb{R}^{n-h+1}$$ -接下来,对特征图采用时间维度上的最大池化(max pooling over time)操作得到此卷积核对应的整句话的特征`$\hat c$`,它是特征图中所有元素的最大值: +接下来,对特征图采用时间维度上的最大池化(max pooling over time)操作得到此卷积核对应的整句话的特征$\hat c$,它是特征图中所有元素的最大值: $$\hat c=max(c)$$ @@ -95,9 +95,9 @@ $$\hat c=max(c)$$ 1. 首先,使用用户特征和电影特征作为神经网络的输入,其中: -- 用户特征融合了四个属性信息,分别是用户ID、性别、职业和年龄。 + - 用户特征融合了四个属性信息,分别是用户ID、性别、职业和年龄。 -- 电影特征融合了三个属性信息,分别是电影ID、电影类型ID和电影名称。 + - 电影特征融合了三个属性信息,分别是电影ID、电影类型ID和电影名称。 2. 对用户特征,将用户ID映射为维度大小为256的向量表示,输入全连接层,并对其他三个属性也做类似的处理。然后将四个属性的特征表示分别全连接并相加。 @@ -105,8 +105,9 @@ $$\hat c=max(c)$$ 4. 得到用户和电影的向量表示后,计算二者的余弦相似度作为推荐系统的打分。最后,用该相似度打分和用户真实打分的差异的平方作为该回归模型的损失函数。 -![rec_regression_network](./image/rec_regression_network.png)

+ +
图4. 融合推荐模型

@@ -141,7 +142,7 @@ movie_info = paddle.dataset.movielens.movie_info() print movie_info.values()[0] ``` - + 这表示,电影的id是1,标题是《Toy Story》,该电影被分为到三个类别中。这三个类别是动画,儿童,喜剧。 @@ -152,13 +153,14 @@ user_info = paddle.dataset.movielens.user_info() print user_info.values()[0] ``` - + 这表示,该用户ID是1,女性,年龄比18岁还年轻。职业ID是10。 其中,年龄使用下列分布 + * 1: "Under 18" * 18: "18-24" * 25: "25-34" @@ -168,6 +170,7 @@ print user_info.values()[0] * 56: "56+" 职业是从下面几种选项里面选则得出: + * 0: "other" or not specified * 1: "academic/educator" * 2: "artist" @@ -203,7 +206,7 @@ mov_id = train_sample[len(user_info[uid].value())] print "User %s rates Movie %s with Score %s"%(user_info[uid], movie_info[mov_id], train_sample[-1]) ``` -User rates Movie with Score [5.0] + User rates Movie with Score [5.0] 即用户1对电影1193的评价为5分。 @@ -214,6 +217,7 @@ User rates Movie 表格 1 电影评论情感分析

-在自然语言处理中,情感分析属于典型的**文本分类**问题,即把需要进行情感分析的文本划分为其所属类别。文本分类涉及文本表示和分类方法两个问题。在深度学习的方法出现之前,主流的文本表示方法为词袋模型BOW(bag of words),话题模型等等;分类方法有SVM(support vector machine), LR(logistic regression)等等。 +在自然语言处理中,情感分析属于典型的**文本分类**问题,即把需要进行情感分析的文本划分为其所属类别。文本分类涉及文本表示和分类方法两个问题。在深度学习的方法出现之前,主流的文本表示方法为词袋模型BOW(bag of words),话题模型等等;分类方法有SVM(support vector machine), LR(logistic regression)等等。 -对于一段文本,BOW表示会忽略其词顺序、语法和句法,将这段文本仅仅看做是一个词集合,因此BOW方法并不能充分表示文本的语义信息。例如,句子“这部电影糟糕透了”和“一个乏味,空洞,没有内涵的作品”在情感分析中具有很高的语义相似度,但是它们的BOW表示的相似度为0。又如,句子“一个空洞,没有内涵的作品”和“一个不空洞而且有内涵的作品”的BOW相似度很高,但实际上它们的意思很不一样。 +对于一段文本,BOW表示会忽略其词顺序、语法和句法,将这段文本仅仅看做是一个词集合,因此BOW方法并不能充分表示文本的语义信息。例如,句子“这部电影糟糕透了”和“一个乏味,空洞,没有内涵的作品”在情感分析中具有很高的语义相似度,但是它们的BOW表示的相似度为0。又如,句子“一个空洞,没有内涵的作品”和“一个不空洞而且有内涵的作品”的BOW相似度很高,但实际上它们的意思很不一样。 本章我们所要介绍的深度学习模型克服了BOW表示的上述缺陷,它在考虑词顺序的基础上把文本映射到低维度的语义空间,并且以端对端(end to end)的方式进行文本表示及分类,其性能相对于传统方法有显著的提升\[[1](#参考文献)\]。 @@ -36,54 +36,54 @@ 循环神经网络是一种能对序列数据进行精确建模的有力工具。实际上,循环神经网络的理论计算能力是图灵完备的\[[4](#参考文献)\]。自然语言是一种典型的序列数据(词序列),近年来,循环神经网络及其变体(如long short term memory\[[5](#参考文献)\]等)在自然语言处理的多个领域,如语言模型、句法解析、语义角色标注(或一般的序列标注)、语义表示、图文生成、对话、机器翻译等任务上均表现优异甚至成为目前效果最好的方法。 -![rnn](./image/rnn.png)

+
图1. 循环神经网络按时间展开的示意图

-循环神经网络按时间展开后如图1所示:在第`$t$`时刻,网络读入第`$t$`个输入`$x_t$`(向量表示)及前一时刻隐层的状态值`$h_{t-1}$`(向量表示,`$h_0$`一般初始化为`$0$`向量),计算得出本时刻隐层的状态值`$h_t$`,重复这一步骤直至读完所有输入。如果将循环神经网络所表示的函数记为`$f$`,则其公式可表示为: +循环神经网络按时间展开后如图1所示:在第$t$时刻,网络读入第$t$个输入$x_t$(向量表示)及前一时刻隐层的状态值$h_{t-1}$(向量表示,$h_0$一般初始化为$0$向量),计算得出本时刻隐层的状态值$h_t$,重复这一步骤直至读完所有输入。如果将循环神经网络所表示的函数记为$f$,则其公式可表示为: $$h_t=f(x_t,h_{t-1})=\sigma(W_{xh}x_t+W_{hh}h_{t-1}+b_h)$$ -其中`$W_{xh}$`是输入到隐层的矩阵参数,`$W_{hh}$`是隐层到隐层的矩阵参数,`$b_h$`为隐层的偏置向量(bias)参数,`$\sigma$`为`$sigmoid$`函数。 +其中$W_{xh}$是输入到隐层的矩阵参数,$W_{hh}$是隐层到隐层的矩阵参数,$b_h$为隐层的偏置向量(bias)参数,$\sigma$为$sigmoid$函数。 -在处理自然语言时,一般会先将词(one-hot表示)映射为其词向量(word embedding)表示,然后再作为循环神经网络每一时刻的输入`$x_t$`。此外,可以根据实际需要的不同在循环神经网络的隐层上连接其它层。如,可以把一个循环神经网络的隐层输出连接至下一个循环神经网络的输入构建深层(deep or stacked)循环神经网络,或者提取最后一个时刻的隐层状态作为句子表示进而使用分类模型等等。 +在处理自然语言时,一般会先将词(one-hot表示)映射为其词向量(word embedding)表示,然后再作为循环神经网络每一时刻的输入$x_t$。此外,可以根据实际需要的不同在循环神经网络的隐层上连接其它层。如,可以把一个循环神经网络的隐层输出连接至下一个循环神经网络的输入构建深层(deep or stacked)循环神经网络,或者提取最后一个时刻的隐层状态作为句子表示进而使用分类模型等等。 ### 长短期记忆网络(LSTM) -对于较长的序列数据,循环神经网络的训练过程中容易出现梯度消失或爆炸现象\[[6](#参考文献)\]。为了解决这一问题,Hochreiter S, Schmidhuber J. (1997)提出了LSTM(long short term memory\[[5](#参考文献)\])。 +对于较长的序列数据,循环神经网络的训练过程中容易出现梯度消失或爆炸现象\[[6](#参考文献)\]。为了解决这一问题,Hochreiter S, Schmidhuber J. (1997)提出了LSTM(long short term memory\[[5](#参考文献)\])。 -相比于简单的循环神经网络,LSTM增加了记忆单元`$c$`、输入门`$i$`、遗忘门`$f$`及输出门`$o$`。这些门及记忆单元组合起来大大提升了循环神经网络处理长序列数据的能力。若将基于LSTM的循环神经网络表示的函数记为`$F$`,则其公式为: +相比于简单的循环神经网络,LSTM增加了记忆单元$c$、输入门$i$、遗忘门$f$及输出门$o$。这些门及记忆单元组合起来大大提升了循环神经网络处理长序列数据的能力。若将基于LSTM的循环神经网络表示的函数记为$F$,则其公式为: $$ h_t=F(x_t,h_{t-1})$$ -`$F$`由下列公式组合而成\[[7](#参考文献)\]: +$F$由下列公式组合而成\[[7](#参考文献)\]: $$ i_t = \sigma{(W_{xi}x_t+W_{hi}h_{t-1}+W_{ci}c_{t-1}+b_i)} $$ $$ f_t = \sigma(W_{xf}x_t+W_{hf}h_{t-1}+W_{cf}c_{t-1}+b_f) $$ $$ c_t = f_t\odot c_{t-1}+i_t\odot tanh(W_{xc}x_t+W_{hc}h_{t-1}+b_c) $$ $$ o_t = \sigma(W_{xo}x_t+W_{ho}h_{t-1}+W_{co}c_{t}+b_o) $$ $$ h_t = o_t\odot tanh(c_t) $$ -其中,`$i_t, f_t, c_t, o_t$`分别表示输入门,遗忘门,记忆单元及输出门的向量值,带角标的`$W$`及`$b$`为模型参数,`$tanh$`为双曲正切函数,`$\odot$`表示逐元素(elementwise)的乘法操作。输入门控制着新输入进入记忆单元`$c$`的强度,遗忘门控制着记忆单元维持上一时刻值的强度,输出门控制着输出记忆单元的强度。三种门的计算方式类似,但有着完全不同的参数,它们各自以不同的方式控制着记忆单元`$c$`,如图2所示: +其中,$i_t, f_t, c_t, o_t$分别表示输入门,遗忘门,记忆单元及输出门的向量值,带角标的$W$及$b$为模型参数,$tanh$为双曲正切函数,$\odot$表示逐元素(elementwise)的乘法操作。输入门控制着新输入进入记忆单元$c$的强度,遗忘门控制着记忆单元维持上一时刻值的强度,输出门控制着输出记忆单元的强度。三种门的计算方式类似,但有着完全不同的参数,它们各自以不同的方式控制着记忆单元$c$,如图2所示: -![lstm](./image/lstm.png)

-图2. 时刻`$t$`的LSTM [7] +
+图2. 时刻$t$的LSTM [7]

LSTM通过给简单的循环神经网络增加记忆及控制门的方式,增强了其处理远距离依赖问题的能力。类似原理的改进还有Gated Recurrent Unit (GRU)\[[8](#参考文献)\],其设计更为简洁一些。**这些改进虽然各有不同,但是它们的宏观描述却与简单的循环神经网络一样(如图2所示),即隐状态依据当前输入及前一时刻的隐状态来改变,不断地循环这一过程直至输入处理完毕:** $$ h_t=Recrurent(x_t,h_{t-1})$$ -其中,`$Recrurent$`可以表示简单的循环神经网络、GRU或LSTM。 +其中,$Recrurent$可以表示简单的循环神经网络、GRU或LSTM。 ### 栈式双向LSTM(Stacked Bidirectional LSTM) -对于正常顺序的循环神经网络,`$h_t$`包含了`$t$`时刻之前的输入信息,也就是上文信息。同样,为了得到下文信息,我们可以使用反方向(将输入逆序处理)的循环神经网络。结合构建深层循环神经网络的方法(深层神经网络往往能得到更抽象和高级的特征表示),我们可以通过构建更加强有力的基于LSTM的栈式双向循环神经网络\[[9](#参考文献)\],来对时序数据进行建模。 +对于正常顺序的循环神经网络,$h_t$包含了$t$时刻之前的输入信息,也就是上文信息。同样,为了得到下文信息,我们可以使用反方向(将输入逆序处理)的循环神经网络。结合构建深层循环神经网络的方法(深层神经网络往往能得到更抽象和高级的特征表示),我们可以通过构建更加强有力的基于LSTM的栈式双向循环神经网络\[[9](#参考文献)\],来对时序数据进行建模。 如图3所示(以三层为例),奇数层LSTM正向,偶数层LSTM反向,高一层的LSTM使用低一层LSTM及之前所有层的信息作为输入,对最高层LSTM序列使用时间维度上的最大池化即可得到文本的定长向量表示(这一表示充分融合了文本的上下文信息,并且对文本进行了深层次抽象),最后我们将文本表示连接至softmax构建分类模型。 -![stacked_lstm](./image/stacked_lstm.jpg)

+
图3. 栈式双向LSTM用于文本分类

@@ -94,11 +94,11 @@ $$ h_t=Recrurent(x_t,h_{t-1})$$ ```text aclImdb |- test -|-- neg -|-- pos + |-- neg + |-- pos |- train -|-- neg -|-- pos + |-- neg + |-- pos ``` Paddle在`dataset/imdb.py`中提实现了imdb数据集的自动下载和读取,并提供了读取字典、训练数据、测试数据等API。 @@ -107,6 +107,7 @@ Paddle在`dataset/imdb.py`中提实现了imdb数据集的自动下载和读取 在该示例中,我们实现了两种文本分类算法,分别基于[推荐系统](https://github.com/PaddlePaddle/book/tree/develop/05.recommender_system)一节介绍过的文本卷积神经网络,以及[栈式双向LSTM](#栈式双向LSTM(Stacked Bidirectional LSTM))。我们首先引入要用到的库和定义全局变量: ```python +from __future__ import print_function import paddle import paddle.fluid as fluid from functools import partial @@ -115,6 +116,7 @@ import numpy as np CLASS_DIM = 2 EMB_DIM = 128 HID_DIM = 512 +STACKED_NUM = 3 BATCH_SIZE = 128 USE_GPU = False ``` @@ -126,23 +128,23 @@ USE_GPU = False ```python def convolution_net(data, input_dim, class_dim, emb_dim, hid_dim): -emb = fluid.layers.embedding( -input=data, size=[input_dim, emb_dim], is_sparse=True) -conv_3 = fluid.nets.sequence_conv_pool( -input=emb, -num_filters=hid_dim, -filter_size=3, -act="tanh", -pool_type="sqrt") -conv_4 = fluid.nets.sequence_conv_pool( -input=emb, -num_filters=hid_dim, -filter_size=4, -act="tanh", -pool_type="sqrt") -prediction = fluid.layers.fc( -input=[conv_3, conv_4], size=class_dim, act="softmax") -return prediction + emb = fluid.layers.embedding( + input=data, size=[input_dim, emb_dim], is_sparse=True) + conv_3 = fluid.nets.sequence_conv_pool( + input=emb, + num_filters=hid_dim, + filter_size=3, + act="tanh", + pool_type="sqrt") + conv_4 = fluid.nets.sequence_conv_pool( + input=emb, + num_filters=hid_dim, + filter_size=4, + act="tanh", + pool_type="sqrt") + prediction = fluid.layers.fc( + input=[conv_3, conv_4], size=class_dim, act="softmax") + return prediction ``` 网络的输入`input_dim`表示的是词典的大小,`class_dim`表示类别数。这里,我们使用[`sequence_conv_pool`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/trainer_config_helpers/networks.py) API实现了卷积和池化操作。 @@ -154,27 +156,26 @@ return prediction ```python def stacked_lstm_net(data, input_dim, class_dim, emb_dim, hid_dim, stacked_num): -emb = fluid.layers.embedding( -input=data, size=[input_dim, emb_dim], is_sparse=True) + emb = fluid.layers.embedding( + input=data, size=[input_dim, emb_dim], is_sparse=True) -fc1 = fluid.layers.fc(input=emb, size=hid_dim) -lstm1, cell1 = fluid.layers.dynamic_lstm(input=fc1, size=hid_dim) + fc1 = fluid.layers.fc(input=emb, size=hid_dim) + lstm1, cell1 = fluid.layers.dynamic_lstm(input=fc1, size=hid_dim) -inputs = [fc1, lstm1] + inputs = [fc1, lstm1] -for i in range(2, stacked_num + 1): -fc = fluid.layers.fc(input=inputs, size=hid_dim) -lstm, cell = fluid.layers.dynamic_lstm( -input=fc, size=hid_dim, is_reverse=(i % 2) == 0) -inputs = [fc, lstm] + for i in range(2, stacked_num + 1): + fc = fluid.layers.fc(input=inputs, size=hid_dim) + lstm, cell = fluid.layers.dynamic_lstm( + input=fc, size=hid_dim, is_reverse=(i % 2) == 0) + inputs = [fc, lstm] -fc_last = fluid.layers.sequence_pool(input=inputs[0], pool_type='max') -lstm_last = fluid.layers.sequence_pool(input=inputs[1], pool_type='max') + fc_last = fluid.layers.sequence_pool(input=inputs[0], pool_type='max') + lstm_last = fluid.layers.sequence_pool(input=inputs[1], pool_type='max') -prediction = fluid.layers.fc(input=[fc_last, lstm_last], -size=class_dim, -act='softmax') -return prediction + prediction = fluid.layers.fc( + input=[fc_last, lstm_last], size=class_dim, act='softmax') + return prediction ``` 以上的栈式双向LSTM抽象出了高级特征并把其映射到和分类类别数同样大小的向量上。`paddle.activation.Softmax`函数用来计算分类属于某个类别的概率。 @@ -184,12 +185,13 @@ return prediction ```python def inference_program(word_dict): -data = fluid.layers.data( -name="words", shape=[1], dtype="int64", lod_level=1) + data = fluid.layers.data( + name="words", shape=[1], dtype="int64", lod_level=1) -dict_dim = len(word_dict) -net = convolution_net(data, dict_dim, CLASS_DIM, EMB_DIM, HID_DIM) -return net + dict_dim = len(word_dict) + net = convolution_net(data, dict_dim, CLASS_DIM, EMB_DIM, HID_DIM) + # net = stacked_lstm_net(data, dict_dim, CLASS_DIM, EMB_DIM, HID_DIM, STACKED_NUM) + return net ``` 我们这里定义了`training_program`。它使用了从`inference_program`返回的结果来计算误差。我们同时定义了优化函数`optimizer_func`。 @@ -200,16 +202,16 @@ return net ```python def train_program(word_dict): -prediction = inference_program(word_dict) -label = fluid.layers.data(name="label", shape=[1], dtype="int64") -cost = fluid.layers.cross_entropy(input=prediction, label=label) -avg_cost = fluid.layers.mean(cost) -accuracy = fluid.layers.accuracy(input=prediction, label=label) -return [avg_cost, accuracy] + prediction = inference_program(word_dict) + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + cost = fluid.layers.cross_entropy(input=prediction, label=label) + avg_cost = fluid.layers.mean(cost) + accuracy = fluid.layers.accuracy(input=prediction, label=label) + return [avg_cost, accuracy] def optimizer_func(): -return fluid.optimizer.Adagrad(learning_rate=0.002) + return fluid.optimizer.Adagrad(learning_rate=0.002) ``` ## 训练模型 @@ -236,9 +238,9 @@ word_dict = paddle.dataset.imdb.word_dict() print ("Reading training data....") train_reader = paddle.batch( -paddle.reader.shuffle( -paddle.dataset.imdb.train(word_dict), buf_size=25000), -batch_size=BATCH_SIZE) + paddle.reader.shuffle( + paddle.dataset.imdb.train(word_dict), buf_size=25000), + batch_size=BATCH_SIZE) ``` ### 构造训练器(trainer) @@ -246,9 +248,9 @@ batch_size=BATCH_SIZE) ```python trainer = fluid.Trainer( -train_func=partial(train_program, word_dict), -place=place, -optimizer_func=optimizer_func) + train_func=partial(train_program, word_dict), + place=place, + optimizer_func=optimizer_func) ``` ### 提供数据 @@ -268,13 +270,13 @@ feed_order = ['words', 'label'] params_dirname = "understand_sentiment_conv.inference.model" def event_handler(event): -if isinstance(event, fluid.EndStepEvent): -print("Step {0}, Epoch {1} Metrics {2}".format( -event.step, event.epoch, map(np.array, event.metrics))) + if isinstance(event, fluid.EndStepEvent): + print("Step {0}, Epoch {1} Metrics {2}".format( + event.step, event.epoch, map(np.array, event.metrics))) -if event.step == 10: -trainer.save_params(params_dirname) -trainer.stop() + if event.step == 10: + trainer.save_params(params_dirname) + trainer.stop() ``` ### 开始训练 @@ -283,10 +285,10 @@ trainer.stop() ```python trainer.train( -num_epochs=1, -event_handler=event_handler, -reader=train_reader, -feed_order=feed_order) + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=feed_order) ``` ## 应用模型 @@ -297,7 +299,7 @@ feed_order=feed_order) ```python inferencer = fluid.Inferencer( -inference_program, param_path=params_dirname, place=place) + infer_func=partial(inference_program, word_dict), param_path=params_dirname, place=place) ``` ### 生成测试用输入数据 @@ -307,14 +309,14 @@ inference_program, param_path=params_dirname, place=place) ```python reviews_str = [ -'read the book forget the movie', 'this is a great movie', 'this is very bad' + 'read the book forget the movie', 'this is a great movie', 'this is very bad' ] reviews = [c.split() for c in reviews_str] UNK = word_dict[''] lod = [] for c in reviews: -lod.append([word_dict.get(words, UNK) for words in c]) + lod.append([word_dict.get(words, UNK) for words in c]) base_shape = [[len(c) for c in lod]] @@ -329,7 +331,7 @@ tensor_words = fluid.create_lod_tensor(lod, base_shape, place) results = inferencer.infer({'words': tensor_words}) for i, r in enumerate(results[0]): -print("Predict probability of ", r[0], " to be positive and ", r[1], " to be negative for review \'", reviews_str[i], "\'") + print("Predict probability of ", r[0], " to be positive and ", r[1], " to be negative for review \'", reviews_str[i], "\'") ``` diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm.png b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm.png deleted file mode 100644 index 98fbea413a98a619004ca669c67f5f867fe974c9..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm_en.png b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm_en.png deleted file mode 100755 index d73a00bf2c1fca2f9b8c26bccf5ea844fa1db50b..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/rnn.png b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/rnn.png deleted file mode 100755 index 26c904102a6e6c4e30f0048b81373ae8c148b355..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/rnn.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm.jpg b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm.jpg deleted file mode 100644 index 6b2adf70f2b5112a2e82505da5cff9f5fd0c6298..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm.jpg and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm_en.png b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm_en.png deleted file mode 100755 index 8b5dbd726178b5555c513294e7b10a81acc96ff5..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/index.md b/doc/fluid/new_docs/beginners_guide/basics/word2vec/README.cn.md similarity index 56% rename from doc/fluid/new_docs/beginners_guide/basics/word2vec/index.md rename to doc/fluid/new_docs/beginners_guide/basics/word2vec/README.cn.md index e73a6334ca1acd49379604f24d3d4e463192a902..0d2d4dbc2caf8f5163d6401b6eec1a7f26e2a12e 100644 --- a/doc/fluid/new_docs/beginners_guide/basics/word2vec/index.md +++ b/doc/fluid/new_docs/beginners_guide/basics/word2vec/README.cn.md @@ -1,7 +1,7 @@ # 词向量 -本教程源代码目录在[book/word2vec](https://github.com/PaddlePaddle/book/tree/develop/04.word2vec), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 +本教程源代码目录在[book/word2vec](https://github.com/PaddlePaddle/book/tree/develop/04.word2vec), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书),更多内容请参考本教程的[视频课堂](http://bit.baidu.com/course/detail/id/175.html)。 ## 背景介绍 @@ -12,17 +12,19 @@ One-hot vector虽然自然,但是用处有限。比如,在互联网广告系统里,如果用户输入的query是“母亲节”,而有一个广告的关键词是“康乃馨”。虽然按照常理,我们知道这两个词之间是有联系的——母亲节通常应该送给母亲一束康乃馨;但是这两个词对应的one-hot vectors之间的距离度量,无论是欧氏距离还是余弦相似度(cosine similarity),由于其向量正交,都认为这两个词毫无相关性。 得出这种与我们相悖的结论的根本原因是:每个词本身的信息量都太小。所以,仅仅给定两个词,不足以让我们准确判别它们是否相关。要想精确计算相关性,我们还需要更多的信息——从大量数据里通过机器学习方法归纳出来的知识。 -在机器学习领域里,各种“知识”被各种模型表示,词向量模型(word embedding model)就是其中的一类。通过词向量模型可将一个 one-hot vector映射到一个维度更低的实数向量(embedding vector),如`$embedding(Mother's\ Day) = [0.3, 4.2, -1.5, ...], embedding(Carnation) = [0.2, 5.6, -2.3, ...]$`。在这个映射到的实数向量表示中,希望两个语义(或用法)上相似的词对应的词向量“更像”,这样如“母亲节”和“康乃馨”的对应词向量的余弦相似度就不再为零了。 +在机器学习领域里,各种“知识”被各种模型表示,词向量模型(word embedding model)就是其中的一类。通过词向量模型可将一个 one-hot vector映射到一个维度更低的实数向量(embedding vector),如$embedding(母亲节) = [0.3, 4.2, -1.5, ...], embedding(康乃馨) = [0.2, 5.6, -2.3, ...]$。在这个映射到的实数向量表示中,希望两个语义(或用法)上相似的词对应的词向量“更像”,这样如“母亲节”和“康乃馨”的对应词向量的余弦相似度就不再为零了。 -词向量模型可以是概率模型、共生矩阵(co-occurrence matrix)模型或神经元网络模型。在用神经网络求词向量之前,传统做法是统计一个词语的共生矩阵`$X$`。`$X$`是一个`$|V| \times |V|$` 大小的矩阵,`$X_{ij}$`表示在所有语料中,词汇表`V`(vocabulary)中第i个词和第j个词同时出现的词数,`$|V|$`为词汇表的大小。对`$X$`做矩阵分解(如奇异值分解,Singular Value Decomposition \[[5](#参考文献)\]),得到的`$U$`即视为所有词的词向量: +词向量模型可以是概率模型、共生矩阵(co-occurrence matrix)模型或神经元网络模型。在用神经网络求词向量之前,传统做法是统计一个词语的共生矩阵$X$。$X$是一个$|V| \times |V|$ 大小的矩阵,$X_{ij}$表示在所有语料中,词汇表`V`(vocabulary)中第i个词和第j个词同时出现的词数,$|V|$为词汇表的大小。对$X$做矩阵分解(如奇异值分解,Singular Value Decomposition \[[5](#参考文献)\]),得到的$U$即视为所有词的词向量: $$X = USV^T$$ -但这样的传统做法有很多问题:
-1) 由于很多词没有出现,导致矩阵极其稀疏,因此需要对词频做额外处理来达到好的矩阵分解效果;
-2) 矩阵非常大,维度太高(通常达到`$10^6*10^6$`的数量级);
-3) 需要手动去掉停用词(如although, a,...),不然这些频繁出现的词也会影响矩阵分解的效果。 +但这样的传统做法有很多问题: + +1) 由于很多词没有出现,导致矩阵极其稀疏,因此需要对词频做额外处理来达到好的矩阵分解效果; +2) 矩阵非常大,维度太高(通常达到$10^6 \times 10^6$的数量级); + +3) 需要手动去掉停用词(如although, a,...),不然这些频繁出现的词也会影响矩阵分解的效果。 基于神经网络的模型不需要计算存储一个在全语料上统计的大表,而是通过学习语义信息得到词向量,因此能很好地解决以上问题。在本章里,我们将展示基于神经网络训练词向量的细节,以及如何用PaddlePaddle训练一个词向量模型。 @@ -31,19 +33,21 @@ $$X = USV^T$$ 本章中,当词向量训练好后,我们可以用数据可视化算法t-SNE\[[4](#参考文献)\]画出词语特征在二维上的投影(如下图所示)。从图中可以看出,语义相关的词语(如a, the, these; big, huge)在投影上距离很近,语意无关的词(如say, business; decision, japan)在投影上的距离很远。 -![2d_similarity](./image/2d_similarity.png)

-图1. 词向量的二维投影 +
+ 图1. 词向量的二维投影

-另一方面,我们知道两个向量的余弦值在`$[-1,1]$`的区间内:两个完全相同的向量余弦值为1, 两个相互垂直的向量之间余弦值为0,两个方向完全相反的向量余弦值为-1,即相关性和余弦值大小成正比。因此我们还可以计算两个词向量的余弦相似度: +另一方面,我们知道两个向量的余弦值在$[-1,1]$的区间内:两个完全相同的向量余弦值为1, 两个相互垂直的向量之间余弦值为0,两个方向完全相反的向量余弦值为-1,即相关性和余弦值大小成正比。因此我们还可以计算两个词向量的余弦相似度: ``` -similarity: 0.899180685161 + please input two words: big huge +similarity: 0.899180685161 please input two words: from company similarity: -0.0997506977351 + ``` 以上结果可以通过运行`calculate_dis.py`, 加载字典里的单词和对应训练特征结果得到,我们将在[应用模型](#应用模型)中详细描述用法。 @@ -56,10 +60,10 @@ similarity: -0.0997506977351 ### 语言模型 在介绍词向量模型之前,我们先来引入一个概念:语言模型。 -语言模型旨在为语句的联合概率函数`$P(w_1, ..., w_T)$`建模, 其中`$w_i$`表示句子中的第i个词。语言模型的目标是,希望模型对有意义的句子赋予大概率,对没意义的句子赋予小概率。 +语言模型旨在为语句的联合概率函数$P(w_1, ..., w_T)$建模, 其中$w_i$表示句子中的第i个词。语言模型的目标是,希望模型对有意义的句子赋予大概率,对没意义的句子赋予小概率。 这样的模型可以应用于很多领域,如机器翻译、语音识别、信息检索、词性标注、手写识别等,它们都希望能得到一个连续序列的概率。 以信息检索为例,当你在搜索“how long is a football bame”时(bame是一个医学名词),搜索引擎会提示你是否希望搜索"how long is a football game", 这是因为根据语言模型计算出“how long is a football bame”的概率很低,而与bame近似的,可能引起错误的词中,game会使该句生成的概率最大。 -对语言模型的目标概率`$P(w_1, ..., w_T)$`,如果假设文本中每个词都是相互独立的,则整句话的联合概率可以表示为其中所有词语条件概率的乘积,即: +对语言模型的目标概率$P(w_1, ..., w_T)$,如果假设文本中每个词都是相互独立的,则整句话的联合概率可以表示为其中所有词语条件概率的乘积,即: $$P(w_1, ..., w_T) = \prod_{t=1}^TP(w_t)$$ @@ -75,7 +79,7 @@ $$P(w_1, ..., w_T) = \prod_{t=1}^TP(w_t | w_1, ... , w_{t-1})$$ Yoshua Bengio等科学家就于2003年在著名论文 Neural Probabilistic Language Models \[[1](#参考文献)\] 中介绍如何学习一个神经元网络表示的词向量模型。文中的神经概率语言模型(Neural Network Language Model,NNLM)通过一个线性映射和一个非线性隐层连接,同时学习了语言模型和词向量,即通过学习大量语料得到词语的向量表达,通过这些向量得到整个句子的概率。用这种方法学习语言模型可以克服维度灾难(curse of dimensionality),即训练和测试数据不同导致的模型不准。注意:由于“神经概率语言模型”说法较为泛泛,我们在这里不用其NNLM的本名,考虑到其具体做法,本文中称该模型为N-gram neural model。 -我们在上文中已经讲到用条件概率建模语言模型,即一句话中第`$t$`个词的概率和该句话的前`$t-1$`个词相关。可实际上越远的词语其实对该词的影响越小,那么如果考虑一个n-gram, 每个词都只受其前面`n-1`个词的影响,则有: +我们在上文中已经讲到用条件概率建模语言模型,即一句话中第$t$个词的概率和该句话的前$t-1$个词相关。可实际上越远的词语其实对该词的影响越小,那么如果考虑一个n-gram, 每个词都只受其前面`n-1`个词的影响,则有: $$P(w_1, ..., w_T) = \prod_{t=n}^TP(w_t|w_{t-1}, w_{t-2}, ..., w_{t-n+1})$$ @@ -83,33 +87,33 @@ $$P(w_1, ..., w_T) = \prod_{t=n}^TP(w_t|w_{t-1}, w_{t-2}, ..., w_{t-n+1})$$ $$\frac{1}{T}\sum_t f(w_t, w_{t-1}, ..., w_{t-n+1};\theta) + R(\theta)$$ -其中`$f(w_t, w_{t-1}, ..., w_{t-n+1})$`表示根据历史n-1个词得到当前词`$w_t$`的条件概率,`$R(\theta)$`表示参数正则项。 +其中$f(w_t, w_{t-1}, ..., w_{t-n+1})$表示根据历史n-1个词得到当前词$w_t$的条件概率,$R(\theta)$表示参数正则项。 -![nnlm](./image/nnlm.png)

-图2. N-gram神经网络模型 +
+ 图2. N-gram神经网络模型

图2展示了N-gram神经网络模型,从下往上看,该模型分为以下几个部分: -- 对于每个样本,模型输入`$w_{t-n+1},...w_{t-1}$`, 输出句子第t个词为字典中`|V|`个词的概率。 + - 对于每个样本,模型输入$w_{t-n+1},...w_{t-1}$, 输出句子第t个词为字典中`|V|`个词的概率。 -每个输入词`$w_{t-n+1},...w_{t-1}$`首先通过映射矩阵映射到词向量`$C(w_{t-n+1}),...C(w_{t-1})$`。 + 每个输入词$w_{t-n+1},...w_{t-1}$首先通过映射矩阵映射到词向量$C(w_{t-n+1}),...C(w_{t-1})$。 -- 然后所有词语的词向量连接成一个大向量,并经过一个非线性映射得到历史词语的隐层表示: + - 然后所有词语的词向量连接成一个大向量,并经过一个非线性映射得到历史词语的隐层表示: -$$g=Utanh(\theta^Tx + b_1) + Wx + b_2$$ + $$g=Utanh(\theta^Tx + b_1) + Wx + b_2$$ -其中,`$x$`为所有词语的词向量连接成的大向量,表示文本历史特征;`$\theta$`、`$U$`、`$b_1$`、`$b_2$`和`$W$`分别为词向量层到隐层连接的参数。`$g$`表示未经归一化的所有输出单词概率,`$g_i$`表示未经归一化的字典中第`$i$`个单词的输出概率。 + 其中,$x$为所有词语的词向量连接成的大向量,表示文本历史特征;$\theta$、$U$、$b_1$、$b_2$和$W$分别为词向量层到隐层连接的参数。$g$表示未经归一化的所有输出单词概率,$g_i$表示未经归一化的字典中第$i$个单词的输出概率。 -- 根据softmax的定义,通过归一化`$g_i$`, 生成目标词`$w_t$`的概率为: + - 根据softmax的定义,通过归一化$g_i$, 生成目标词$w_t$的概率为: -$$P(w_t | w_1, ..., w_{t-n+1}) = \frac{e^{g_{w_t}}}{\sum_i^{|V|} e^{g_i}}$$ + $$P(w_t | w_1, ..., w_{t-n+1}) = \frac{e^{g_{w_t}}}{\sum_i^{|V|} e^{g_i}}$$ -- 整个网络的损失值(cost)为多类分类交叉熵,用公式表示为 + - 整个网络的损失值(cost)为多类分类交叉熵,用公式表示为 -$$J(\theta) = -\sum_{i=1}^N\sum_{c=1}^{|V|}y_k^{i}log(softmax(g_k^i))$$ + $$J(\theta) = -\sum_{i=1}^N\sum_{c=1}^{|V|}y_k^{i}log(softmax(g_k^i))$$ -其中`$y_k^i$`表示第`$i$`个样本第`$k$`类的真实标签(0或1),`$softmax(g_k^i)$`表示第i个样本第k类softmax输出的概率。 + 其中$y_k^i$表示第$i$个样本第$k$类的真实标签(0或1),$softmax(g_k^i)$表示第i个样本第k类softmax输出的概率。 @@ -117,27 +121,27 @@ $$J(\theta) = -\sum_{i=1}^N\sum_{c=1}^{|V|}y_k^{i}log(softmax(g_k^i))$$ CBOW模型通过一个词的上下文(各N个词)预测当前词。当N=2时,模型如下图所示: -![cbow](./image/cbow.png)

-图3. CBOW模型 +
+ 图3. CBOW模型

具体来说,不考虑上下文的词语输入顺序,CBOW是用上下文词语的词向量的均值来预测当前词。即: $$context = \frac{x_{t-1} + x_{t-2} + x_{t+1} + x_{t+2}}{4}$$ -其中`$x_t$`为第`$t$`个词的词向量,分类分数(score)向量 `$z=U*context$`,最终的分类`$y$`采用softmax,损失函数采用多类分类交叉熵。 +其中$x_t$为第$t$个词的词向量,分类分数(score)向量 $z=U*context$,最终的分类$y$采用softmax,损失函数采用多类分类交叉熵。 ### Skip-gram model CBOW的好处是对上下文词语的分布在词向量上进行了平滑,去掉了噪声,因此在小数据集上很有效。而Skip-gram的方法中,用一个词预测其上下文,得到了当前词上下文的很多样本,因此可用于更大的数据集。 -![skipgram](./image/skipgram.png)

-图4. Skip-gram模型 +
+ 图4. Skip-gram模型

-如上图所示,Skip-gram模型的具体做法是,将一个词的词向量映射到`$2n$`个词的词向量(`$2n$`表示当前输入词的前后各`$n$`个词),然后分别通过softmax得到这`$2n$`个词的分类损失值之和。 +如上图所示,Skip-gram模型的具体做法是,将一个词的词向量映射到$2n$个词的词向量($2n$表示当前输入词的前后各$n$个词),然后分别通过softmax得到这$2n$个词的分类损失值之和。 ## 数据准备 @@ -148,21 +152,21 @@ CBOW的好处是对上下文词语的分布在词向量上进行了平滑,去

- - - - - - - - - - - - - - - + + + + + + + + + + + + + + +
训练数据验证数据测试数据
ptb.train.txtptb.valid.txtptb.test.txt
42068句3370句3761句
训练数据验证数据测试数据
ptb.train.txtptb.valid.txtptb.test.txt
42068句3370句3761句

@@ -189,9 +193,9 @@ dream that one day 本配置的模型结构如下图所示: -![ngram](./image/ngram.png)

-图5. 模型配置中的N-gram神经网络模型 +
+ 图5. 模型配置中的N-gram神经网络模型

首先,加载所需要的包: @@ -204,6 +208,7 @@ from functools import partial import math import os import sys +from __future__ import print_function ``` 然后,定义参数: @@ -226,57 +231,57 @@ dict_size = len(word_dict) ```python def inference_program(is_sparse): -first_word = fluid.layers.data(name='firstw', shape=[1], dtype='int64') -second_word = fluid.layers.data(name='secondw', shape=[1], dtype='int64') -third_word = fluid.layers.data(name='thirdw', shape=[1], dtype='int64') -fourth_word = fluid.layers.data(name='fourthw', shape=[1], dtype='int64') - -embed_first = fluid.layers.embedding( -input=first_word, -size=[dict_size, EMBED_SIZE], -dtype='float32', -is_sparse=is_sparse, -param_attr='shared_w') -embed_second = fluid.layers.embedding( -input=second_word, -size=[dict_size, EMBED_SIZE], -dtype='float32', -is_sparse=is_sparse, -param_attr='shared_w') -embed_third = fluid.layers.embedding( -input=third_word, -size=[dict_size, EMBED_SIZE], -dtype='float32', -is_sparse=is_sparse, -param_attr='shared_w') -embed_fourth = fluid.layers.embedding( -input=fourth_word, -size=[dict_size, EMBED_SIZE], -dtype='float32', -is_sparse=is_sparse, -param_attr='shared_w') - -concat_embed = fluid.layers.concat( -input=[embed_first, embed_second, embed_third, embed_fourth], axis=1) -hidden1 = fluid.layers.fc(input=concat_embed, -size=HIDDEN_SIZE, -act='sigmoid') -predict_word = fluid.layers.fc(input=hidden1, size=dict_size, act='softmax') -return predict_word + first_word = fluid.layers.data(name='firstw', shape=[1], dtype='int64') + second_word = fluid.layers.data(name='secondw', shape=[1], dtype='int64') + third_word = fluid.layers.data(name='thirdw', shape=[1], dtype='int64') + fourth_word = fluid.layers.data(name='fourthw', shape=[1], dtype='int64') + + embed_first = fluid.layers.embedding( + input=first_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + embed_second = fluid.layers.embedding( + input=second_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + embed_third = fluid.layers.embedding( + input=third_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + embed_fourth = fluid.layers.embedding( + input=fourth_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + + concat_embed = fluid.layers.concat( + input=[embed_first, embed_second, embed_third, embed_fourth], axis=1) + hidden1 = fluid.layers.fc(input=concat_embed, + size=HIDDEN_SIZE, + act='sigmoid') + predict_word = fluid.layers.fc(input=hidden1, size=dict_size, act='softmax') + return predict_word ``` - 基于以上的神经网络结构,我们可以如下定义我们的`训练`方法 ```python def train_program(is_sparse): -# The declaration of 'next_word' must be after the invoking of inference_program, -# or the data input order of train program would be [next_word, firstw, secondw, -# thirdw, fourthw], which is not correct. -predict_word = inference_program(is_sparse) -next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64') -cost = fluid.layers.cross_entropy(input=predict_word, label=next_word) -avg_cost = fluid.layers.mean(cost) -return avg_cost + # The declaration of 'next_word' must be after the invoking of inference_program, + # or the data input order of train program would be [next_word, firstw, secondw, + # thirdw, fourthw], which is not correct. + predict_word = inference_program(is_sparse) + next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64') + cost = fluid.layers.cross_entropy(input=predict_word, label=next_word) + avg_cost = fluid.layers.mean(cost) + return avg_cost ``` - 现在我们可以开始训练啦。如今的版本较之以前就简单了许多。我们有现成的训练和测试集:`paddle.dataset.imikolov.train()`和`paddle.dataset.imikolov.test()`。两者都会返回一个读取器。在PaddlePaddle中,读取器是一个Python的函数,每次调用,会读取下一条数据。它是一个Python的generator。 @@ -285,59 +290,59 @@ return avg_cost ```python def optimizer_func(): -# Note here we need to choose more sophisticated optimizers -# such as AdaGrad with a decay rate. The normal SGD converges -# very slowly. -# optimizer=fluid.optimizer.SGD(learning_rate=0.001), -return fluid.optimizer.AdagradOptimizer( -learning_rate=3e-3, -regularization=fluid.regularizer.L2DecayRegularizer(8e-4)) + # Note here we need to choose more sophisticated optimizers + # such as AdaGrad with a decay rate. The normal SGD converges + # very slowly. + # optimizer=fluid.optimizer.SGD(learning_rate=0.001), + return fluid.optimizer.AdagradOptimizer( + learning_rate=3e-3, + regularization=fluid.regularizer.L2DecayRegularizer(8e-4)) def train(use_cuda, train_program, params_dirname): -train_reader = paddle.batch( -paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) -test_reader = paddle.batch( -paddle.dataset.imikolov.test(word_dict, N), BATCH_SIZE) - -place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - -def event_handler(event): -if isinstance(event, fluid.EndStepEvent): -# We output cost every 10 steps. -if event.step % 10 == 0: -outs = trainer.test( -reader=test_reader, -feed_order=['firstw', 'secondw', 'thirdw', 'fourthw', 'nextw']) -avg_cost = outs[0] - -print "Step %d: Average Cost %f" % (event.step, avg_cost) - -# If average cost is lower than 5.8, we consider the model good enough to stop. -# Note 5.8 is a relatively high value. In order to get a better model, one should -# aim for avg_cost lower than 3.5. But the training could take longer time. -if avg_cost < 5.8: -trainer.save_params(params_dirname) -trainer.stop() - -if math.isnan(avg_cost): -sys.exit("got NaN loss, training failed.") - -trainer = fluid.Trainer( -train_func=train_program, -optimizer_func=optimizer_func, -place=place) - -trainer.train( -reader=train_reader, -num_epochs=1, -event_handler=event_handler, -feed_order=['firstw', 'secondw', 'thirdw', 'fourthw', 'nextw']) + train_reader = paddle.batch( + paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) + test_reader = paddle.batch( + paddle.dataset.imikolov.test(word_dict, N), BATCH_SIZE) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + def event_handler(event): + if isinstance(event, fluid.EndStepEvent): + # We output cost every 10 steps. + if event.step % 10 == 0: + outs = trainer.test( + reader=test_reader, + feed_order=['firstw', 'secondw', 'thirdw', 'fourthw', 'nextw']) + avg_cost = outs[0] + + print("Step %d: Average Cost %f" % (event.step, avg_cost)) + + # If average cost is lower than 5.8, we consider the model good enough to stop. + # Note 5.8 is a relatively high value. In order to get a better model, one should + # aim for avg_cost lower than 3.5. But the training could take longer time. + if avg_cost < 5.8: + trainer.save_params(params_dirname) + trainer.stop() + + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + + trainer = fluid.Trainer( + train_func=train_program, + optimizer_func=optimizer_func, + place=place) + + trainer.train( + reader=train_reader, + num_epochs=1, + event_handler=event_handler, + feed_order=['firstw', 'secondw', 'thirdw', 'fourthw', 'nextw']) ``` - `trainer.train`将会开始训练。从`event_handler`返回的监控情况如下: -```python +```text Step 0: Average Cost 7.337213 Step 10: Average Cost 6.136128 Step 20: Average Cost 5.766995 @@ -352,50 +357,49 @@ Step 20: Average Cost 5.766995 ```python def infer(use_cuda, inference_program, params_dirname=None): -place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() -inferencer = fluid.Inferencer( -infer_func=inference_program, param_path=params_dirname, place=place) - -# Setup inputs by creating 4 LoDTensors representing 4 words. Here each word -# is simply an index to look up for the corresponding word vector and hence -# the shape of word (base_shape) should be [1]. The length-based level of -# detail (lod) info of each LoDtensor should be [[1]] meaning there is only -# one lod_level and there is only one sequence of one word on this level. -# Note that lod info should be a list of lists. - -data1 = [[211]] # 'among' -data2 = [[6]] # 'a' -data3 = [[96]] # 'group' -data4 = [[4]] # 'of' -lod = [[1]] - -first_word = fluid.create_lod_tensor(data1, lod, place) -second_word = fluid.create_lod_tensor(data2, lod, place) -third_word = fluid.create_lod_tensor(data3, lod, place) -fourth_word = fluid.create_lod_tensor(data4, lod, place) - -result = inferencer.infer( -{ -'firstw': first_word, -'secondw': second_word, -'thirdw': third_word, -'fourthw': fourth_word -}, -return_numpy=False) - -print(numpy.array(result[0])) -most_possible_word_index = numpy.argmax(result[0]) -print(most_possible_word_index) -print([ -key for key, value in word_dict.iteritems() -if value == most_possible_word_index -][0]) + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path=params_dirname, place=place) + + # Setup inputs by creating 4 LoDTensors representing 4 words. Here each word + # is simply an index to look up for the corresponding word vector and hence + # the shape of word (base_shape) should be [1]. The length-based level of + # detail (lod) info of each LoDtensor should be [[1]] meaning there is only + # one lod_level and there is only one sequence of one word on this level. + # Note that lod info should be a list of lists. + + data1 = [[211]] # 'among' + data2 = [[6]] # 'a' + data3 = [[96]] # 'group' + data4 = [[4]] # 'of' + lod = [[1]] + + first_word = fluid.create_lod_tensor(data1, lod, place) + second_word = fluid.create_lod_tensor(data2, lod, place) + third_word = fluid.create_lod_tensor(data3, lod, place) + fourth_word = fluid.create_lod_tensor(data4, lod, place) + + result = inferencer.infer( + { + 'firstw': first_word, + 'secondw': second_word, + 'thirdw': third_word, + 'fourthw': fourth_word + }, + return_numpy=False) + + print(numpy.array(result[0])) + most_possible_word_index = numpy.argmax(result[0]) + print(most_possible_word_index) + print([ + key for key, value in word_dict.iteritems() + if value == most_possible_word_index + ][0]) ``` 在经历3分钟的短暂训练后,我们得到如下的预测。我们的模型预测 `among a group of` 的下一个词是`a`。这比较符合文法规律。如果我们训练时间更长,比如几个小时,那么我们会得到的下一个预测是 `workers`。 - -```python +```text [[0.00106646 0.0007907 0.00072041 ... 0.00049024 0.00041355 0.00084464]] 6 a @@ -405,20 +409,20 @@ a ```python def main(use_cuda, is_sparse): -if use_cuda and not fluid.core.is_compiled_with_cuda(): -return + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return -params_dirname = "word2vec.inference.model" + params_dirname = "word2vec.inference.model" -train( -use_cuda=use_cuda, -train_program=partial(train_program, is_sparse), -params_dirname=params_dirname) + train( + use_cuda=use_cuda, + train_program=partial(train_program, is_sparse), + params_dirname=params_dirname) -infer( -use_cuda=use_cuda, -inference_program=partial(inference_program, is_sparse), -params_dirname=params_dirname) + infer( + use_cuda=use_cuda, + inference_program=partial(inference_program, is_sparse), + params_dirname=params_dirname) main(use_cuda=use_cuda, is_sparse=True) diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/2d_similarity.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/2d_similarity.png deleted file mode 100644 index 384f59919a2c8dedb198e97d51434616648932e1..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/2d_similarity.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow.png deleted file mode 100644 index 76b7d4bc0f99372465bd9aa34721513d39ad0776..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow_en.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow_en.png deleted file mode 100755 index d985c393e618e9b79df05e4ff0ae57ccc93744d0..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.en.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.en.png deleted file mode 100755 index 2e16ab2f443732b8ef5404a8e7cd2457bc5eee23..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.png deleted file mode 100644 index 2449dce6a86b43b1b997ff418ed0dba56848463f..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm.png deleted file mode 100644 index 1e0b40a8f7aefdf46d42761305511f281c08e595..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm_en.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm_en.png deleted file mode 100755 index 158bd64b8f8729dea67834a8d591d21bce8b8564..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/sentence_emb.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/sentence_emb.png deleted file mode 100644 index ce4a8bf4769183cbaff91793753d2350a3ce936c..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/sentence_emb.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram.png deleted file mode 100644 index a3ab385845d3dc8b5c670bae91225bc8dd47a8bb..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram_en.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram_en.png deleted file mode 100755 index 3c36c6d1f66eb98ea78c0673965d02a4ee3aa288..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram_en.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/README.cn.md b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/README.cn.md index ba43ada5100ed1db7192de9c795b4b8a6596d705..84d971a458038080e3f3883ac35e57059be2786d 100644 --- a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/README.cn.md +++ b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/README.cn.md @@ -1,144 +1,75 @@ -```eval_rst -.. _quick_start_fit_a_line: -``` # 线性回归 让我们从经典的线性回归(Linear Regression \[[1](#参考文献)\])模型开始这份教程。在这一章里,你将使用真实的数据集建立起一个房价预测模型,并且了解到机器学习中的若干重要概念。 -本教程源代码目录在[book/fit_a_line](https://github.com/PaddlePaddle/book/tree/develop/01.fit_a_line), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 +本教程源代码目录在[book/fit_a_line](https://github.com/PaddlePaddle/book/tree/develop/01.fit_a_line), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书),更多内容请参考本教程的[视频课堂](http://bit.baidu.com/course/detail/id/137.html)。 ## 背景介绍 -给定一个大小为`$n$`的数据集 `${\{y_{i}, x_{i1}, ..., x_{id}\}}_{i=1}^{n}$`,其中`$x_{i1}, \ldots, x_{id}$`是第`$i$`个样本`$d$`个属性上的取值,`$y_i$`是该样本待预测的目标。线性回归模型假设目标`$y_i$`可以被属性间的线性组合描述,即 +给定一个大小为$n$的数据集 ${\{y_{i}, x_{i1}, ..., x_{id}\}}_{i=1}^{n}$,其中$x_{i1}, \ldots, x_{id}$是第$i$个样本$d$个属性上的取值,$y_i$是该样本待预测的目标。线性回归模型假设目标$y_i$可以被属性间的线性组合描述,即 $$y_i = \omega_1x_{i1} + \omega_2x_{i2} + \ldots + \omega_dx_{id} + b, i=1,\ldots,n$$ -例如,在我们将要建模的房价预测问题里,`$x_{ij}$`是描述房子`$i$`的各种属性(比如房间的个数、周围学校和医院的个数、交通状况等),而 `$y_i$`是房屋的价格。 +例如,在我们将要建模的房价预测问题里,$x_{ij}$是描述房子$i$的各种属性(比如房间的个数、周围学校和医院的个数、交通状况等),而 $y_i$是房屋的价格。 初看起来,这个假设实在过于简单了,变量间的真实关系很难是线性的。但由于线性回归模型有形式简单和易于建模分析的优点,它在实际问题中得到了大量的应用。很多经典的统计学习、机器学习书籍\[[2,3,4](#参考文献)\]也选择对线性模型独立成章重点讲解。 ## 效果展示 我们使用从[UCI Housing Data Set](https://archive.ics.uci.edu/ml/datasets/Housing)获得的波士顿房价数据集进行模型的训练和预测。下面的散点图展示了使用模型对部分房屋价格进行的预测。其中,每个点的横坐标表示同一类房屋真实价格的中位数,纵坐标表示线性回归模型根据特征预测的结果,当二者值完全相等的时候就会落在虚线上。所以模型预测得越准确,则点离虚线越近。 - -![BostonHousePricePredictions](./image/predictions.png) -

图1. 预测值 V.S. 真实值

+

+
+ 图1. 预测值 V.S. 真实值 +

## 模型概览 ### 模型定义 -在波士顿房价数据集中,和房屋相关的值共有14个:前13个用来描述房屋相关的各种信息,即模型中的 `$x_i$`;最后一个值为我们要预测的该类房屋价格的中位数,即模型中的 `$y_i$`。因此,我们的模型就可以表示成: +在波士顿房价数据集中,和房屋相关的值共有14个:前13个用来描述房屋相关的各种信息,即模型中的 $x_i$;最后一个值为我们要预测的该类房屋价格的中位数,即模型中的 $y_i$。因此,我们的模型就可以表示成: $$\hat{Y} = \omega_1X_{1} + \omega_2X_{2} + \ldots + \omega_{13}X_{13} + b$$ -`$\hat{Y}$` 表示模型的预测结果,用来和真实值`$Y$`区分。模型要学习的参数即:`$\omega_1, \ldots, \omega_{13}, b$`。 +$\hat{Y}$ 表示模型的预测结果,用来和真实值$Y$区分。模型要学习的参数即:$\omega_1, \ldots, \omega_{13}, b$。 -建立模型后,我们需要给模型一个优化目标,使得学到的参数能够让预测值`$\hat{Y}$`尽可能地接近真实值`$Y$`。这里我们引入损失函数([Loss Function](https://en.wikipedia.org/wiki/Loss_function),或Cost Function)这个概念。 输入任意一个数据样本的目标值`$y_{i}$`和模型给出的预测值`$\hat{y_{i}}$`,损失函数输出一个非负的实值。这个实值通常用来反映模型误差的大小。 +建立模型后,我们需要给模型一个优化目标,使得学到的参数能够让预测值$\hat{Y}$尽可能地接近真实值$Y$。这里我们引入损失函数([Loss Function](https://en.wikipedia.org/wiki/Loss_function),或Cost Function)这个概念。 输入任意一个数据样本的目标值$y_{i}$和模型给出的预测值$\hat{y_{i}}$,损失函数输出一个非负的实值。这个实值通常用来反映模型误差的大小。 对于线性回归模型来讲,最常见的损失函数就是均方误差(Mean Squared Error, [MSE](https://en.wikipedia.org/wiki/Mean_squared_error))了,它的形式是: $$MSE=\frac{1}{n}\sum_{i=1}^{n}{(\hat{Y_i}-Y_i)}^2$$ -即对于一个大小为`$n$`的测试集,`$MSE$`是`$n$`个数据预测结果误差平方的均值。 +即对于一个大小为$n$的测试集,$MSE$是$n$个数据预测结果误差平方的均值。 ### 训练过程 定义好模型结构之后,我们要通过以下几个步骤进行模型训练 -1. 初始化参数,其中包括权重`$\omega_i$`和偏置`$b$`,对其进行初始化(如0均值,1方差)。 -2. 网络正向传播计算网络输出和损失函数。 -3. 根据损失函数进行反向误差传播 ([backpropagation](https://en.wikipedia.org/wiki/Backpropagation)),将网络误差从输出层依次向前传递, 并更新网络中的参数。 -4. 重复2~3步骤,直至网络训练误差达到规定的程度或训练轮次达到设定值。 + 1. 初始化参数,其中包括权重$\omega_i$和偏置$b$,对其进行初始化(如0均值,1方差)。 + 2. 网络正向传播计算网络输出和损失函数。 + 3. 根据损失函数进行反向误差传播 ([backpropagation](https://en.wikipedia.org/wiki/Backpropagation)),将网络误差从输出层依次向前传递, 并更新网络中的参数。 + 4. 重复2~3步骤,直至网络训练误差达到规定的程度或训练轮次达到设定值。 ## 数据集 ### 数据集介绍 这份数据集共506行,每行包含了波士顿郊区的一类房屋的相关信息及该类房屋价格的中位数。其各维属性的意义如下: -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
属性名解释类型
CRIM该镇的人均犯罪率连续值
ZN占地面积超过25,000平方呎的住宅用地比例连续值
INDUS非零售商业用地比例连续值
CHAS是否邻近 Charles River离散值,1=邻近;0=不邻近
NOX一氧化氮浓度连续值
RM每栋房屋的平均客房数连续值
AGE1940年之前建成的自用单位比例连续值
DIS到波士顿5个就业中心的加权距离连续值
RAD到径向公路的可达性指数连续值
TAX全值财产税率连续值
PTRATIO学生与教师的比例连续值
B1000(BK - 0.63)^2,其中BK为黑人占比连续值
LSTAT低收入人群占比连续值
MEDV同类房屋价格的中位数连续值
-

+| 属性名 | 解释 | 类型 | +| ------| ------ | ------ | +| CRIM | 该镇的人均犯罪率 | 连续值 | +| ZN | 占地面积超过25,000平方呎的住宅用地比例 | 连续值 | +| INDUS | 非零售商业用地比例 | 连续值 | +| CHAS | 是否邻近 Charles River | 离散值,1=邻近;0=不邻近 | +| NOX | 一氧化氮浓度 | 连续值 | +| RM | 每栋房屋的平均客房数 | 连续值 | +| AGE | 1940年之前建成的自用单位比例 | 连续值 | +| DIS | 到波士顿5个就业中心的加权距离 | 连续值 | +| RAD | 到径向公路的可达性指数 | 连续值 | +| TAX | 全值财产税率 | 连续值 | +| PTRATIO | 学生与教师的比例 | 连续值 | +| B | 1000(BK - 0.63)^2,其中BK为黑人占比 | 连续值 | +| LSTAT | 低收入人群占比 | 连续值 | +| MEDV | 同类房屋价格的中位数 | 连续值 | ### 数据预处理 #### 连续值与离散值 -观察一下数据,我们的第一个发现是:所有的13维属性中,有12维的连续值和1维的离散值(CHAS)。离散值虽然也常使用类似0、1、2这样的数字表示,但是其含义与连续值是不同的,因为这里的差值没有实际意义。例如,我们用0、1、2来分别表示红色、绿色和蓝色的话,我们并不能因此说“蓝色和红色”比“绿色和红色”的距离更远。所以通常对一个有`$d$`个可能取值的离散属性,我们会将它们转为`$d$`个取值为0或1的二值属性或者将每个可能取值映射为一个多维向量。不过就这里而言,因为CHAS本身就是一个二值属性,就省去了这个麻烦。 +观察一下数据,我们的第一个发现是:所有的13维属性中,有12维的连续值和1维的离散值(CHAS)。离散值虽然也常使用类似0、1、2这样的数字表示,但是其含义与连续值是不同的,因为这里的差值没有实际意义。例如,我们用0、1、2来分别表示红色、绿色和蓝色的话,我们并不能因此说“蓝色和红色”比“绿色和红色”的距离更远。所以通常对一个有$d$个可能取值的离散属性,我们会将它们转为$d$个取值为0或1的二值属性或者将每个可能取值映射为一个多维向量。不过就这里而言,因为CHAS本身就是一个二值属性,就省去了这个麻烦。 #### 属性的归一化 另外一个稍加观察即可发现的事实是,各维属性的取值范围差别很大(如图2所示)。例如,属性B的取值范围是[0.32, 396.90],而属性NOX的取值范围是[0.3850, 0.8170]。这里就要用到一个常见的操作-归一化(normalization)了。归一化的目标是把各位属性的取值范围放缩到差不多的区间,例如[-0.5,0.5]。这里我们使用一种很常见的操作方法:减掉均值,然后除以原取值范围。 @@ -148,11 +79,13 @@ $$MSE=\frac{1}{n}\sum_{i=1}^{n}{(\hat{Y_i}-Y_i)}^2$$ - 不同的数值范围会导致不同属性对模型的重要性不同(至少在训练的初始阶段如此),而这个隐含的假设常常是不合理的。这会对优化的过程造成困难,使训练时间大大的加长。 - 很多的机器学习技巧/模型(例如L1,L2正则项,向量空间模型-Vector Space Model)都基于这样的假设:所有的属性取值都差不多是以0为均值且取值范围相近的。 -![featureScale](./image/ranges.png) -

图2. 各维属性的取值范围

+

+
+ 图2. 各维属性的取值范围 +

#### 整理训练集与测试集 -我们将数据集分割为两份:一份用于调整模型的参数,即进行模型的训练,模型在这份数据集上的误差被称为**训练误差**;另外一份被用来测试,模型在这份数据集上的误差被称为**测试误差**。我们训练模型的目的是为了通过从训练数据中找到规律来预测未知的新数据,所以测试误差是更能反映模型表现的指标。分割数据的比例要考虑到两个因素:更多的训练数据会降低参数估计的方差,从而得到更可信的模型;而更多的测试数据会降低测试误差的方差,从而得到更可信的测试误差。我们这个例子中设置的分割比例为`$8:2$` +我们将数据集分割为两份:一份用于调整模型的参数,即进行模型的训练,模型在这份数据集上的误差被称为**训练误差**;另外一份被用来测试,模型在这份数据集上的误差被称为**测试误差**。我们训练模型的目的是为了通过从训练数据中找到规律来预测未知的新数据,所以测试误差是更能反映模型表现的指标。分割数据的比例要考虑到两个因素:更多的训练数据会降低参数估计的方差,从而得到更可信的模型;而更多的测试数据会降低测试误差的方差,从而得到更可信的测试误差。我们这个例子中设置的分割比例为$8:2$ 在更复杂的模型训练过程中,我们往往还会多使用一种数据集:验证集。因为复杂的模型中常常还有一些超参数([Hyperparameter](https://en.wikipedia.org/wiki/Hyperparameter_optimization))需要调节,所以我们会尝试多种超参数的组合来分别训练多个模型,然后对比它们在验证集上的表现选择相对最好的一组超参数,最后才使用这组参数下训练的模型在测试集上评估测试误差。由于本章训练的模型比较简单,我们暂且忽略掉这个过程。 @@ -167,6 +100,7 @@ $$MSE=\frac{1}{n}\sum_{i=1}^{n}{(\hat{Y_i}-Y_i)}^2$$ import paddle import paddle.fluid as fluid import numpy +from __future__ import print_function ``` 我们通过uci_housing模块引入了数据集合[UCI Housing Data Set](https://archive.ics.uci.edu/ml/datasets/Housing) @@ -182,14 +116,14 @@ import numpy BATCH_SIZE = 20 train_reader = paddle.batch( -paddle.reader.shuffle( -paddle.dataset.uci_housing.train(), buf_size=500), -batch_size=BATCH_SIZE) + paddle.reader.shuffle( + paddle.dataset.uci_housing.train(), buf_size=500), + batch_size=BATCH_SIZE) test_reader = paddle.batch( -paddle.reader.shuffle( -paddle.dataset.uci_housing.test(), buf_size=500), -batch_size=BATCH_SIZE) + paddle.reader.shuffle( + paddle.dataset.uci_housing.test(), buf_size=500), + batch_size=BATCH_SIZE) ``` ### 配置训练程序 @@ -197,16 +131,25 @@ batch_size=BATCH_SIZE) ```python def train_program(): -y = fluid.layers.data(name='y', shape=[1], dtype='float32') + y = fluid.layers.data(name='y', shape=[1], dtype='float32') -# feature vector of length 13 -x = fluid.layers.data(name='x', shape=[13], dtype='float32') -y_predict = fluid.layers.fc(input=x, size=1, act=None) + # feature vector of length 13 + x = fluid.layers.data(name='x', shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) -loss = fluid.layers.square_error_cost(input=y_predict, label=y) -avg_loss = fluid.layers.mean(loss) + loss = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_loss = fluid.layers.mean(loss) -return avg_loss + return avg_loss +``` + +### Optimizer Function 配置 + +在下面的 `SGD optimizer`,`learning_rate` 是训练的速度,与网络的训练收敛速度有关系。 + +```python +def optimizer_program(): + return fluid.optimizer.SGD(learning_rate=0.001) ``` ### 定义运算场所 @@ -222,9 +165,9 @@ place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() ```python trainer = fluid.Trainer( -train_func=train_program, -place=place, -optimizer_func=fluid.optimizer.SGD(learning_rate=0.001)) + train_func=train_program, + place=place, + optimizer_func=optimizer_program) ``` ### 开始提供数据 @@ -237,7 +180,7 @@ feed_order=['x', 'y'] 除此之外,可以定义一个事件相应器来处理类似`打印训练进程`的事件: ```python -# Specify the directory path to save the parameters +# Specify the directory to save the parameters params_dirname = "fit_a_line.inference.model" # Plot data @@ -248,27 +191,30 @@ plot_cost = Ploter(train_title, test_title) step = 0 -# event_handler to print training and testing info +# event_handler prints training and testing info def event_handler_plot(event): -global step -if isinstance(event, fluid.EndStepEvent): -if event.step % 10 == 0: # every 10 batches, record a test cost -test_metrics = trainer.test( -reader=test_reader, feed_order=feed_order) - -plot_cost.append(test_title, step, test_metrics[0]) -plot_cost.plot() - -if test_metrics[0] < 10.0: -# If the accuracy is good enough, we can stop the training. -print('loss is less than 10.0, stop') -trainer.stop() - -# We can save the trained parameters for the inferences later -if params_dirname is not None: -trainer.save_params(params_dirname) - -step += 1 + global step + if isinstance(event, fluid.EndStepEvent): + if step % 10 == 0: # record a train cost every 10 batches + plot_cost.append(train_title, step, event.metrics[0]) + + if step % 100 == 0: # record a test cost every 100 batches + test_metrics = trainer.test( + reader=test_reader, feed_order=feed_order) + plot_cost.append(test_title, step, test_metrics[0]) + plot_cost.plot() + + if test_metrics[0] < 10.0: + # If the accuracy is good enough, we can stop the training. + print('loss is less than 10.0, stop') + trainer.stop() + step += 1 + + if isinstance(event, fluid.EndEpochEvent): + if event.epoch % 10 == 0: + # We can save the trained parameters for the inferences later + if params_dirname is not None: + trainer.save_params(params_dirname) ``` ### 开始训练 @@ -279,13 +225,16 @@ step += 1 # The training could take up to a few minutes. trainer.train( -reader=train_reader, -num_epochs=100, -event_handler=event_handler_plot, -feed_order=feed_order) + reader=train_reader, + num_epochs=100, + event_handler=event_handler_plot, + feed_order=feed_order) ``` +
+
+图3 训练结果 +
-![trainTestCost](./image/train_and_test.png) ## 预测 提供一个`inference_program`和一个`params_dirname`来初始化预测器。`params_dirname`用来存储我们的参数。 @@ -296,9 +245,9 @@ feed_order=feed_order) ```python def inference_program(): -x = fluid.layers.data(name='x', shape=[13], dtype='float32') -y_predict = fluid.layers.fc(input=x, size=1, act=None) -return y_predict + x = fluid.layers.data(name='x', shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + return y_predict ``` ### 预测 @@ -306,13 +255,23 @@ return y_predict ```python inferencer = fluid.Inferencer( -infer_func=inference_program, param_path=params_dirname, place=place) + infer_func=inference_program, param_path=params_dirname, place=place) batch_size = 10 -tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32") +test_reader = paddle.batch(paddle.dataset.uci_housing.test(),batch_size=batch_size) +test_data = test_reader().next() +test_x = numpy.array([data[0] for data in test_data]).astype("float32") +test_y = numpy.array([data[1] for data in test_data]).astype("float32") + +results = inferencer.infer({'x': test_x}) + +print("infer results: (House Price)") +for idx, val in enumerate(results[0]): + print("%d: %.2f" % (idx, val)) -results = inferencer.infer({'x': tensor_x}) -print("infer results: ", results[0]) +print("\nground truth:") +for idx, val in enumerate(test_y): + print("%d: %.2f" % (idx, val)) ``` ## 总结 diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/predictions.png b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/predictions.png deleted file mode 100644 index 27e4acb1313794f52ad9ad9e874cdadd197ff41f..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/predictions.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/ranges.png b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/ranges.png deleted file mode 100644 index 5d86b12715f46afbafb7d50e2938e184219b5b95..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/ranges.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/train_and_test.png b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/train_and_test.png deleted file mode 100644 index bcd304a6a0baf30ecfbc43e08fc0aca179d05958..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/train_and_test.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/README.cn.md b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/README.cn.md index c04a949a3f6550048f2a3447070829aeb640b995..6672a3170e2132f3ce5df2f5e6eb54fbd77d40e3 100644 --- a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/README.cn.md +++ b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/README.cn.md @@ -1,12 +1,14 @@ # 识别数字 -本教程源代码目录在[book/recognize_digits](https://github.com/PaddlePaddle/book/tree/develop/02.recognize_digits), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 +本教程源代码目录在[book/recognize_digits](https://github.com/PaddlePaddle/book/tree/develop/02.recognize_digits), 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书),更多内容请参考本教程的[视频课堂](http://bit.baidu.com/course/detail/id/167.html)。 ## 背景介绍 当我们学习编程的时候,编写的第一个程序一般是实现打印"Hello World"。而机器学习(或深度学习)的入门教程,一般都是 [MNIST](http://yann.lecun.com/exdb/mnist/) 数据库上的手写识别问题。原因是手写识别属于典型的图像分类问题,比较简单,同时MNIST数据集也很完备。MNIST数据集作为一个简单的计算机视觉数据集,包含一系列如图1所示的手写数字图片和对应的标签。图片是28x28的像素矩阵,标签则对应着0~9的10个数字。每张图片都经过了大小归一化和居中处理。 -![MNIST](./image/mnist_example_image.png) -

图1. MNIST图片示例

+

+
+图1. MNIST图片示例 +

MNIST数据集是从 [NIST](https://www.nist.gov/srd/nist-special-database-19) 的Special Database 3(SD-3)和Special Database 1(SD-1)构建而来。由于SD-3是由美国人口调查局的员工进行标注,SD-1是由美国高中生进行标注,因此SD-3比SD-1更干净也更容易识别。Yann LeCun等人从SD-1和SD-3中各取一半作为MNIST的训练集(60000条数据)和测试集(10000条数据),其中训练集来自250位不同的标注员,此外还保证了训练集和测试集的标注员是不完全相同的。 @@ -20,86 +22,96 @@ Yann LeCun早先在手写字符识别上做了很多研究,并在研究过程 ## 模型概览 基于MNIST数据训练一个分类器,在介绍本教程使用的三个基本图像分类网络前,我们先给出一些定义: -- `$X$`是输入:MNIST图片是`$28\times28$` 的二维图像,为了进行计算,我们将其转化为`$784$`维向量,即`$X=\left ( x_0, x_1, \dots, x_{783} \right )$`。 -- `$Y$`是输出:分类器的输出是10类数字(0-9),即`$Y=\left ( y_0, y_1, \dots, y_9 \right )$`,每一维`$y_i$`代表图片分类为第`$i$`类数字的概率。 -- `$L$`是图片的真实标签:`$L=\left ( l_0, l_1, \dots, l_9 \right )$`也是10维,但只有一维为1,其他都为0。 +- $X$是输入:MNIST图片是$28\times28$ 的二维图像,为了进行计算,我们将其转化为$784$维向量,即$X=\left ( x_0, x_1, \dots, x_{783} \right )$。 +- $Y$是输出:分类器的输出是10类数字(0-9),即$Y=\left ( y_0, y_1, \dots, y_9 \right )$,每一维$y_i$代表图片分类为第$i$类数字的概率。 +- $L$是图片的真实标签:$L=\left ( l_0, l_1, \dots, l_9 \right )$也是10维,但只有一维为1,其他都为0。 ### Softmax回归(Softmax Regression) 最简单的Softmax回归模型是先将输入层经过一个全连接层得到的特征,然后直接通过softmax 函数进行多分类\[[9](#参考文献)\]。 -输入层的数据`$X$`传到输出层,在激活操作之前,会乘以相应的权重 `$W$` ,并加上偏置变量 `$b$` ,具体如下: +输入层的数据$X$传到输出层,在激活操作之前,会乘以相应的权重 $W$ ,并加上偏置变量 $b$ ,具体如下: $$ y_i = \text{softmax}(\sum_j W_{i,j}x_j + b_i) $$ -其中 `$ \text{softmax}(x_i) = \frac{e^{x_i}}{\sum_j e^{x_j}} $` +其中 $ \text{softmax}(x_i) = \frac{e^{x_i}}{\sum_j e^{x_j}} $ -对于有 `$N$` 个类别的多分类问题,指定 `$N$` 个输出节点,`$N$` 维结果向量经过softmax将归一化为 `$N$` 个[0,1]范围内的实数值,分别表示该样本属于这 `$N$` 个类别的概率。此处的 `$y_i$` 即对应该图片为数字 `$i$` 的预测概率。 +对于有 $N$ 个类别的多分类问题,指定 $N$ 个输出节点,$N$ 维结果向量经过softmax将归一化为 $N$ 个[0,1]范围内的实数值,分别表示该样本属于这 $N$ 个类别的概率。此处的 $y_i$ 即对应该图片为数字 $i$ 的预测概率。 -在分类问题中,我们一般采用交叉熵代价损失函数(cross entropy),公式如下: +在分类问题中,我们一般采用交叉熵代价损失函数(cross entropy loss),公式如下: -$$ \text{crossentropy}(label, y) = -\sum_i label_ilog(y_i) $$ +$$ L_{cross-entropy}(label, y) = -\sum_i label_ilog(y_i) $$ 图2为softmax回归的网络图,图中权重用蓝线表示、偏置用红线表示、+1代表偏置参数的系数为1。 -![softmaxRegression](./image/softmax_regression.png) -

图2. softmax回归网络结构图

+

+
+图2. softmax回归网络结构图
+

### 多层感知器(Multilayer Perceptron, MLP) Softmax回归模型采用了最简单的两层神经网络,即只有输入层和输出层,因此其拟合能力有限。为了达到更好的识别效果,我们考虑在输入层和输出层中间加上若干个隐藏层\[[10](#参考文献)\]。 -1. 经过第一个隐藏层,可以得到 `$ H_1 = \phi(W_1X + b_1) $`,其中`$\phi$`代表激活函数,常见的有sigmoid、tanh或ReLU等函数。 -2. 经过第二个隐藏层,可以得到 `$ H_2 = \phi(W_2H_1 + b_2) $`。 -3. 最后,再经过输出层,得到的`$Y=\text{softmax}(W_3H_2 + b_3)$`,即为最后的分类结果向量。 +1. 经过第一个隐藏层,可以得到 $ H_1 = \phi(W_1X + b_1) $,其中$\phi$代表激活函数,常见的有sigmoid、tanh或ReLU等函数。 +2. 经过第二个隐藏层,可以得到 $ H_2 = \phi(W_2H_1 + b_2) $。 +3. 最后,再经过输出层,得到的$Y=\text{softmax}(W_3H_2 + b_3)$,即为最后的分类结果向量。 图3为多层感知器的网络结构图,图中权重用蓝线表示、偏置用红线表示、+1代表偏置参数的系数为1。 -![multilayerPerceptron](./image/mlp.png) -

图3. 多层感知器网络结构图

+

+
+图3. 多层感知器网络结构图
+

### 卷积神经网络(Convolutional Neural Network, CNN) 在多层感知器模型中,将图像展开成一维向量输入到网络中,忽略了图像的位置和结构信息,而卷积神经网络能够更好的利用图像的结构信息。[LeNet-5](http://yann.lecun.com/exdb/lenet/)是一个较简单的卷积神经网络。图4显示了其结构:输入的二维图像,先经过两次卷积层到池化层,再经过全连接层,最后使用softmax分类作为输出层。下面我们主要介绍卷积层和池化层。 -![cnnStructure](./image/cnn.png) -

图4. LeNet-5卷积神经网络结构

+

+
+图4. LeNet-5卷积神经网络结构
+

#### 卷积层 卷积层是卷积神经网络的核心基石。在图像识别里我们提到的卷积是二维卷积,即离散二维滤波器(也称作卷积核)与二维图像做卷积操作,简单的讲是二维滤波器滑动到二维图像上所有位置,并在每个位置上与该像素点及其领域像素点做内积。卷积操作被广泛应用与图像处理领域,不同卷积核可以提取不同的特征,例如边沿、线性、角等特征。在深层卷积神经网络中,通过卷积操作可以提取出图像低级到复杂的特征。 -![cnn](https://raw.githubusercontent.com/PaddlePaddle/book/develop/02.recognize_digits/image/conv_layer.png) -

图5. 卷积层图片

+

+
+图5. 卷积层图片
+

-图5给出一个卷积计算过程的示例图,输入图像大小为`$H=5,W=5,D=3$`,即`$5 \times 5$`大小的3通道(RGB,也称作深度)彩色图像。这个示例图中包含两(用`$K$`表示)组卷积核,即图中滤波器`$W_0$`和`$W_1$`。在卷积计算中,通常对不同的输入通道采用不同的卷积核,如图示例中每组卷积核包含(`$D=3$`)个`$3 \times 3$`(用`$F \times F$`表示)大小的卷积核。另外,这个示例中卷积核在图像的水平方向(`$W$`方向)和垂直方向(`$H$`方向)的滑动步长为2(用`$S$`表示);对输入图像周围各填充1(用`$P$`表示)个0,即图中输入层原始数据为蓝色部分,灰色部分是进行了大小为1的扩展,用0来进行扩展。经过卷积操作得到输出为`$3 \times 3 \times 2$`(用`$H_{o} \times W_{o} \times K$`表示)大小的特征图,即`$3 \times 3$`大小的2通道特征图,其中`$H_o$`计算公式为:`$H_o = (H - F + 2 \times P)/S + 1$`,`$W_o$`同理。 而输出特征图中的每个像素,是每组滤波器与输入图像每个特征图的内积再求和,再加上偏置`$b_o$`,偏置通常对于每个输出特征图是共享的。输出特征图`$o[:,:,0]$`中的最后一个`$-2$`计算如图5右下角公式所示。 +图5给出一个卷积计算过程的示例图,输入图像大小为$H=5,W=5,D=3$,即$5 \times 5$大小的3通道(RGB,也称作深度)彩色图像。这个示例图中包含两(用$K$表示)组卷积核,即图中滤波器$W_0$和$W_1$。在卷积计算中,通常对不同的输入通道采用不同的卷积核,如图示例中每组卷积核包含($D=3)$个$3 \times 3$(用$F \times F$表示)大小的卷积核。另外,这个示例中卷积核在图像的水平方向($W$方向)和垂直方向($H$方向)的滑动步长为2(用$S$表示);对输入图像周围各填充1(用$P$表示)个0,即图中输入层原始数据为蓝色部分,灰色部分是进行了大小为1的扩展,用0来进行扩展。经过卷积操作得到输出为$3 \times 3 \times 2$(用$H_{o} \times W_{o} \times K$表示)大小的特征图,即$3 \times 3$大小的2通道特征图,其中$H_o$计算公式为:$H_o = (H - F + 2 \times P)/S + 1$,$W_o$同理。 而输出特征图中的每个像素,是每组滤波器与输入图像每个特征图的内积再求和,再加上偏置$b_o$,偏置通常对于每个输出特征图是共享的。输出特征图$o[:,:,0]$中的最后一个$-2$计算如图5右下角公式所示。 -在卷积操作中卷积核是可学习的参数,经过上面示例介绍,每层卷积的参数大小为`$D \times F \times F \times K$`。在多层感知器模型中,神经元通常是全部连接,参数较多。而卷积层的参数较少,这也是由卷积层的主要特性即局部连接和共享权重所决定。 +在卷积操作中卷积核是可学习的参数,经过上面示例介绍,每层卷积的参数大小为$D \times F \times F \times K$。在多层感知器模型中,神经元通常是全部连接,参数较多。而卷积层的参数较少,这也是由卷积层的主要特性即局部连接和共享权重所决定。 - 局部连接:每个神经元仅与输入神经元的一块区域连接,这块局部区域称作感受野(receptive field)。在图像卷积操作中,即神经元在空间维度(spatial dimension,即上图示例H和W所在的平面)是局部连接,但在深度上是全部连接。对于二维图像本身而言,也是局部像素关联较强。这种局部连接保证了学习后的过滤器能够对于局部的输入特征有最强的响应。局部连接的思想,也是受启发于生物学里面的视觉系统结构,视觉皮层的神经元就是局部接受信息的。 -- 权重共享:计算同一个深度切片的神经元时采用的滤波器是共享的。例如图4中计算`$o[:,:,0]$`的每个每个神经元的滤波器均相同,都为`$W_0$`,这样可以很大程度上减少参数。共享权重在一定程度上讲是有意义的,例如图片的底层边缘特征与特征在图中的具体位置无关。但是在一些场景中是无意的,比如输入的图片是人脸,眼睛和头发位于不同的位置,希望在不同的位置学到不同的特征 (参考[斯坦福大学公开课]( http://cs231n.github.io/convolutional-networks/))。请注意权重只是对于同一深度切片的神经元是共享的,在卷积层,通常采用多组卷积核提取不同特征,即对应不同深度切片的特征,不同深度切片的神经元权重是不共享。另外,偏重对同一深度切片的所有神经元都是共享的。 +- 权重共享:计算同一个深度切片的神经元时采用的滤波器是共享的。例如图4中计算$o[:,:,0]$的每个每个神经元的滤波器均相同,都为$W_0$,这样可以很大程度上减少参数。共享权重在一定程度上讲是有意义的,例如图片的底层边缘特征与特征在图中的具体位置无关。但是在一些场景中是无意的,比如输入的图片是人脸,眼睛和头发位于不同的位置,希望在不同的位置学到不同的特征 (参考[斯坦福大学公开课]( http://cs231n.github.io/convolutional-networks/))。请注意权重只是对于同一深度切片的神经元是共享的,在卷积层,通常采用多组卷积核提取不同特征,即对应不同深度切片的特征,不同深度切片的神经元权重是不共享。另外,偏重对同一深度切片的所有神经元都是共享的。 通过介绍卷积计算过程及其特性,可以看出卷积是线性操作,并具有平移不变性(shift-invariant),平移不变性即在图像每个位置执行相同的操作。卷积层的局部连接和权重共享使得需要学习的参数大大减小,这样也有利于训练较大卷积神经网络。 #### 池化层 -![pooling](./image/max_pooling.png) -

图6. 池化层图片

+

+
+图6. 池化层图片
+

池化是非线性下采样的一种形式,主要作用是通过减少网络的参数来减小计算量,并且能够在一定程度上控制过拟合。通常在卷积层的后面会加上一个池化层。池化包括最大池化、平均池化等。其中最大池化是用不重叠的矩形框将输入层分成不同的区域,对于每个矩形框的数取最大值作为输出层,如图6所示。 更详细的关于卷积神经网络的具体知识可以参考[斯坦福大学公开课]( http://cs231n.github.io/convolutional-networks/ )和[图像分类](https://github.com/PaddlePaddle/book/blob/develop/image_classification/README.md)教程。 -### 常见激活函数介绍 -- sigmoid激活函数: `$ f(x) = sigmoid(x) = \frac{1}{1+e^{-x}} $` +### 常见激活函数介绍 +- sigmoid激活函数: $ f(x) = sigmoid(x) = \frac{1}{1+e^{-x}} $ -- tanh激活函数: `$ f(x) = tanh(x) = \frac{e^x-e^{-x}}{e^x+e^{-x}} $` +- tanh激活函数: $ f(x) = tanh(x) = \frac{e^x-e^{-x}}{e^x+e^{-x}} $ -实际上,tanh函数只是规模变化的sigmoid函数,将sigmoid函数值放大2倍之后再向下平移1个单位:tanh(x) = 2sigmoid(2x) - 1 。 + 实际上,tanh函数只是规模变化的sigmoid函数,将sigmoid函数值放大2倍之后再向下平移1个单位:tanh(x) = 2sigmoid(2x) - 1 。 -- ReLU激活函数: `$ f(x) = max(0, x) $` +- ReLU激活函数: $ f(x) = max(0, x) $ 更详细的介绍请参考[维基百科激活函数](https://en.wikipedia.org/wiki/Activation_function)。 @@ -107,35 +119,13 @@ Softmax回归模型采用了最简单的两层神经网络,即只有输入层 PaddlePaddle在API中提供了自动加载[MNIST](http://yann.lecun.com/exdb/mnist/)数据的模块`paddle.dataset.mnist`。加载后的数据位于`/home/username/.cache/paddle/dataset/mnist`下: -

- - - - - - - - - - - - - - - - - - - - - - - - - - -
文件名称说明
train-images-idx3-ubyte训练数据图片,60,000条数据
train-labels-idx1-ubyte训练数据标签,60,000条数据
t10k-images-idx3-ubyte测试数据图片,10,000条数据
t10k-labels-idx1-ubyte测试数据标签,10,000条数据
-

+ +| 文件名称 | 说明 | +|----------------------|-------------------------| +|train-images-idx3-ubyte| 训练数据图片,60,000条数据 | +|train-labels-idx1-ubyte| 训练数据标签,60,000条数据 | +|t10k-images-idx3-ubyte | 测试数据图片,10,000条数据 | +|t10k-labels-idx1-ubyte | 测试数据标签,10,000条数据 | ## Fluid API 概述 @@ -165,6 +155,7 @@ PaddlePaddle在API中提供了自动加载[MNIST](http://yann.lecun.com/exdb/mni ```python import paddle import paddle.fluid as fluid +from __future__ import print_function ``` ### Program Functions 配置 @@ -177,51 +168,51 @@ import paddle.fluid as fluid ```python def softmax_regression(): -img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') -predict = fluid.layers.fc( -input=img, size=10, act='softmax') -return predict + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + predict = fluid.layers.fc( + input=img, size=10, act='softmax') + return predict ``` - 多层感知器:下面代码实现了一个含有两个隐藏层(即全连接层)的多层感知器。其中两个隐藏层的激活函数均采用ReLU,输出层的激活函数用Softmax。 ```python def multilayer_perceptron(): -img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') -# 第一个全连接层,激活函数为ReLU -hidden = fluid.layers.fc(input=img, size=200, act='relu') -# 第二个全连接层,激活函数为ReLU -hidden = fluid.layers.fc(input=hidden, size=200, act='relu') -# 以softmax为激活函数的全连接输出层,输出层的大小必须为数字的个数10 -prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') -return prediction + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + # 第一个全连接层,激活函数为ReLU + hidden = fluid.layers.fc(input=img, size=200, act='relu') + # 第二个全连接层,激活函数为ReLU + hidden = fluid.layers.fc(input=hidden, size=200, act='relu') + # 以softmax为激活函数的全连接输出层,输出层的大小必须为数字的个数10 + prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') + return prediction ``` - 卷积神经网络LeNet-5: 输入的二维图像,首先经过两次卷积层到池化层,再经过全连接层,最后使用以softmax为激活函数的全连接层作为输出层。 ```python def convolutional_neural_network(): -img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') -# 第一个卷积-池化层 -conv_pool_1 = fluid.nets.simple_img_conv_pool( -input=img, -filter_size=5, -num_filters=20, -pool_size=2, -pool_stride=2, -act="relu") -conv_pool_1 = fluid.layers.batch_norm(conv_pool_1) -# 第二个卷积-池化层 -conv_pool_2 = fluid.nets.simple_img_conv_pool( -input=conv_pool_1, -filter_size=5, -num_filters=50, -pool_size=2, -pool_stride=2, -act="relu") -# 以softmax为激活函数的全连接输出层,输出层的大小必须为数字的个数10 -prediction = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax') -return prediction + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + # 第一个卷积-池化层 + conv_pool_1 = fluid.nets.simple_img_conv_pool( + input=img, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu") + conv_pool_1 = fluid.layers.batch_norm(conv_pool_1) + # 第二个卷积-池化层 + conv_pool_2 = fluid.nets.simple_img_conv_pool( + input=conv_pool_1, + filter_size=5, + num_filters=50, + pool_size=2, + pool_stride=2, + act="relu") + # 以softmax为激活函数的全连接输出层,输出层的大小必须为数字的个数10 + prediction = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax') + return prediction ``` #### Train Program 配置 @@ -234,18 +225,17 @@ return prediction ```python def train_program(): -label = fluid.layers.data(name='label', shape=[1], dtype='int64') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') -# predict = softmax_regression() # uncomment for Softmax回归 -# predict = multilayer_perceptron() # uncomment for 多层感知器 -predict = convolutional_neural_network() # uncomment for LeNet5卷积神经网络 -cost = fluid.layers.cross_entropy(input=predict, label=label) -avg_cost = fluid.layers.mean(cost) -acc = fluid.layers.accuracy(input=predict, label=label) -return [avg_cost, acc] + # predict = softmax_regression() # uncomment for Softmax回归 + # predict = multilayer_perceptron() # uncomment for 多层感知器 + predict = convolutional_neural_network() # uncomment for LeNet5卷积神经网络 + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(cost) + acc = fluid.layers.accuracy(input=predict, label=label) + return [avg_cost, acc] -# 该模型运行在单个CPU上 ``` #### Optimizer Function 配置 @@ -254,25 +244,25 @@ return [avg_cost, acc] ```python def optimizer_program(): -return fluid.optimizer.Adam(learning_rate=0.001) + return fluid.optimizer.Adam(learning_rate=0.001) ``` ### 数据集 Feeders 配置 下一步,我们开始训练过程。`paddle.dataset.movielens.train()`和`paddle.dataset.movielens.test()`分别做训练和测试数据集。这两个函数各自返回一个reader——PaddlePaddle中的reader是一个Python函数,每次调用的时候返回一个Python yield generator。 -下面`shuffle`是一个reader decorator,它接受一个reader A,返回另一个reader B —— reader B 每次读入`buffer_size`条训练数据到一个buffer里,然后随机打乱其顺序,并且逐条输出。 +下面`shuffle`是一个reader decorator,它接受一个reader A,返回另一个reader B。reader B 每次读入`buffer_size`条训练数据到一个buffer里,然后随机打乱其顺序,并且逐条输出。 -`batch`是一个特殊的decorator,它的输入是一个reader,输出是一个batched reader —— 在PaddlePaddle里,一个reader每次yield一条训练数据,而一个batched reader每次yield一个minibatch。 +`batch`是一个特殊的decorator,它的输入是一个reader,输出是一个batched reader。在PaddlePaddle里,一个reader每次yield一条训练数据,而一个batched reader每次yield一个minibatch。 ```python train_reader = paddle.batch( -paddle.reader.shuffle( -paddle.dataset.mnist.train(), buf_size=500), -batch_size=64) + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=500), + batch_size=64) test_reader = paddle.batch( -paddle.dataset.mnist.test(), batch_size=64) + paddle.dataset.mnist.test(), batch_size=64) ``` ### Trainer 配置 @@ -285,7 +275,7 @@ use_cuda = False # set to True if training with GPU place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( -train_func=train_program, place=place, optimizer_func=optimizer_program) + train_func=train_program, place=place, optimizer_func=optimizer_program) ``` #### Event Handler 配置 @@ -300,27 +290,31 @@ Fluid API 在训练期间为回调函数提供了一个钩子。用户能够通 params_dirname = "recognize_digits_network.inference.model" lists = [] def event_handler(event): -if isinstance(event, fluid.EndStepEvent): -if event.step % 100 == 0: -# event.metrics maps with train program return arguments. -# event.metrics[0] will yeild avg_cost and event.metrics[1] will yeild acc in this example. -print "Pass %d, Batch %d, Cost %f" % ( -event.step, event.epoch, event.metrics[0]) - -if isinstance(event, fluid.EndEpochEvent): -avg_cost, acc = trainer.test( -reader=test_reader, feed_order=['img', 'label']) - -print("Test with Epoch %d, avg_cost: %s, acc: %s" % (event.epoch, avg_cost, acc)) - -# save parameters -trainer.save_params(params_dirname) -lists.append((event.epoch, avg_cost, acc)) + if isinstance(event, fluid.EndStepEvent): + if event.step % 100 == 0: + # event.metrics maps with train program return arguments. + # event.metrics[0] will yeild avg_cost and event.metrics[1] will yeild acc in this example. + print("Pass %d, Batch %d, Cost %f" % ( + event.step, event.epoch, event.metrics[0])) + + if isinstance(event, fluid.EndEpochEvent): + avg_cost, acc = trainer.test( + reader=test_reader, feed_order=['img', 'label']) + + print("Test with Epoch %d, avg_cost: %s, acc: %s" % (event.epoch, avg_cost, acc)) + + # save parameters + trainer.save_params(params_dirname) + lists.append((event.epoch, avg_cost, acc)) ``` `event_handler_plot` 可以用来在训练过程中画图如下: -![png](./image/train_and_test.png) +
+
+图7 训练结果 +
+ ```python from paddle.v2.plot import Ploter @@ -333,22 +327,22 @@ lists = [] # event_handler to plot a figure def event_handler_plot(event): -global step -if isinstance(event, fluid.EndStepEvent): -if step % 100 == 0: -# event.metrics maps with train program return arguments. -# event.metrics[0] will yeild avg_cost and event.metrics[1] will yeild acc in this example. -cost_ploter.append(train_title, step, event.metrics[0]) -cost_ploter.plot() -step += 1 -if isinstance(event, fluid.EndEpochEvent): -# save parameters -trainer.save_params(params_dirname) - -avg_cost, acc = trainer.test( -reader=test_reader, feed_order=['img', 'label']) -cost_ploter.append(test_title, step, avg_cost) -lists.append((event.epoch, avg_cost, acc)) + global step + if isinstance(event, fluid.EndStepEvent): + if step % 100 == 0: + # event.metrics maps with train program return arguments. + # event.metrics[0] will yeild avg_cost and event.metrics[1] will yeild acc in this example. + cost_ploter.append(train_title, step, event.metrics[0]) + cost_ploter.plot() + step += 1 + if isinstance(event, fluid.EndEpochEvent): + # save parameters + trainer.save_params(params_dirname) + + avg_cost, acc = trainer.test( + reader=test_reader, feed_order=['img', 'label']) + cost_ploter.append(test_title, step, avg_cost) + lists.append((event.epoch, avg_cost, acc)) ``` #### 开始训练 @@ -359,10 +353,10 @@ lists.append((event.epoch, avg_cost, acc)) ```python trainer.train( -num_epochs=5, -event_handler=event_handler, -reader=train_reader, -feed_order=['img', 'label']) + num_epochs=5, + event_handler=event_handler, + reader=train_reader, + feed_order=['img', 'label']) ``` 训练过程是完全自动的,event_handler里打印的日志类似如下所示: @@ -395,11 +389,11 @@ Test with Epoch 0, avg_cost: 0.053097883707459624, acc: 0.9822850318471338 ```python inferencer = fluid.Inferencer( -# infer_func=softmax_regression, # uncomment for softmax regression -# infer_func=multilayer_perceptron, # uncomment for MLP -infer_func=convolutional_neural_network, # uncomment for LeNet5 -param_path=params_dirname, -place=place) + # infer_func=softmax_regression, # uncomment for softmax regression + # infer_func=multilayer_perceptron, # uncomment for MLP + infer_func=convolutional_neural_network, # uncomment for LeNet5 + param_path=params_dirname, + place=place) ``` ### 生成预测输入数据 @@ -412,11 +406,11 @@ import os import numpy as np from PIL import Image def load_image(file): -im = Image.open(file).convert('L') -im = im.resize((28, 28), Image.ANTIALIAS) -im = np.array(im).reshape(1, 1, 28, 28).astype(np.float32) -im = im / 255.0 * 2.0 - 1.0 -return im + im = Image.open(file).convert('L') + im = im.resize((28, 28), Image.ANTIALIAS) + im = np.array(im).reshape(1, 1, 28, 28).astype(np.float32) + im = im / 255.0 * 2.0 - 1.0 + return im cur_dir = cur_dir = os.getcwd() img = load_image(cur_dir + '/image/infer_3.png') @@ -429,7 +423,7 @@ img = load_image(cur_dir + '/image/infer_3.png') ```python results = inferencer.infer({'img': img}) lab = np.argsort(results) # probs and lab are the results of one batch data -print "Label of image/infer_3.png is: %d" % lab[0][0][-1] +print ("Inference result of image/infer_3.png is: %d" % lab[0][0][-1]) ``` ## 总结 diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn.png deleted file mode 100644 index 3f5cdaacdc6acce41c5c6c99649be46685cf9903..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn_train_log.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn_train_log.png deleted file mode 100644 index 65bd17eacd41bbdbdb042bd1ba366eb53663b410..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn_train_log.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/infer_3.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/infer_3.png deleted file mode 100644 index 030cd60d3b4af9aecd4941204da4ad15f6e1189f..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/infer_3.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/max_pooling.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/max_pooling.png deleted file mode 100644 index 90b02fa2a735cfcc9efb2de90906325dedcb358c..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/max_pooling.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp.png deleted file mode 100644 index 9f4d26cd8da32201d0a5e9c72d466301dd2b42a1..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp_train_log.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp_train_log.png deleted file mode 100644 index f5a478fdc24f29c17555a2f1451f3f5a079faed9..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp_train_log.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mnist_example_image.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mnist_example_image.png deleted file mode 100644 index 4edd7cabf8a2282f6392ac1421c7ca4afb288589..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mnist_example_image.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_regression.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_regression.png deleted file mode 100644 index 40b98298288b9c406fce1cbca9c913753020a94d..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_regression.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_train_log.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_train_log.png deleted file mode 100644 index 47204941af7f22e68386a70a06ec4f122b83e262..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_train_log.png and /dev/null differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/train_and_test.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/train_and_test.png deleted file mode 100644 index 5cb87b450d0398bcfaec0e647c362052069797e7..0000000000000000000000000000000000000000 Binary files a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/train_and_test.png and /dev/null differ diff --git a/doc/fluid/new_docs/advanced_usage/deploy/build_and_install_lib_cn.rst b/doc/fluid/new_docs/user_guides/howto/inference/build_and_install_lib_cn.rst similarity index 100% rename from doc/fluid/new_docs/advanced_usage/deploy/build_and_install_lib_cn.rst rename to doc/fluid/new_docs/user_guides/howto/inference/build_and_install_lib_cn.rst diff --git a/doc/fluid/new_docs/user_guides/howto/inference/index.rst b/doc/fluid/new_docs/user_guides/howto/inference/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..45e1a2883773b92ed47ef8d51417bbdcd060b4ec --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/inference/index.rst @@ -0,0 +1,11 @@ +############ +模型预测部署 +############ + +PaddlePaddle Fluid 提供了 C++ API 来支持模型的部署上线 + +.. toctree:: + :maxdepth: 2 + + build_and_install_lib_cn.rst + native_infer.rst diff --git a/doc/fluid/new_docs/advanced_usage/deploy/native_infer.rst b/doc/fluid/new_docs/user_guides/howto/inference/native_infer.rst similarity index 94% rename from doc/fluid/new_docs/advanced_usage/deploy/native_infer.rst rename to doc/fluid/new_docs/user_guides/howto/inference/native_infer.rst index aa9377c112856693cda72779bd399f2415d716f0..21a6fe5cf54d0c0c760ade4ba602024ffa29675f 100644 --- a/doc/fluid/new_docs/advanced_usage/deploy/native_infer.rst +++ b/doc/fluid/new_docs/user_guides/howto/inference/native_infer.rst @@ -4,11 +4,13 @@ Paddle 预测 API 为了更简单方便的预测部署,Fluid 提供了一套高层 API 用来隐藏底层不同的优化实现。 -`预测库相关代码 `__ +`预测库相关代码 `__ 包括 - 头文件 ``paddle_inference_api.h`` 定义了所有的接口 - 库文件\ ``libpaddle_fluid.so`` 或 ``libpaddle_fluid.a`` +- 库文件 ``libpaddle_inference_api.so`` 或 + ``libpaddle_inference_api.a`` 编译和依赖可以参考 :ref:`install_or_build_cpp_inference_lib` 。 @@ -95,7 +97,8 @@ engine CHECK(predictor->Run(slots, &outputs)); // 获取 outputs ... -编译时,联编 ``libpaddle_fluid.a/.so`` 即可。 +编译时,联编 ``libpaddle_fluid.a/.so`` 和 +``libpaddle_inference_api.a/.so`` 便可。 详细代码参考 ------------ diff --git a/doc/fluid/new_docs/user_guides/index.rst b/doc/fluid/new_docs/user_guides/index.rst index 453cb71cfdf72e031ce0f0517e2db936eca38dfc..377631109d8f65c149b12cd2a0e4da920fdf4def 100644 --- a/doc/fluid/new_docs/user_guides/index.rst +++ b/doc/fluid/new_docs/user_guides/index.rst @@ -15,4 +15,5 @@ howto/training/index howto/debug/index howto/evaluation/index + howto/inference/index models/index.rst diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index e963902a50200b785284e8f233fcca1abf459140..f7cc8ca9b7a75a56414656a17ee3f0b53fd0e128 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -55,9 +55,10 @@ paddle.fluid.Inferencer.__init__ ArgSpec(args=['self', 'infer_func', 'param_path paddle.fluid.Inferencer.infer ArgSpec(args=['self', 'inputs', 'return_numpy'], varargs=None, keywords=None, defaults=(True,)) paddle.fluid.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) -paddle.fluid.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.DistributeTranspiler.get_pserver_programs ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) +paddle.fluid.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.DistributeTranspiler.get_trainer_program ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) -paddle.fluid.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True)) +paddle.fluid.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode', 'startup_program'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True, None)) paddle.fluid.InferenceTranspiler.__init__ paddle.fluid.InferenceTranspiler.transpile ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.memory_optimize ArgSpec(args=['input_program', 'skip_opt_set', 'print_log', 'level'], varargs=None, keywords=None, defaults=(None, False, 0)) @@ -113,6 +114,7 @@ paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)) paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)) paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None)) +paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None)) paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) @@ -190,7 +192,7 @@ paddle.fluid.layers.argsort ArgSpec(args=['input', 'axis', 'name'], varargs=None paddle.fluid.layers.ones ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,)) paddle.fluid.layers.zeros ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,)) paddle.fluid.layers.reverse ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=None) -paddle.fluid.layers.While.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.While.__init__ ArgSpec(args=['self', 'cond', 'is_test', 'name'], varargs=None, keywords=None, defaults=(False, None)) paddle.fluid.layers.While.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.Switch.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.Switch.case ArgSpec(args=['self', 'condition'], varargs=None, keywords=None, defaults=None) @@ -329,9 +331,10 @@ paddle.fluid.contrib.BeamSearchDecoder.update_array ArgSpec(args=['self', 'array paddle.fluid.contrib.memory_usage ArgSpec(args=['program', 'batch_size'], varargs=None, keywords=None, defaults=None) paddle.fluid.transpiler.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.transpiler.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) -paddle.fluid.transpiler.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.transpiler.DistributeTranspiler.get_pserver_programs ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) +paddle.fluid.transpiler.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.transpiler.DistributeTranspiler.get_trainer_program ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) -paddle.fluid.transpiler.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True)) +paddle.fluid.transpiler.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode', 'startup_program'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True, None)) paddle.fluid.transpiler.InferenceTranspiler.__init__ paddle.fluid.transpiler.InferenceTranspiler.transpile ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.transpiler.memory_optimize ArgSpec(args=['input_program', 'skip_opt_set', 'print_log', 'level'], varargs=None, keywords=None, defaults=(None, False, 0)) diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc index bc61b0eacbf6c8a1fd4487ad5a442fed1b536345..0bfff745493d069e948e6d277ec2bbfb0673a70b 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc @@ -625,19 +625,11 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph, void MultiDevSSAGraphBuilder::CreateScaleLossGradOp( ir::Graph *result, const std::string &loss_grad_name) const { for (size_t i = 0; i < places_.size(); ++i) { -// Insert ScaleCost OpHandle -#ifdef PADDLE_WITH_CUDA - auto *communication_dev_ctx = - nccl_ctxs_ ? nccl_ctxs_->DevCtx(places_[i]) - : platform::DeviceContextPool::Instance().Get(places_[i]); -#else - auto *communication_dev_ctx = - platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); -#endif + // Insert ScaleCost OpHandle + auto *dev_ctx = platform::DeviceContextPool::Instance().Get(places_[i]); auto *op_handle = new ScaleLossGradOpHandle( result->CreateEmptyNode("scale_loss_grad", ir::Node::Type::kOperation), - local_scopes_.size(), local_scopes_[i], places_[i], - communication_dev_ctx); + local_scopes_.size(), local_scopes_[i], places_[i], dev_ctx); result->Get(kGraphOps).emplace_back(op_handle); // FIXME: Currently ScaleLossGradOp only use device_count as scale @@ -744,7 +736,7 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result, .emplace(varname, op_dev_id); } } else { - PADDLE_ENFORCE( + PADDLE_THROW( "the distribute training related op should be in [split_byref, " "concat]."); } @@ -754,17 +746,26 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result, node->Op()->Type()); CreateComputationalOp(result, node, op_dev_id); - if (node->Op()->Type() == "concat") { - ConnectOp(result, result->Get(kGraphOps).back().get(), - "fetch_barrier"); +} + +void SetOpInputsAllPlaces(ir::Graph *result, ir::Node *node, int num_places) { + auto *op_handle = result->Get(kGraphOps).back().get(); + for (ir::Node *input : node->inputs) { + VarHandle *var = nullptr; + for (int place_offset = 0; place_offset < num_places; ++place_offset) { + auto &var_holders = result->Get(kGraphVars)[place_offset]; + auto &var_holder = var_holders[input->Name()]; + if (!var_holder.empty()) { + var = var_holder.rbegin()->get(); + op_handle->AddInput(var); + } + } } } // Create RPC related op handles that connects its in ops and out ops. void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result, ir::Node *node) const { - // FIXME(typhoonzero): Cleanup this deps for both sync mode and async mode - // put them into transpiler. int op_dev_id = -1; if (node->Op()->Type() == "send") { // TODO(paddle-dev): getting the first var is not safe. @@ -799,8 +800,6 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result, } auto recv_param_grad = boost::get>( node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName())); - // FIXME(typhoonzero): assume each recv op output one param - // Use the same place as send. if (recv_param_grad.size() == 2U) { op_dev_id = GetVarDeviceID(*result, recv_param_grad[1]); VLOG(10) << "recv param " << recv_param_grad[0] @@ -814,34 +813,44 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result, .emplace(varname, op_dev_id); } } else { - // send_barrier and fetch_barrier op can be scheduled on device 0 + // send_barrier, fetch_barrier will run on place 0; op_dev_id = 0; } PADDLE_ENFORCE(op_dev_id != -1, "can not find the right place for rpc op: %s", node->Op()->Type()); - result->Get(kGraphOps).emplace_back(new RPCOpHandle( result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id], node->Op()->Type(), places_[op_dev_id])); - // TODO(panyx0718): This might not be needed anymore. - if (node->Op()->Type() == "send_barrier") { - ConnectOp(result, result->Get(kGraphOps).back().get(), "send"); - } else if (node->Op()->Type() == "recv") { - ConnectOp(result, result->Get(kGraphOps).back().get(), - "send_barrier"); - } else if (node->Op()->Type() == "fetch_barrier") { - ConnectOp(result, result->Get(kGraphOps).back().get(), "recv"); - } else if (node->Op()->Type() == "send") { - // do nothing + if (node->Op()->Type() == "send") { + CreateOpHandleIOs(result, node, op_dev_id); } else { - PADDLE_THROW( - "rpc op should be in [" - "send, send_barrier. recv, fetch_barrier]"); - } + // send_barrier, recv, fetch_barrier's inputs are deps var, get them from + // all places + auto p = places_[op_dev_id]; + auto *op_handle = result->Get(kGraphOps).back().get(); + op_handle->SetDeviceContext(p, + platform::DeviceContextPool::Instance().Get(p)); - CreateOpHandleIOs(result, node, op_dev_id); + SetOpInputsAllPlaces(result, node, places_.size()); + for (ir::Node *output : node->outputs) { + int outvar_dev_id = op_dev_id; + if (node->Op()->Type() == "fetch_barrier") { + outvar_dev_id = GetVarDeviceID(*result, output->Name()); + PADDLE_ENFORCE_NE(outvar_dev_id, -1); + } + p = places_[outvar_dev_id]; + ir::Node *new_node = nullptr; + if (output->Var()) { + new_node = result->CreateVarNode(output->Var()); + } else { + new_node = + result->CreateEmptyNode(output->Name(), ir::Node::Type::kVariable); + } + CreateOpOutput(result, op_handle, new_node, p, outvar_dev_id); + } + } } bool MultiDevSSAGraphBuilder::IsScaleLossOp(ir::Node *node) const { diff --git a/paddle/fluid/framework/ir/graph.cc b/paddle/fluid/framework/ir/graph.cc index 2a6bf4ac230df81b38751000bf4b663f24984db3..39b0f2f038380b4631728e28031511205c4b40f2 100644 --- a/paddle/fluid/framework/ir/graph.cc +++ b/paddle/fluid/framework/ir/graph.cc @@ -132,63 +132,6 @@ Graph::Graph(const ProgramDesc &program) : program_(program) { } } - std::vector send_ops; - ir::Node *send_bar = nullptr; - std::vector recv_ops; - ir::Node *fetch_bar = nullptr; - for (ir::Node *node : Nodes()) { - if (node->Name() == "send") { - send_ops.push_back(node); - } else if (node->Name() == "send_barrier") { - PADDLE_ENFORCE(!send_bar, "only has one send barrier"); - send_bar = node; - } else if (node->Name() == "recv") { - recv_ops.push_back(node); - } else if (node->Name() == "fetch_barrier") { - PADDLE_ENFORCE(!fetch_bar, "only has one fetch barrier"); - fetch_bar = node; - } - } - if (send_bar) { - for (ir::Node *send : send_ops) { - ir::Node *dep_var = CreateControlDepVar(); - send->outputs.push_back(dep_var); - dep_var->inputs.push_back(send); - send_bar->inputs.push_back(dep_var); - dep_var->outputs.push_back(send_bar); - } - for (ir::Node *recv : recv_ops) { - ir::Node *dep_var = CreateControlDepVar(); - recv->inputs.push_back(dep_var); - dep_var->outputs.push_back(recv); - send_bar->outputs.push_back(dep_var); - dep_var->inputs.push_back(send_bar); - } - } - if (fetch_bar) { - for (ir::Node *recv : recv_ops) { - ir::Node *dep_var = CreateControlDepVar(); - recv->outputs.push_back(dep_var); - dep_var->inputs.push_back(recv); - fetch_bar->inputs.push_back(dep_var); - dep_var->outputs.push_back(fetch_bar); - } - } - - std::vector send_vars = FindDistTrainSendVars(send_ops); - std::vector recv_vars = FindDistTrainRecvVars(recv_ops); - for (ir::Node *node : Nodes()) { - if (IsDistTrainOp(node, send_vars, recv_vars)) { - if (fetch_bar && node->Name() == "concat") { - ir::Node *dep_var = CreateControlDepVar(); - fetch_bar->outputs.push_back(dep_var); - dep_var->inputs.push_back(fetch_bar); - node->inputs.push_back(dep_var); - dep_var->outputs.push_back(node); - } - } - } - /** * We should handle write after read(WAR) and write after write(WAW) here. * Because some of the operators of the program can be executed parallelly. diff --git a/paddle/fluid/inference/analysis/analyzer.cc b/paddle/fluid/inference/analysis/analyzer.cc index 9318f1089781b30468cf4d3c7151d0dd26e50a9c..e51d6cfeb931d50a9a573df29c916ebd3da403d1 100644 --- a/paddle/fluid/inference/analysis/analyzer.cc +++ b/paddle/fluid/inference/analysis/analyzer.cc @@ -24,7 +24,7 @@ namespace paddle { -DEFINE_bool(inference_analysis_enable_tensorrt_subgraph_engine, true, +DEFINE_bool(inference_analysis_enable_tensorrt_subgraph_engine, false, "Enable subgraph to TensorRT engine for acceleration"); DEFINE_string(inference_analysis_graphviz_log_root, "./", @@ -44,7 +44,8 @@ class DfgPassManagerImpl final : public DfgPassManager { if (FLAGS_inference_analysis_enable_tensorrt_subgraph_engine) { auto trt_teller = [&](const Node* node) { std::unordered_set teller_set( - {"elementwise_add", "mul", "conv2d", "pool2d", "relu", "softmax"}); + {"elementwise_add", "mul", "conv2d", "pool2d", "relu", "softmax", + "depthwise_conv2d", "batch_norm", "concat"}); if (!node->IsFunction()) return false; const auto* func = static_cast(node); diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc index 18c32fa09199003f17183207828cdfe4e627ae1a..ce0639a6162da6347ed130ecb1586c9a2d4071d5 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc @@ -23,9 +23,6 @@ namespace paddle { namespace inference { -DEFINE_int32(tensorrt_max_batchsize, 3, "TensorRT maximum batch size"); -DEFINE_int32(tensorrt_workspace_size, 2048, "TensorRT workspace size"); - namespace analysis { using framework::proto::ProgramDesc; @@ -52,7 +49,6 @@ bool DataFlowGraphToFluidPass::Initialize(Argument *argument) { bool DataFlowGraphToFluidPass::Finalize() { return true; } void DataFlowGraphToFluidPass::Run(DataFlowGraph *graph) { - FilterRedundantOutputOfSubGraph(graph); LOG(INFO) << "graph.inputs " << graph->inputs.size(); for (auto &node : GraphTraits(graph).nodes_in_TS()) { if (node.deleted()) continue; @@ -191,8 +187,6 @@ void CreateTrtEngineOp(Node *node, const DataFlowGraph &graph, // Set attrs SetAttr(desc.Proto(), "subgraph", block->SerializeAsString()); SetAttr(desc.Proto(), "engine_uniq_key", "trt-" + std::to_string(counter++)); - SetAttr(desc.Proto(), "max_batch", FLAGS_tensorrt_max_batchsize); - SetAttr(desc.Proto(), "max_workspace", FLAGS_tensorrt_workspace_size); SetAttr(desc.Proto(), "parameters", ExtractParameters(graph.nodes.nodes())); SetAttr(desc.Proto(), "output_name_mapping", output_mapping); node->SetPbMsg(desc.Proto()->SerializeAsString()); diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h index 59c47365aa6c8ad5886c4515850d264f69cc4670..0c9a8a0b7cae17bf2eaa714348ea1c9b5e43611b 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h @@ -27,9 +27,6 @@ namespace paddle { namespace inference { -DECLARE_int32(tensorrt_max_batchsize); -DECLARE_int32(tensorrt_workspace_size); - namespace analysis { class DataFlowGraphToFluidPass final : public DataFlowGraphPass { public: diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc index 511631d3e067f14bc1230d9e4b4d92dbe604e1d4..16d82b5aa1acaf87d1cd78ad5b79faa65143ad7d 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc @@ -92,6 +92,7 @@ void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) { auto *in = graph->nodes.GetMutable(var2id.at(in_var.arguments(k))); in->outlinks.push_back(o); o->inlinks.push_back(in); + unique_written_vars.insert(in); } } for (int j = 0; j < op.outputs_size(); j++) { @@ -112,7 +113,6 @@ void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) { } out->inlinks.push_back(o); o->outlinks.push_back(out); - unique_written_vars.insert(out); } } } diff --git a/paddle/fluid/inference/analysis/subgraph_splitter.cc b/paddle/fluid/inference/analysis/subgraph_splitter.cc index 80809d4c43ca08298bad25cf614dcb4117d3f99a..9146c0e45e77b5f120d3be622f74e3008bca2b6f 100644 --- a/paddle/fluid/inference/analysis/subgraph_splitter.cc +++ b/paddle/fluid/inference/analysis/subgraph_splitter.cc @@ -153,6 +153,7 @@ void SubGraphFuse::ReplaceNodesWithSubGraphs() { inlink_or_outlink_cleaner(o->inlinks); } } + FilterRedundantOutputOfSubGraph(graph_); } } // namespace analysis diff --git a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc index 45b5a7638b7dc6a54bbd905766fd5c284cb6aea1..5967402055c61dd480055497640a16ab7e94a746 100644 --- a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc @@ -15,6 +15,7 @@ #include "paddle/fluid/inference/analysis/analyzer.h" #include "paddle/fluid/inference/api/api_impl.h" #include "paddle/fluid/inference/api/paddle_inference_api.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/utils/singleton.h" #include "paddle/fluid/operators/tensorrt_engine_op.h" @@ -32,7 +33,9 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor { bool Init(const std::shared_ptr& parent_scope) { VLOG(3) << "Predictor::init()"; - + FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = true; + FLAGS_tensorrt_max_batch_size = config_.max_batch_size; + FLAGS_tensorrt_workspace_size = config_.workspace_size; if (config_.use_gpu) { place_ = paddle::platform::CUDAPlace(config_.device); } else { @@ -150,3 +153,13 @@ CreatePaddlePredictor( } } // namespace paddle + +USE_TRT_CONVERTER(elementwise_add_weight); +USE_TRT_CONVERTER(mul); +USE_TRT_CONVERTER(conv2d); +USE_TRT_CONVERTER(relu); +USE_TRT_CONVERTER(fc); +USE_TRT_CONVERTER(pool2d); +USE_TRT_CONVERTER(softmax); +USE_TRT_CONVERTER(batch_norm); +USE_TRT_CONVERTER(concat); diff --git a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc index fcbf9b89d608e7961e3ef81ac1c70e083dae1cc0..b3892b5dd89fef055f87f9170a0e1f24cc93d0a0 100644 --- a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc @@ -37,6 +37,7 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) { config1.use_gpu = true; config1.fraction_of_gpu_memory = 0.3; config1.device = 0; + config1.max_batch_size = 10; auto predictor0 = CreatePaddlePredictor(config0); diff --git a/paddle/fluid/inference/api/paddle_inference_api.h b/paddle/fluid/inference/api/paddle_inference_api.h index 794534467be066e91db2b4c204913ab2cf12dbfd..da6c2cfc21809aa08bba79880dc898bb056d16a0 100644 --- a/paddle/fluid/inference/api/paddle_inference_api.h +++ b/paddle/fluid/inference/api/paddle_inference_api.h @@ -137,6 +137,14 @@ struct AnakinConfig : public PaddlePredictor::Config { struct TensorRTConfig : public NativeConfig { // Determine whether a subgraph will be executed by TRT. int min_subgraph_size{1}; + // While TensorRT allows an engine optimized for a given max batch size + // to run at any smaller size, the performance for those smaller + // sizes may not be as well-optimized. Therefore, Max batch is best + // equivalent to the runtime batch size. + int max_batch_size{1}; + // For workspace_size, refer it from here: + // https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#troubleshooting + int workspace_size{1 << 30}; }; // A factory to help create different predictors. diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt index b52d083f280e5e7713600a7b748dedd37aca0a1e..a610687a5b11999a7cb7426dbe961e5972ee1746 100644 --- a/paddle/fluid/inference/tensorrt/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/CMakeLists.txt @@ -1,4 +1,4 @@ -nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto) +nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto device_context) nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine) add_subdirectory(convert) diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index 6863b035d8cd9dfb21aed3947226a796778912a4..9d7be2d03cf7bb12afe7e52d9630f184d689dc25 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,7 +1,7 @@ # Add TRT tests nv_library(tensorrt_converter SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc -activation_op.cc softmax_op.cc +batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc DEPS tensorrt_engine operator scope framework_proto op_registry) nv_test(test_op_converter SRCS test_op_converter.cc DEPS @@ -18,9 +18,12 @@ nv_test(test_trt_conv_op SRCS test_conv2d_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine conv_op SERIAL) nv_test(test_trt_pool2d_op SRCS test_pool2d_op.cc pool2d_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine pool_op SERIAL) - nv_test(test_trt_elementwise_op SRCS test_elementwise_op.cc elementwise_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine elementwise_add_op SERIAL) - nv_test(test_trt_softmax_op SRCS test_softmax_op.cc softmax_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine softmax_op SERIAL) +nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL) + +nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL) diff --git a/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc b/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..94f8b0ae5606d39a722ffe28501645c9b6fc5d2e --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc @@ -0,0 +1,136 @@ +/* Copyright (c) 2018 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. */ + +#include +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +class BatchNormOpConverter : public OpConverter { + public: + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + LOG(INFO) << "convert a fluid batch norm op to tensorrt batch_norm"; + + framework::OpDesc op_desc(op, nullptr); + PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); + PADDLE_ENFORCE_EQ(op_desc.Input("Bias").size(), 1); // Bias is a weight + PADDLE_ENFORCE_EQ(op_desc.Input("Mean").size(), 1); // Mean is a weight + PADDLE_ENFORCE_EQ(op_desc.Input("Scale").size(), 1); // Scale is a weight + PADDLE_ENFORCE_EQ(op_desc.Input("Variance").size(), + 1); // Variance is a weight + PADDLE_ENFORCE_EQ(op_desc.Output("Y").size(), 1); + + auto* X = engine_->GetITensor(op_desc.Input("X").front()); + // Declare weights + auto* Bias_v = scope.FindVar(op_desc.Input("Bias").front()); + auto* Mean_v = scope.FindVar(op_desc.Input("Mean").front()); + auto* Scale_v = scope.FindVar(op_desc.Input("Scale").front()); + auto* Variance_v = scope.FindVar(op_desc.Input("Variance").front()); + const float eps = boost::get(op_desc.GetAttr("epsilon")); + + PADDLE_ENFORCE_NOT_NULL(Bias_v); + PADDLE_ENFORCE_NOT_NULL(Mean_v); + PADDLE_ENFORCE_NOT_NULL(Scale_v); + PADDLE_ENFORCE_NOT_NULL(Variance_v); + + // get tensor + auto* Bias_t = Bias_v->GetMutable(); + auto* Mean_t = Mean_v->GetMutable(); + auto* Scale_t = Scale_v->GetMutable(); + auto* Variance_t = Variance_v->GetMutable(); + + // create temp tensor for weights + framework::LoDTensor bias_tensor; + framework::LoDTensor mean_tensor; + framework::LoDTensor scale_tensor; + framework::LoDTensor variance_tensor; + + bias_tensor.Resize(Bias_t->dims()); + mean_tensor.Resize(Mean_t->dims()); + scale_tensor.Resize(Scale_t->dims()); + variance_tensor.Resize(Variance_t->dims()); + + platform::CPUPlace cpu_place; + // copy data from gpu to cpu + TensorCopySync((*Bias_t), cpu_place, &bias_tensor); + TensorCopySync((*Mean_t), cpu_place, &mean_tensor); + TensorCopySync((*Scale_t), cpu_place, &scale_tensor); + TensorCopySync((*Variance_t), cpu_place, &variance_tensor); + + auto* bias_data = bias_tensor.mutable_data(platform::CPUPlace()); + auto* mean_data = mean_tensor.mutable_data(platform::CPUPlace()); + auto* scale_data = scale_tensor.mutable_data(platform::CPUPlace()); + auto* variance_data = + variance_tensor.mutable_data(platform::CPUPlace()); + + std::unique_ptr combile_scale_tensor( + new framework::LoDTensor()); + std::unique_ptr combile_bias_tensor( + new framework::LoDTensor()); + + combile_scale_tensor->Resize(scale_tensor.dims()); + combile_bias_tensor->Resize(bias_tensor.dims()); + + auto* combile_scale_data = + combile_scale_tensor->mutable_data(platform::CPUPlace()); + auto* combile_bias_data = + combile_bias_tensor->mutable_data(platform::CPUPlace()); + + size_t ele_num = combile_scale_tensor->memory_size() / sizeof(float); + + for (size_t i = 0; i < ele_num; i++) { + float scale = scale_data[i]; + float bias = bias_data[i]; + float mean = mean_data[i]; + float variance = variance_data[i]; + combile_scale_data[i] = scale / sqrtf(variance + eps); + combile_bias_data[i] = bias - mean * combile_scale_data[i]; + } + + TensorRTEngine::Weight scale_weights{ + nvinfer1::DataType::kFLOAT, static_cast(combile_scale_data), + combile_scale_tensor->memory_size() / sizeof(float)}; + TensorRTEngine::Weight shift_weights{ + nvinfer1::DataType::kFLOAT, static_cast(combile_bias_data), + combile_bias_tensor->memory_size() / sizeof(float)}; + TensorRTEngine::Weight power_weights{nvinfer1::DataType::kFLOAT, nullptr, + 0}; + + nvinfer1::IScaleLayer* layer = + TRT_ENGINE_ADD_LAYER(engine_, Scale, *const_cast(X), + nvinfer1::ScaleMode::kCHANNEL, shift_weights.get(), + scale_weights.get(), power_weights.get()); + + auto output_name = op_desc.Output("Y").front(); + engine_->weight_map[op_desc.Input("Bias").front()] = + std::move(combile_bias_tensor); + engine_->weight_map[op_desc.Input("Scale").front()] = + std::move(combile_scale_tensor); + + engine_->SetITensor(output_name, layer->getOutput(0)); + + if (test_mode) { + engine_->DeclareOutput(output_name); + } + } +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +REGISTER_TRT_OP_CONVERTER(batch_norm, BatchNormOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/concat_op.cc b/paddle/fluid/inference/tensorrt/convert/concat_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..bb9627bf957b63993b2c8d23e7ec8122eb004eaf --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/concat_op.cc @@ -0,0 +1,57 @@ +/* Copyright (c) 2018 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. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +/* + * MulOp, IMatrixMultiplyLayer in TRT. This Layer doesn't has weights. + */ +class ConcatOpConverter : public OpConverter { + public: + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + VLOG(4) << "convert a fluid mul op to tensorrt mul layer without bias"; + + framework::OpDesc op_desc(op, nullptr); + // Declare inputs + std::vector itensors; + for (auto& input_name : op_desc.Input("X")) { + itensors.push_back(engine_->GetITensor(input_name)); + } + int axis = boost::get(op_desc.GetAttr("axis")); + PADDLE_ENFORCE(axis > 0, + "The axis attr of Concat op should be large than 0 for trt"); + + auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Concatenation, itensors.data(), + itensors.size()); + axis = axis - 1; // Remove batch dim + layer->setAxis(axis); + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { // the test framework can not determine which is the + // output, so place the declaration inside. + engine_->DeclareOutput(output_name); + } + } +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +REGISTER_TRT_OP_CONVERTER(concat, ConcatOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc index dba1d50b2d1c487ced8e6ca51f2d257641ad5fc7..841a95db38ce7cf0cb5961ff04cb569ee2633e6f 100644 --- a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc @@ -35,12 +35,20 @@ class Conv2dOpConverter : public OpConverter { auto* Y_v = scope.FindVar(op_desc.Input("Filter").front()); PADDLE_ENFORCE_NOT_NULL(Y_v); auto* Y_t = Y_v->GetMutable(); - auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); - PADDLE_ENFORCE_EQ(Y_t->dims().size(), 4UL); - const int n_output = Y_t->dims()[0]; - const int filter_h = Y_t->dims()[2]; - const int filter_w = Y_t->dims()[3]; + platform::CPUPlace cpu_place; + std::unique_ptr weight_tensor( + new framework::LoDTensor()); + weight_tensor->Resize(Y_t->dims()); + TensorCopySync((*Y_t), cpu_place, weight_tensor.get()); + + auto* weight_data = + weight_tensor->mutable_data(platform::CPUPlace()); + + PADDLE_ENFORCE_EQ(weight_tensor->dims().size(), 4UL); + const int n_output = weight_tensor->dims()[0]; + const int filter_h = weight_tensor->dims()[2]; + const int filter_w = weight_tensor->dims()[3]; const int groups = boost::get(op_desc.GetAttr("groups")); const std::vector dilations = @@ -57,7 +65,7 @@ class Conv2dOpConverter : public OpConverter { TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT, static_cast(weight_data), - Y_t->memory_size() / sizeof(float)}; + weight_tensor->memory_size() / sizeof(float)}; TensorRTEngine::Weight bias{nvinfer1::DataType::kFLOAT, nullptr, 0}; auto* layer = TRT_ENGINE_ADD_LAYER( @@ -70,6 +78,8 @@ class Conv2dOpConverter : public OpConverter { layer->setNbGroups(groups); auto output_name = op_desc.Output("Output").front(); + engine_->weight_map[op_desc.Input("Filter").front()] = + std::move(weight_tensor); engine_->SetITensor(output_name, layer->getOutput(0)); if (test_mode) { engine_->DeclareOutput(output_name); diff --git a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc index 3744550f60a1696aedd8a3ecd24f1b21d22325b9..60a72b4eb5c75b5cd12305f13763a9a1a567213f 100644 --- a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc @@ -12,7 +12,6 @@ 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. */ -#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" namespace paddle { @@ -40,10 +39,17 @@ class ElementwiseWeightOpConverter : public OpConverter { auto* Y_v = scope.FindVar(op_desc.Input("Y").front()); PADDLE_ENFORCE_NOT_NULL(Y_v); auto* Y_t = Y_v->GetMutable(); - auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); + + platform::CPUPlace cpu_place; + std::unique_ptr weight_tensor( + new framework::LoDTensor()); + weight_tensor->Resize(Y_t->dims()); + TensorCopySync((*Y_t), cpu_place, weight_tensor.get()); + auto* weight_data = + weight_tensor->mutable_data(platform::CPUPlace()); auto scale_mode = nvinfer1::ScaleMode::kELEMENTWISE; - std::vector dims_y = framework::vectorize2int(Y_t->dims()); + std::vector dims_y = framework::vectorize2int(weight_tensor->dims()); if (static_cast(dims_y.size()) == dims_x.nbDims + 1) { if (dims_y[0] == 1) dims_y.erase(dims_y.begin()); } @@ -70,9 +76,9 @@ class ElementwiseWeightOpConverter : public OpConverter { PADDLE_THROW("TensorRT unsupported weight Shape for Elementwise op!"); } - TensorRTEngine::Weight shift_weights{nvinfer1::DataType::kFLOAT, - static_cast(weight_data), - Y_t->memory_size() / sizeof(float)}; + TensorRTEngine::Weight shift_weights{ + nvinfer1::DataType::kFLOAT, static_cast(weight_data), + weight_tensor->memory_size() / sizeof(float)}; TensorRTEngine::Weight scale_weights{nvinfer1::DataType::kFLOAT, nullptr, 0}; TensorRTEngine::Weight power_weights{nvinfer1::DataType::kFLOAT, nullptr, @@ -82,6 +88,8 @@ class ElementwiseWeightOpConverter : public OpConverter { engine_, Scale, *const_cast(X), scale_mode, shift_weights.get(), scale_weights.get(), power_weights.get()); auto output_name = op_desc.Output("Out")[0]; + + engine_->weight_map[op_desc.Input("Y").front()] = std::move(weight_tensor); engine_->SetITensor(output_name, layer->getOutput(0)); if (test_mode) { // the test framework can not determine which is the // output, so place the declaration inside. diff --git a/paddle/fluid/inference/tensorrt/convert/fc_op.cc b/paddle/fluid/inference/tensorrt/convert/fc_op.cc index 39fe1f609d7b94638506877fc301f19ef33ec8ac..ad98d85aae9cf594922aca00c43718ccfbce2278 100644 --- a/paddle/fluid/inference/tensorrt/convert/fc_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/fc_op.cc @@ -12,12 +12,7 @@ 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. */ -#include "paddle/fluid/framework/eigen.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" -#include "paddle/fluid/inference/tensorrt/engine.h" -#include "paddle/fluid/platform/place.h" namespace paddle { namespace inference { @@ -73,19 +68,26 @@ class FcOpConverter : public OpConverter { auto* Y_t = Y_v->GetMutable(); // This may trigger a GPU->CPU copy, because TRT's weight can only be // assigned from CPU memory, that can't be avoided. - auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); - PADDLE_ENFORCE_EQ(Y_t->dims().size(), 2UL); // a matrix - size_t n_output = Y_t->dims()[1]; + platform::CPUPlace cpu_place; + framework::LoDTensor weight_tensor; + weight_tensor.Resize(Y_t->dims()); + TensorCopySync((*Y_t), cpu_place, &weight_tensor); - framework::LoDTensor tmp; - tmp.Resize(Y_t->dims()); - memcpy(tmp.mutable_data(platform::CPUPlace()), weight_data, + auto* weight_data = weight_tensor.mutable_data(platform::CPUPlace()); + + PADDLE_ENFORCE_EQ(weight_tensor.dims().size(), 2UL); // a matrix + size_t n_output = weight_tensor.dims()[1]; + + std::unique_ptr tmp(new framework::LoDTensor()); + tmp->Resize(weight_tensor.dims()); + + memcpy(tmp->mutable_data(platform::CPUPlace()), weight_data, Y_t->dims()[0] * Y_t->dims()[1] * sizeof(float)); TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT, static_cast(weight_data), Y_t->memory_size() / sizeof(float)}; TensorRTEngine::Weight tmp_weight(nvinfer1::DataType::kFLOAT, - static_cast(tmp.data()), + static_cast(tmp->data()), Y_t->memory_size() / sizeof(float)); weight.dims.assign({Y_t->dims()[0], Y_t->dims()[1]}); tmp_weight.dims = weight.dims; @@ -106,6 +108,7 @@ class FcOpConverter : public OpConverter { auto output_name = op_desc.Output("Out").front(); engine_->SetITensor(output_name, layer->getOutput(0)); + engine_->weight_map[op_desc.Input("Y").front()] = std::move(tmp); if (test_mode) { engine_->DeclareOutput(output_name); } diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h index 41faaf7212accaaec238062b1340e8da8fa6be33..d309d94c560f2b484fac6b6cd40cc2704d641069 100644 --- a/paddle/fluid/inference/tensorrt/convert/op_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -79,6 +79,14 @@ class OpConverter { it = Registry::Lookup("elementwise_" + op_type + "_tensor"); } + PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", + op_desc.Type()); + } + + if (op_desc.Type() == "depthwise_conv2d") { + it = Registry::Lookup("conv2d"); + PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", + op_desc.Type()); } if (!it) { diff --git a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc index 11cad95361867476c6f775af778015da37f1cfb1..73f1b28ddf73403862e55d102a259d7b6cf67b1f 100644 --- a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc @@ -33,6 +33,7 @@ class Pool2dOpConverter : public OpConverter { PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1); auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]); + bool global_pooling = boost::get(op_desc.GetAttr("global_pooling")); std::string pool_type = boost::get(op_desc.GetAttr("pooling_type")); std::vector ksize = @@ -42,7 +43,13 @@ class Pool2dOpConverter : public OpConverter { std::vector paddings = boost::get>(op_desc.GetAttr("paddings")); - const nvinfer1::DimsHW nv_ksize(ksize[0], ksize[1]); + nvinfer1::DimsHW nv_ksize(ksize[0], ksize[1]); + if (global_pooling == true) { + nvinfer1::Dims input_shape = input1->getDimensions(); + int nbDims = input_shape.nbDims; + nv_ksize.d[0] = input_shape.d[nbDims - 2]; + nv_ksize.d[1] = input_shape.d[nbDims - 1]; + } const nvinfer1::DimsHW nv_strides(strides[0], strides[1]); const nvinfer1::DimsHW nv_paddings(paddings[0], paddings[1]); diff --git a/paddle/fluid/inference/tensorrt/convert/test_batch_norm_op.cc b/paddle/fluid/inference/tensorrt/convert/test_batch_norm_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..41412cb079540da72760558379b158b6538aa6a8 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_batch_norm_op.cc @@ -0,0 +1,71 @@ +/* Copyright (c) 2018 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. */ + +#include +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(batch_norm_op, test) { + std::unordered_set parameters( + {"batch_norm_scale", "batch_norm_bias", "batch_norm_mean", + "batch_norm_variance"}); + framework::Scope scope; + TRTConvertValidation validator(5, parameters, scope, 1 << 15); + std::vector param_shape{2}; + + validator.DeclInputVar("batch_norm_X", nvinfer1::DimsCHW(2, 5, 5)); + validator.DeclParamVar("batch_norm_scale", param_shape); + validator.DeclParamVar("batch_norm_bias", param_shape); + validator.DeclParamVar("batch_norm_mean", param_shape); + validator.DeclParamVar("batch_norm_variance", param_shape); + validator.DeclOutputVar("batch_norm_Y", nvinfer1::DimsCHW(2, 5, 5)); + validator.DeclOutputVar("batch_norm_save_mean", param_shape); + validator.DeclOutputVar("batch_norm_save_variance", param_shape); + + // Prepare Op description + framework::OpDesc desc; + + desc.SetType("batch_norm"); + desc.SetInput("X", {"batch_norm_X"}); + desc.SetInput("Scale", {"batch_norm_scale"}); + desc.SetInput("Bias", {"batch_norm_bias"}); + desc.SetInput("Mean", {"batch_norm_mean"}); + desc.SetInput("Variance", {"batch_norm_variance"}); + desc.SetOutput("Y", {"batch_norm_Y"}); + desc.SetOutput("MeanOut", {"batch_norm_mean"}); + desc.SetOutput("VarianceOut", {"batch_norm_variance"}); + desc.SetOutput("SavedMean", {"batch_norm_save_mean"}); + desc.SetOutput("SavedVariance", {"batch_norm_save_variance"}); + + float eps = 1e-5f; + bool is_test = true; + desc.SetAttr("epsilon", eps); + desc.SetAttr("is_test", is_test); + + validator.SetOp(*desc.Proto()); + + std::unordered_set neglected_output = { + "batch_norm_save_mean", "batch_norm_save_variance", "batch_norm_mean", + "batch_norm_variance"}; + validator.Execute(3, neglected_output); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle +USE_OP(batch_norm); diff --git a/paddle/fluid/inference/tensorrt/convert/test_concat_op.cc b/paddle/fluid/inference/tensorrt/convert/test_concat_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..4f284a4db5758e072915d7fd0f16115b8a36ba8b --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_concat_op.cc @@ -0,0 +1,49 @@ +/* Copyright (c) 2018 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. */ + +#include +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(concat_op, test) { + std::unordered_set parameters({""}); + framework::Scope scope; + TRTConvertValidation validator(10, parameters, scope, 1000); + validator.DeclInputVar("concat_x1", nvinfer1::DimsCHW(10, 3, 1)); + validator.DeclInputVar("concat_x2", nvinfer1::DimsCHW(3, 3, 1)); + validator.DeclInputVar("concat_x3", nvinfer1::DimsCHW(7, 3, 1)); + validator.DeclOutputVar("concat_out", nvinfer1::DimsCHW(20, 3, 1)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("concat"); + desc.SetInput("X", {"concat_x1", "concat_x2", "concat_x3"}); + desc.SetOutput("Out", {"concat_out"}); + + int axis = 1; + desc.SetAttr("axis", axis); + + validator.SetOp(*desc.Proto()); + + validator.Execute(5); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle +USE_OP(concat); diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc index d6651a5b244ba31a01220e6299cb2016ae61fe64..01d7f700da9cc67d0ebbd3d9649e3823f58a8811 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc @@ -57,6 +57,7 @@ TEST(OpConverter, ConvertBlock) { auto* x = scope.Var("conv2d-Y"); auto* x_tensor = x->GetMutable(); x_tensor->Resize(framework::make_ddim(dim_vec)); + x_tensor->mutable_data(platform::CUDAPlace(0)); OpConverter converter; converter.ConvertBlock(*block->Proto(), {"conv2d-Y"}, scope, diff --git a/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc index c5dddbc8cd37b9fb1ba39382af2da5ad045f3af2..aedd6b62df040eeee4e48f628128511cd8bf4439 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc @@ -20,7 +20,7 @@ namespace paddle { namespace inference { namespace tensorrt { -TEST(Pool2dOpConverter, main) { +void test_pool2d(bool global_pooling) { framework::Scope scope; std::unordered_set parameters; TRTConvertValidation validator(5, parameters, scope, 1 << 15); @@ -28,7 +28,10 @@ TEST(Pool2dOpConverter, main) { // The ITensor's Dims should not contain the batch size. // So, the ITensor's Dims of input and output should be C * H * W. validator.DeclInputVar("pool2d-X", nvinfer1::Dims3(3, 4, 4)); - validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 2, 2)); + if (global_pooling) + validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 1, 1)); + else + validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 2, 2)); // Prepare Op description framework::OpDesc desc; @@ -45,6 +48,7 @@ TEST(Pool2dOpConverter, main) { desc.SetAttr("ksize", ksize); desc.SetAttr("strides", strides); desc.SetAttr("paddings", paddings); + desc.SetAttr("global_pooling", global_pooling); LOG(INFO) << "set OP"; validator.SetOp(*desc.Proto()); @@ -53,6 +57,10 @@ TEST(Pool2dOpConverter, main) { validator.Execute(3); } +TEST(Pool2dOpConverter, normal) { test_pool2d(false); } + +TEST(Pool2dOpConverter, test_global_pooling) { test_pool2d(true); } + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/ut_helper.h b/paddle/fluid/inference/tensorrt/convert/ut_helper.h index 4265f33f28fe36b1745baf4761c3c85e3a281d6b..0a6f171fc40a838fd81d6a51aca0430d5526f188 100644 --- a/paddle/fluid/inference/tensorrt/convert/ut_helper.h +++ b/paddle/fluid/inference/tensorrt/convert/ut_helper.h @@ -24,6 +24,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/inference/analysis/helper.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/tensorrt/engine.h" @@ -48,11 +49,17 @@ void RandomizeTensor(framework::LoDTensor* tensor, const platform::Place& place, auto dims = tensor->dims(); size_t num_elements = analysis::AccuDims(dims, dims.size()); PADDLE_ENFORCE_GT(num_elements, 0); - auto* data = tensor->mutable_data(place); + + platform::CPUPlace cpu_place; + framework::LoDTensor temp_tensor; + temp_tensor.Resize(dims); + auto* temp_data = temp_tensor.mutable_data(cpu_place); for (size_t i = 0; i < num_elements; i++) { - *(data + i) = random(0., 1.); + *(temp_data + i) = random(0., 1.); } + + TensorCopySync(temp_tensor, place, tensor); } /* @@ -91,18 +98,26 @@ class TRTConvertValidation { engine_->DeclareInput(name, nvinfer1::DataType::kFLOAT, dims); } + void DeclParamVar(const std::string& name, const std::vector dim_vec) { + DeclVar(name, dim_vec); + } + // Declare a parameter varaible in the scope. void DeclParamVar(const std::string& name, const nvinfer1::Dims& dims) { DeclVar(name, dims, true); } + void DeclOutputVar(const std::string& name, const std::vector dim_vec) { + DeclVar(name, dim_vec); + } + void DeclOutputVar(const std::string& name, const nvinfer1::Dims& dims) { DeclVar(name, dims); } void DeclVar(const std::string& name, const std::vector dim_vec) { - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); auto* x = scope_.Var(name); auto* x_tensor = x->GetMutable(); @@ -141,18 +156,22 @@ class TRTConvertValidation { PADDLE_ENFORCE(var); auto tensor = var->GetMutable(); - engine_->SetInputFromCPU( + engine_->SetInputFromGPU( input, static_cast(tensor->data()), sizeof(float) * analysis::AccuDims(tensor->dims(), tensor->dims().size())); } } - void Execute(int batch_size) { + // We use the set 'neglected_output' here, because some Ops like batch norm, + // the outputs specified in the op des are only used during training, + // so we should neglect those output during inference. + void Execute(int batch_size, + std::unordered_set neglected_output = {}) { // Execute Fluid Op PADDLE_ENFORCE_LE(batch_size, max_batch_size_); - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); op_->Run(scope_, place); // Execute TRT. engine_->Execute(batch_size); @@ -161,6 +180,7 @@ class TRTConvertValidation { ASSERT_FALSE(op_desc_->OutputArgumentNames().empty()); const size_t output_space_size = 3000; for (const auto& output : op_desc_->OutputArgumentNames()) { + if (neglected_output.count(output)) continue; std::vector fluid_out; std::vector trt_out(output_space_size); engine_->GetOutputInCPU(output, &trt_out[0], output_space_size); diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index b821c3d0bf425c46fae634fbf53f7ee63100ca5c..14e9e14d33d637ee68e37593cc48721e5169499f 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -33,6 +33,7 @@ void TensorRTEngine::Build(const DescType &paddle_model) { } void TensorRTEngine::Execute(int batch_size) { + freshDeviceId(); batch_size_ = batch_size; std::vector buffers; for (auto &buf : buffers_) { @@ -60,6 +61,7 @@ TensorRTEngine::~TensorRTEngine() { } void TensorRTEngine::FreezeNetwork() { + freshDeviceId(); PADDLE_ENFORCE(infer_builder_ != nullptr, "Call InitNetwork first to initialize network."); PADDLE_ENFORCE(infer_network_ != nullptr, @@ -241,6 +243,13 @@ void TensorRTEngine::SetRuntimeBatch(size_t batch_size) { int TensorRTEngine::GetRuntimeBatch() { return runtime_batch_; } +void TensorRTEngine::freshDeviceId() { + int count; + cudaGetDeviceCount(&count); + PADDLE_ENFORCE_LT(device_, count); + cudaSetDevice(device_); +} + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index 694468c419c20089de1cdecff1a903ad0cc6e99f..bd3ba4cea6551a7f6651e311e2649de191a6faa1 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -19,6 +19,7 @@ limitations under the License. */ #include #include #include +#include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/inference/engine.h" #include "paddle/fluid/inference/tensorrt/helper.h" #include "paddle/fluid/inference/utils/singleton.h" @@ -52,13 +53,15 @@ class TensorRTEngine : public EngineBase { }; TensorRTEngine(int max_batch, int max_workspace, - cudaStream_t* stream = nullptr, + cudaStream_t* stream = nullptr, int device = 0, nvinfer1::ILogger& logger = NaiveLogger::Global()) : max_batch_(max_batch), max_workspace_(max_workspace), stream_(stream ? stream : &default_stream_), - logger_(logger) { - cudaStreamCreate(&default_stream_); + logger_(logger), + device_(device) { + freshDeviceId(); + cudaStreamCreate(stream_); } virtual ~TensorRTEngine(); @@ -119,6 +122,15 @@ class TensorRTEngine : public EngineBase { nvinfer1::INetworkDefinition* network() { return infer_network_.get(); } void SetRuntimeBatch(size_t batch_size); int GetRuntimeBatch(); + int GetDevice() { return device_; } + + // A pointer to CPU memory is needed of the TRT weight. + // Before TRT runs, fluid loads weight into GPU storage. + // so we need to copy the weights from GPU to CPU in our op converter. + // We use a map to store these weights for the weight memory is not released + // in advance, which affecting the construction of TRT Op. + std::unordered_map> + weight_map; private: // the max batch size @@ -140,6 +152,8 @@ class TensorRTEngine : public EngineBase { std::unordered_map buffer_sizes_; std::unordered_map itensor_map_; + // The specific GPU id that the TensorRTEngine bounded to. + int device_; // TensorRT related internal members template @@ -156,6 +170,10 @@ class TensorRTEngine : public EngineBase { infer_ptr infer_network_; infer_ptr infer_engine_; infer_ptr infer_context_; + // Each ICudaEngine object is bound to a specific GPU when it is instantiated, + // ensure that the thread is associated with the correct device by calling + // freshDeviceId(). + void freshDeviceId(); }; // class TensorRTEngine // Add an layer__ into engine__ with args ARGS. @@ -188,8 +206,8 @@ class TRT_EngineManager { // Create or get an engine called `name` TensorRTEngine* Create(int max_batch, int max_workspace, cudaStream_t* stream, - const std::string& name) { - auto* p = new TensorRTEngine(max_batch, max_workspace, stream); + const std::string& name, int gpu_device = 0) { + auto* p = new TensorRTEngine(max_batch, max_workspace, stream, gpu_device); engines_[name].reset(p); return p; } diff --git a/paddle/fluid/inference/tensorrt/test_engine.cc b/paddle/fluid/inference/tensorrt/test_engine.cc index dc03702990587bf5e65d28da662d10df4d882110..da1f6535cb3b2476cd475797861d6d2bb6d88856 100644 --- a/paddle/fluid/inference/tensorrt/test_engine.cc +++ b/paddle/fluid/inference/tensorrt/test_engine.cc @@ -27,7 +27,7 @@ namespace tensorrt { class TensorRTEngineTest : public ::testing::Test { protected: void SetUp() override { - ASSERT_EQ(0, cudaStreamCreate(&stream_)); + // ASSERT_EQ(0, cudaStreamCreate(&stream_)); engine_ = new TensorRTEngine(10, 1 << 10, &stream_); engine_->InitNetwork(); } diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index e8b5dec9d49f5613cec92441d19ab7dc1a1ad90c..f456720f981e8e17bf04c15e2f3fbc121a5b66b3 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -100,7 +100,8 @@ function(op_library TARGET) endif() # Define operators that don't need pybind here. - foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op" "tensor_array_read_write_op") + foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op" +"tensor_array_read_write_op" "tensorrt_engine_op") if ("${TARGET}" STREQUAL "${manual_pybind_op}") set(pybind_flag 1) endif() @@ -248,6 +249,7 @@ op_library(softmax_op DEPS softmax) op_library(sequence_softmax_op DEPS softmax) if (WITH_GPU AND TENSORRT_FOUND) op_library(tensorrt_engine_op DEPS tensorrt_engine tensorrt_converter) + file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(tensorrt_engine);\n") nv_test(test_tensorrt_engine_op SRCS tensorrt_engine_op_test.cc DEPS tensorrt_engine_op analysis) @@ -280,6 +282,7 @@ op_library(unsqueeze_op DEPS reshape_op) op_library(squeeze_op DEPS reshape_op) op_library(extract_rows_op DEPS memory) op_library(flatten_op DEPS reshape_op) +op_library(sequence_pad_op DEPS sequence_padding) if (WITH_GPU) op_library(conv_op DEPS vol2col depthwise_conv im2col) diff --git a/paddle/fluid/operators/fetch_barrier_op.cc b/paddle/fluid/operators/fetch_barrier_op.cc index d9cd956dfdff3d009d38ee5088f5396080580483..9d7ac7ab6194593747548fac3cefc8d4ed3058d8 100644 --- a/paddle/fluid/operators/fetch_barrier_op.cc +++ b/paddle/fluid/operators/fetch_barrier_op.cc @@ -52,6 +52,8 @@ class FetchBarrierOp : public framework::OperatorBase { class FetchBarrierOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() { + AddOutput("Out", "(Any) Dummy outputs, used for control dependency") + .AsDuplicable(); AddComment(R"DOC( SendBarrier operator diff --git a/paddle/fluid/operators/math/sequence_padding.cc b/paddle/fluid/operators/math/sequence_padding.cc index d63c6c4ed55331235188c1c750468d4e75b9b7f2..25f06a25a0638cbb394df58d35f88307941d117f 100644 --- a/paddle/fluid/operators/math/sequence_padding.cc +++ b/paddle/fluid/operators/math/sequence_padding.cc @@ -18,65 +18,86 @@ namespace paddle { namespace operators { namespace math { +template +void CopyValidData(framework::Tensor* dst_tensor, + const framework::Tensor* src_tensor, + const framework::Vector& seq_offsets, + int pad_seq_len, int step_width, bool norm_by_len, + CopyType type, PadLayout layout) { + int seq_num = seq_offsets.size() - 1; + const T* src_data = src_tensor->data(); + T* dst_data = dst_tensor->data(); + + int seq_cpy_gap = step_width; + int pad_cpy_gap = + layout == kBatchLengthWidth ? step_width : seq_num * step_width; + for (int seq_idx = 0; seq_idx < seq_num; ++seq_idx) { + int valid_seq_len = seq_offsets[seq_idx + 1] - seq_offsets[seq_idx]; + PADDLE_ENFORCE_GE( + pad_seq_len, valid_seq_len, + "The padded sequence length can not be less than its original length."); + int seq_data_offset = seq_offsets[seq_idx] * step_width; + int pad_data_offset = layout == kBatchLengthWidth + ? seq_idx * pad_seq_len * step_width + : seq_idx * step_width; + float scale = 1.0f / static_cast(valid_seq_len); + + for (int step_idx = 0; step_idx < valid_seq_len; ++step_idx) { + const T* src = + src_data + (type == kSeqToPad ? seq_data_offset : pad_data_offset); + T* dst = + dst_data + (type == kSeqToPad ? pad_data_offset : seq_data_offset); + memcpy(dst, src, step_width * sizeof(T)); + if (norm_by_len) { + for (int i = 0; i < step_width; ++i) { + *(dst + i) *= scale; + } + } + seq_data_offset += seq_cpy_gap; + pad_data_offset += pad_cpy_gap; + } + } +} + template class PaddingLoDTensorFunctor { public: void operator()(const platform::CPUDeviceContext& context, - const framework::LoDTensor& seq, framework::Tensor* padding, - bool norm_by_times) { - auto lod = seq.lod(); - PADDLE_ENFORCE_GT(lod.size(), 0UL, - "The LoD of LoDTensor seq should not be null."); - - const size_t level = 0; - framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - - auto seq_dims = seq.dims(); - PADDLE_ENFORCE_EQ(seq_dims[0], - static_cast(abs_offset_lod[level].back()), - "The first dimension of LoDTensor seq should be " - "equal to the sum of all sequences's length."); - - auto padding_dims = padding->dims(); - PADDLE_ENFORCE_EQ(padding_dims.size(), 3UL, - "The input padding should be a 3-D Tensor of shape " - "[max_sequence_length, num_sequences, sequence_width]."); - - const int64_t max_sequence_length = MaximumSequenceLength(lod, level); - PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length, - "The first dimension of Tensor padding should be the " - "maximum length of all sequences in LoDTensor seq."); - - const int64_t num_sequences = abs_offset_lod[level].size() - 1; - PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences, - "The second dimension of Tensor padding should be the " - "number of sequences in LoDTensor seq."); - - const int64_t sequence_width = seq.numel() / seq_dims[0]; - PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width, - "The third dimension of Tensor padding should be the " - "width of sequence in LoDTensor seq."); - - const T* seq_data = seq.data(); - T* padding_data = padding->data(); - for (int64_t i = 0; i < max_sequence_length; ++i) { - for (int64_t j = 0; j < num_sequences; ++j) { - int64_t start_pos = abs_offset_lod[level][j]; - int64_t sequence_length = abs_offset_lod[level][j + 1] - start_pos; - if (i < sequence_length) { - // i > 0 => sequence_length > 0 - T scale = - norm_by_times ? (1.0f / static_cast(sequence_length)) : 1.0f; - for (int64_t k = 0; k < sequence_width; ++k) { - padding_data[(i * num_sequences + j) * sequence_width + k] = - seq_data[(start_pos + i) * sequence_width + k] * scale; - } - } else { - memset(padding_data + (i * num_sequences + j) * sequence_width, 0, - sequence_width * sizeof(T)); - } + const framework::LoDTensor& seq_tensor, + framework::LoDTensor* pad_tensor, + const framework::LoDTensor& pad_value, int pad_seq_len = -1, + int lod_level = 0, bool norm_by_times = false, + const PadLayout layout = kBatchLengthWidth) { + auto seq_lod = seq_tensor.lod(); + const auto seq_offsets = framework::ToAbsOffset(seq_lod)[lod_level]; + const auto& seq_tensor_dims = seq_tensor.dims(); + const auto& pad_tensor_dims = pad_tensor->dims(); + if (pad_seq_len == -1) { + pad_seq_len = MaximumSequenceLength(seq_offsets); + } + int step_width = seq_tensor.numel() / seq_tensor_dims[0]; + + CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len, + step_width, layout); + PADDLE_ENFORCE(pad_value.numel() == 1 || pad_value.numel() == step_width, + "The numel of 'pad_value' can only be 1 or be equal to the " + "'step_width'."); + + // fill padding value + T* pad_data = pad_tensor->data(); + const T* pad_value_data = pad_value.data(); + if (pad_value.numel() == 1) { + for (int i = 0; i < pad_tensor->numel(); ++i) { + pad_data[i] = *pad_value_data; + } + } else { + for (int i = 0; i < pad_tensor->numel(); i += step_width) { + memcpy(pad_data + i, pad_value_data, step_width * sizeof(T)); } } + + CopyValidData(pad_tensor, &seq_tensor, seq_offsets, pad_seq_len, + step_width, norm_by_times, kSeqToPad, layout); } }; @@ -84,62 +105,35 @@ template class UnpaddingLoDTensorFunctor { public: void operator()(const platform::CPUDeviceContext& context, - framework::LoDTensor* seq, const framework::Tensor& padding, - bool norm_by_times) { - auto lod = seq->lod(); - PADDLE_ENFORCE_GT(lod.size(), 0UL, - "The LoD of LoDTensor seq should not be null."); - - const size_t level = 0; - framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - - auto seq_dims = seq->dims(); - PADDLE_ENFORCE_EQ(seq_dims[0], - static_cast(abs_offset_lod[level].back()), - "The first dimension of LoDTensor seq should be " - "equal to the sum of all sequences's length."); - - auto padding_dims = padding.dims(); - PADDLE_ENFORCE_EQ(padding_dims.size(), 3UL, - "The input padding should be a 3-D Tensor of shape " - "[max_sequnece_length, num_sequences, sequence_width]."); - - const int64_t max_sequence_length = MaximumSequenceLength(lod, level); - PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length, - "The first dimension of Tensor padding should be " - "the maximum length of all sequences in LoDTensor seq."); - - const int64_t num_sequences = abs_offset_lod[level].size() - 1; - PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences, - "The second dimension of Tensor padding should be " - "the number of sequences in LoDTensor seq."); - - const int64_t sequence_width = seq->numel() / seq_dims[0]; - PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width, - "The third dimension of Tensor padding should be the " - "width of sequence in LoDTensor seq."); - - const T* padding_data = padding.data(); - T* seq_data = seq->data(); - for (int64_t i = 0; i < num_sequences; ++i) { - int64_t start_pos = abs_offset_lod[level][i]; - int64_t sequence_length = abs_offset_lod[level][i + 1] - start_pos; - for (int64_t j = 0; j < sequence_length; ++j) { - // sequence_width > j > 0 - T scale = - norm_by_times ? (1.0f / static_cast(sequence_length)) : 1.0f; - for (int64_t k = 0; k < sequence_width; ++k) { - seq_data[(start_pos + j) * sequence_width + k] = - padding_data[(j * num_sequences + i) * sequence_width + k] * - scale; - } - } + const framework::LoDTensor& pad_tensor, + framework::LoDTensor* seq_tensor, int pad_seq_len = -1, + int lod_level = 0, bool norm_by_times = false, + const PadLayout layout = kBatchLengthWidth) { + auto seq_offsets = framework::ToAbsOffset(seq_tensor->lod())[lod_level]; + const auto& seq_tensor_dims = seq_tensor->dims(); + const auto& pad_tensor_dims = pad_tensor.dims(); + if (pad_seq_len == -1) { + pad_seq_len = MaximumSequenceLength(seq_offsets); } + int step_width = seq_tensor->numel() / seq_tensor_dims[0]; + + CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len, + step_width, layout); + + CopyValidData(seq_tensor, &pad_tensor, seq_offsets, pad_seq_len, + step_width, norm_by_times, kPadToSeq, layout); } }; +template class PaddingLoDTensorFunctor; +template class PaddingLoDTensorFunctor; template class PaddingLoDTensorFunctor; +template class PaddingLoDTensorFunctor; + +template class UnpaddingLoDTensorFunctor; +template class UnpaddingLoDTensorFunctor; template class UnpaddingLoDTensorFunctor; +template class UnpaddingLoDTensorFunctor; } // namespace math } // namespace operators diff --git a/paddle/fluid/operators/math/sequence_padding.cu b/paddle/fluid/operators/math/sequence_padding.cu index 0956a0c17d387f4a174c7ed4e9b1b1f816dcf4ae..035e10dcbe4e2083723e47d7dda75ce267a9f141 100644 --- a/paddle/fluid/operators/math/sequence_padding.cu +++ b/paddle/fluid/operators/math/sequence_padding.cu @@ -19,41 +19,32 @@ namespace paddle { namespace operators { namespace math { -template -__global__ void SequencePaddingKernel(T* padding, T* sequence, - const size_t* sequence_start_positions, - const size_t sequence_width, - const size_t max_sequence_length, - const size_t num_sequences) { - size_t padding_idx = blockIdx.y; - size_t start_pos = sequence_start_positions[padding_idx]; - size_t sequence_length = - sequence_start_positions[padding_idx + 1] - start_pos; - - size_t sequence_idx = blockIdx.x * blockDim.y + threadIdx.y; - size_t padding_base_idx = - (sequence_idx * num_sequences + padding_idx) * sequence_width; - size_t sequence_base_idx = (start_pos + sequence_idx) * sequence_width; - - if (sequence_idx < sequence_length) { - T scale = NormByTimes ? (1.0f / static_cast(sequence_length)) : 1.0f; - if (Padding) { - /* sequence -> padding */ - for (size_t i = threadIdx.x; i < sequence_width; i += blockDim.x) { - padding[padding_base_idx + i] = scale * sequence[sequence_base_idx + i]; - } - } else { - /* padding -> sequence */ - for (size_t i = threadIdx.x; i < sequence_width; i += blockDim.x) { - sequence[sequence_base_idx + i] = scale * padding[padding_base_idx + i]; - } +template +__global__ void SequencePaddingKernel( + T* dst, const T* src, const T* pad_value, bool is_constant_pad, + const size_t* seq_offsets, const size_t seq_num, const size_t pad_seq_len, + const size_t step_width, bool norm_by_len, const PadLayout layout) { + size_t seq_idx = blockIdx.y; + size_t seq_len = seq_offsets[seq_idx + 1] - seq_offsets[seq_idx]; + + size_t step_idx = blockIdx.x * blockDim.y + threadIdx.y; + size_t seq_data_offset = (seq_offsets[seq_idx] + step_idx) * step_width; + size_t pad_data_offset = layout == kBatchLengthWidth + ? (seq_idx * pad_seq_len + step_idx) * step_width + : (step_idx * seq_num + seq_idx) * step_width; + + T* dst_data = dst + (Type == kSeqToPad ? pad_data_offset : seq_data_offset); + const T* src_data = + src + (Type == kSeqToPad ? seq_data_offset : pad_data_offset); + + if (step_idx < seq_len) { + float scale = norm_by_len ? (1.0f / static_cast(seq_len)) : 1.0f; + for (size_t i = threadIdx.x; i < step_width; i += blockDim.x) { + dst_data[i] = scale * src_data[i]; } - } else if (sequence_idx < max_sequence_length) { - if (Padding) { - /* sequence -> padding */ - for (size_t i = threadIdx.x; i < sequence_width; i += blockDim.x) { - padding[padding_base_idx + i] = 0; - } + } else if (step_idx < pad_seq_len && Type == kSeqToPad) { + for (size_t i = threadIdx.x; i < step_width; i += blockDim.x) { + dst_data[i] = is_constant_pad ? pad_value[0] : pad_value[i]; } } } @@ -62,74 +53,59 @@ template class PaddingLoDTensorFunctor { public: void operator()(const platform::CUDADeviceContext& context, - const framework::LoDTensor& seq, framework::Tensor* padding, - bool norm_by_times) { - auto lod = seq.lod(); - PADDLE_ENFORCE_GT(lod.size(), 0UL, - "The lod of LoDTensor seq should not be null."); - - const size_t level = 0; - framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - - auto seq_dims = seq.dims(); - PADDLE_ENFORCE_EQ(seq_dims[0], - static_cast(abs_offset_lod[level].back()), - "The first dimension of LoDTensor seq should be " - "equal to the sum of all sequences's length."); - - auto padding_dims = padding->dims(); - PADDLE_ENFORCE_EQ(padding_dims.size(), 3UL, - "The input padding should be a 3-D Tensor of shape " - "[max_sequence_length, num_sequences, sequence_width]."); - - int64_t max_sequence_length = MaximumSequenceLength(lod, level); - PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length, - "The first dimension of Tensor padding should be the " - "maximum length of all sequences in LoDTensor seq."); - - const int64_t num_sequences = abs_offset_lod[level].size() - 1; - PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences, - "The second dimension of Tensor padding should be the " - "number of sequences in LoDTensor seq."); - - const int64_t sequence_width = seq.numel() / seq_dims[0]; - PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width, - "The third dimension of Tensor padding should be the " - "width of sequence in LoDTensor seq."); - - if (!norm_by_times && num_sequences == 1UL) { - TensorCopy(seq, context.GetPlace(), context, padding); - padding->Resize(padding_dims); + const framework::LoDTensor& seq_tensor, + framework::LoDTensor* pad_tensor, + const framework::LoDTensor& pad_value, int pad_seq_len = -1, + int lod_level = 0, bool norm_by_times = false, + const PadLayout layout = kBatchLengthWidth) { + auto seq_lod = seq_tensor.lod(); + const auto seq_offsets = framework::ToAbsOffset(seq_lod)[lod_level]; + const auto& seq_tensor_dims = seq_tensor.dims(); + const auto& pad_tensor_dims = pad_tensor->dims(); + int max_seq_len = MaximumSequenceLength(seq_offsets); + if (pad_seq_len == -1) { + pad_seq_len = max_seq_len; + } + PADDLE_ENFORCE_GE(pad_seq_len, max_seq_len, + "The pad_seq_len must be equal to or greater than the " + "original max sequence length."); + int step_width = seq_tensor.numel() / seq_tensor_dims[0]; + int seq_num = seq_offsets.size() - 1; + + CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len, + step_width, layout); + PADDLE_ENFORCE(pad_value.numel() == 1 || pad_value.numel() == step_width, + "The numel of 'pad_value' can only be 1 or be equal to the " + "'step_width'."); + + if (!norm_by_times && seq_num == 1UL && pad_seq_len == max_seq_len) { + TensorCopy(seq_tensor, context.GetPlace(), context, pad_tensor); + pad_tensor->Resize(pad_tensor_dims); return; } - const int64_t kBlockSize = 512; + const int kBlockSize = 512; /* At least use 32 threads to copy sequence_width elements, * and at least 8 elements for each thread. */ size_t block_dim_x = - std::min(((((sequence_width + 7) >> 3) + 31) >> 5) << 5, kBlockSize); + std::min(((((step_width + 7) >> 3) + 31) >> 5) << 5, kBlockSize); size_t block_dim_y = kBlockSize / block_dim_x; dim3 threads(block_dim_x, block_dim_y); - size_t grid_dim_x = (max_sequence_length + block_dim_y - 1) / block_dim_y; - size_t grid_dim_y = num_sequences; + size_t grid_dim_x = (pad_seq_len + block_dim_y - 1) / block_dim_y; + size_t grid_dim_y = seq_num; dim3 grid(grid_dim_x, grid_dim_y); - const T* seq_data = seq.data(); - T* padding_data = padding->data(); - if (norm_by_times) { - SequencePaddingKernel<<>>( - padding_data, const_cast(seq_data), - abs_offset_lod[level].CUDAData(context.GetPlace()), sequence_width, - max_sequence_length, num_sequences); - } else { - SequencePaddingKernel<<>>( - padding_data, const_cast(seq_data), - abs_offset_lod[level].CUDAData(context.GetPlace()), sequence_width, - max_sequence_length, num_sequences); - } + const T* seq_data = seq_tensor.data(); + T* pad_data = pad_tensor->data(); + const T* pad_value_data = pad_value.data(); + + SequencePaddingKernel<<>>( + pad_data, seq_data, pad_value_data, pad_value.numel() == 1, + seq_offsets.CUDAData(context.GetPlace()), seq_num, pad_seq_len, + step_width, norm_by_times, layout); } }; @@ -137,79 +113,62 @@ template class UnpaddingLoDTensorFunctor { public: void operator()(const platform::CUDADeviceContext& context, - framework::LoDTensor* seq, const framework::Tensor& padding, - bool norm_by_times) { - auto lod = seq->lod(); - PADDLE_ENFORCE_GT(lod.size(), 0UL, - "The lod of LoDTensor seq should not be null."); - - const size_t level = 0; - framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - - auto seq_dims = seq->dims(); - PADDLE_ENFORCE_EQ(seq_dims[0], - static_cast(abs_offset_lod[level].back()), - "The first dimension of LoDTensor seq should be " - "equal to the sum of all sequences's length."); - - auto padding_dims = padding.dims(); - PADDLE_ENFORCE_EQ(padding_dims.size(), 3UL, - "The input padding should be a 3-D Tensor of shape " - "[max_sequnece_length, num_sequences, sequence_width]."); - - int64_t max_sequence_length = MaximumSequenceLength(lod, level); - PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length, - "The first dimension of Tensor padding should be " - "the maximum length of all sequences in LoDTensor seq."); - - const int64_t num_sequences = abs_offset_lod[level].size() - 1; - PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences, - "The second dimension of Tensor padding should be " - "the number of sequences in LoDTensor seq."); - - const int64_t sequence_width = seq->numel() / seq_dims[0]; - PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width, - "The third dimension of Tensor padding should be the " - "width of sequence in LoDTensor seq."); - - if (!norm_by_times && num_sequences == 1UL) { - TensorCopy(padding, context.GetPlace(), context, seq); - seq->Resize(seq_dims); + const framework::LoDTensor& pad_tensor, + framework::LoDTensor* seq_tensor, int pad_seq_len = -1, + int lod_level = 0, bool norm_by_times = false, + const PadLayout layout = kBatchLengthWidth) { + auto seq_offsets = framework::ToAbsOffset(seq_tensor->lod())[lod_level]; + const auto& seq_tensor_dims = seq_tensor->dims(); + const auto& pad_tensor_dims = pad_tensor.dims(); + int max_seq_len = MaximumSequenceLength(seq_offsets); + if (pad_seq_len == -1) { + pad_seq_len = max_seq_len; + } + int step_width = seq_tensor->numel() / seq_tensor_dims[0]; + int seq_num = seq_offsets.size() - 1; + + CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len, + step_width, layout); + + if (!norm_by_times && seq_num == 1UL && pad_seq_len == max_seq_len) { + TensorCopy(pad_tensor, context.GetPlace(), context, seq_tensor); + seq_tensor->Resize(seq_tensor_dims); return; } - const int64_t kBlockSize = 512; + const int kBlockSize = 512; /* At least use 32 threads to copy sequence_width elements, * and at least 8 elements for each thread. */ size_t block_dim_x = - std::min(((((sequence_width + 7) >> 3) + 31) >> 5) << 5, kBlockSize); + std::min(((((step_width + 7) >> 3) + 31) >> 5) << 5, kBlockSize); size_t block_dim_y = kBlockSize / block_dim_x; dim3 threads(block_dim_x, block_dim_y); - size_t grid_dim_x = (max_sequence_length + block_dim_y - 1) / block_dim_y; - size_t grid_dim_y = num_sequences; + size_t grid_dim_x = (pad_seq_len + block_dim_y - 1) / block_dim_y; + size_t grid_dim_y = seq_num; dim3 grid(grid_dim_x, grid_dim_y); - const T* padding_data = padding.data(); - T* seq_data = seq->data(); - if (norm_by_times) { - SequencePaddingKernel<<>>( - const_cast(padding_data), seq_data, - abs_offset_lod[level].CUDAData(context.GetPlace()), sequence_width, - max_sequence_length, num_sequences); - } else { - SequencePaddingKernel<<>>( - const_cast(padding_data), seq_data, - abs_offset_lod[level].CUDAData(context.GetPlace()), sequence_width, - max_sequence_length, num_sequences); - } + const T* pad_data = pad_tensor.data(); + T* seq_data = seq_tensor->data(); + + SequencePaddingKernel<<>>( + seq_data, pad_data, nullptr, false, + seq_offsets.CUDAData(context.GetPlace()), seq_num, pad_seq_len, + step_width, norm_by_times, layout); } }; +template class PaddingLoDTensorFunctor; +template class PaddingLoDTensorFunctor; template class PaddingLoDTensorFunctor; +template class PaddingLoDTensorFunctor; + +template class UnpaddingLoDTensorFunctor; +template class UnpaddingLoDTensorFunctor; template class UnpaddingLoDTensorFunctor; +template class UnpaddingLoDTensorFunctor; } // namespace math } // namespace operators diff --git a/paddle/fluid/operators/math/sequence_padding.h b/paddle/fluid/operators/math/sequence_padding.h index b56e6db1ebdac1a00561c07845c03bb8fbd8d35a..e752aa58979dddba4d010071d2c4b5dc3e0c6756 100644 --- a/paddle/fluid/operators/math/sequence_padding.h +++ b/paddle/fluid/operators/math/sequence_padding.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include +#include #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/platform/device_context.h" @@ -22,17 +23,33 @@ namespace paddle { namespace operators { namespace math { -inline static size_t MaximumSequenceLength(const framework::LoD& lod, - const size_t level) { - const size_t num_sequences = lod[level].size() - 1; - size_t max_sequence_length = 0; - framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - for (size_t i = 0; i < num_sequences; ++i) { - max_sequence_length = - std::max(max_sequence_length, - abs_offset_lod[level][i + 1] - abs_offset_lod[level][i]); +enum PadLayout { kBatchLengthWidth = 0, kLengthBatchWidth }; + +enum CopyType { kSeqToPad, kPadToSeq }; + +inline static size_t MaximumSequenceLength( + const framework::Vector& seq_offset) { + size_t seq_num = seq_offset.size() - 1; + size_t max_seq_len = 0; + for (size_t i = 0; i < seq_num; ++i) { + max_seq_len = std::max(max_seq_len, seq_offset[i + 1] - seq_offset[i]); } - return max_sequence_length; + return max_seq_len; +} + +inline static void CheckDims(const framework::DDim& seq_tensor_dims, + const framework::DDim& pad_tensor_dims, + const framework::Vector& seq_offset, + int64_t padded_seq_len, int64_t step_width, + const PadLayout& layout) { + PADDLE_ENFORCE_EQ(static_cast(seq_tensor_dims[0]), seq_offset.back(), + "Value of 1st dimension of the sequence tensor should be " + "equal to sum of lengths of all sequences."); + + PADDLE_ENFORCE(seq_tensor_dims.size() + 1 == pad_tensor_dims.size() || + seq_tensor_dims.size() == pad_tensor_dims.size(), + "pad_tensor's rank should be 1 greater than seq_tensor's " + "rank, or be equal with it."); } /* @@ -64,15 +81,22 @@ inline static size_t MaximumSequenceLength(const framework::LoD& lod, template class PaddingLoDTensorFunctor { public: - void operator()(const DeviceContext& context, const framework::LoDTensor& seq, - framework::Tensor* padding, bool norm_by_times); + void operator()(const DeviceContext& context, + const framework::LoDTensor& seq_tensor, + framework::LoDTensor* pad_tensor, + const framework::LoDTensor& pad_value, int pad_seq_len = -1, + int lod_level = 0, bool norm_by_times = false, + const PadLayout layout = kBatchLengthWidth); }; template class UnpaddingLoDTensorFunctor { public: - void operator()(const DeviceContext& context, framework::LoDTensor* seq, - const framework::Tensor& padding, bool norm_by_times); + void operator()(const DeviceContext& context, + const framework::LoDTensor& pad_tensor, + framework::LoDTensor* seq_tensor, int pad_seq_len = -1, + int lod_level = 0, bool norm_by_times = false, + const PadLayout layout = kBatchLengthWidth); }; } // namespace math diff --git a/paddle/fluid/operators/math/sequence_padding_test.cc b/paddle/fluid/operators/math/sequence_padding_test.cc index b0c201db0ccbe81d8f57cd984d2cdfd2f6a48f25..4f61b1029c65aedaf4fce771866964fe1d0d6112 100644 --- a/paddle/fluid/operators/math/sequence_padding_test.cc +++ b/paddle/fluid/operators/math/sequence_padding_test.cc @@ -23,7 +23,9 @@ void TestSequencePadding(const paddle::framework::LoD& lod, paddle::framework::LoDTensor cpu_seq_back; paddle::framework::LoDTensor seq; paddle::framework::LoDTensor seq_back; - paddle::framework::Tensor padding; + paddle::framework::LoDTensor padding; + paddle::framework::LoDTensor cpu_pad_value; + paddle::framework::LoDTensor pad_value; const size_t level = lod.size() - 1; auto seq_dims = @@ -46,20 +48,33 @@ void TestSequencePadding(const paddle::framework::LoD& lod, } const size_t max_sequence_length = - paddle::operators::math::MaximumSequenceLength(lod, level); + paddle::operators::math::MaximumSequenceLength(lod[level]); const size_t num_sequences = lod[level].size() - 1; auto padding_dims = paddle::framework::make_ddim({static_cast(max_sequence_length), static_cast(num_sequences), static_cast(sequence_width)}); + padding.mutable_data(padding_dims, *place); + + T* pad_value_data = + cpu_pad_value.mutable_data({1}, paddle::platform::CPUPlace()); + *pad_value_data = static_cast(0); + if (paddle::platform::is_cpu_place(*place)) { + pad_value = cpu_pad_value; + } else { + TensorCopySync(cpu_pad_value, *place, &pad_value); + } + paddle::operators::math::PaddingLoDTensorFunctor()( - *context, seq, &padding, false); + *context, seq, &padding, pad_value, -1, 0, false, + paddle::operators::math::kLengthBatchWidth); seq_back.set_lod(lod); seq_back.mutable_data(seq_dims, *place); paddle::operators::math::UnpaddingLoDTensorFunctor()( - *context, &seq_back, padding, false); + *context, padding, &seq_back, -1, 0, false, + paddle::operators::math::kLengthBatchWidth); if (paddle::platform::is_cpu_place(*place)) { cpu_seq_back = seq_back; diff --git a/paddle/fluid/operators/scale_op.cc b/paddle/fluid/operators/scale_op.cc index 7f8822e40053b5bcd394f446138a2292d80b69bf..c614de2eac143b3a545c60226aefa93dd72dea4f 100644 --- a/paddle/fluid/operators/scale_op.cc +++ b/paddle/fluid/operators/scale_op.cc @@ -13,8 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/scale_op.h" + #include +#include "paddle/fluid/operators/detail/safe_ref.h" + namespace paddle { namespace operators { @@ -52,6 +55,21 @@ $$Out = scale*X$$ } }; +class ScaleOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + auto &in_var_name = op_desc.Input("X").front(); + auto &in_var = detail::Ref(block->FindVarRecursive(in_var_name)); + + auto out_var_name = op_desc.Output("Out").front(); + auto *out_var = block->FindVarRecursive(out_var_name); + + out_var->SetType(in_var.GetType()); + out_var->SetDataType(in_var.GetDataType()); + } +}; + class ScaleGradMaker : public framework::SingleGradOpDescMaker { public: using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; @@ -71,7 +89,8 @@ class ScaleGradMaker : public framework::SingleGradOpDescMaker { namespace ops = paddle::operators; -REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, ops::ScaleGradMaker); +REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, ops::ScaleGradMaker, + ops::ScaleOpVarTypeInference); REGISTER_OP_CPU_KERNEL( scale, ops::ScaleKernel, ops::ScaleKernel, diff --git a/paddle/fluid/operators/scale_op.h b/paddle/fluid/operators/scale_op.h index c6a59b76adcd6b4d3e7db5e7c7185f266f46841f..fe035aba81dd74d21539974beed255275be3013b 100644 --- a/paddle/fluid/operators/scale_op.h +++ b/paddle/fluid/operators/scale_op.h @@ -22,17 +22,29 @@ namespace operators { template class ScaleKernel : public framework::OpKernel { public: - virtual void Compute(const framework::ExecutionContext& context) const { - auto* tensor = context.Output("Out"); - auto* in = context.Input("X"); - tensor->mutable_data(in->place()); + virtual void Compute(const framework::ExecutionContext& ctx) const { + auto* in_var = ctx.InputVar("X"); + auto* in = ctx.Input("X"); - auto scale = static_cast(context.Attr("scale")); + auto* out_var = ctx.OutputVar("Out"); + auto* out = ctx.Output("Out"); + out->mutable_data(in->place()); - auto eigen_out = framework::EigenVector::Flatten(*tensor); + PADDLE_ENFORCE_EQ(in->dims(), out->dims(), + "in and out should have the same dim"); + + auto scale = static_cast(ctx.Attr("scale")); + + if (in_var->IsType() && in_var != out_var) { + auto& in_slr = in_var->Get(); + auto* out_slr = out_var->GetMutable(); + out_slr->set_rows(in_slr.rows()); + out_slr->set_height(in_slr.height()); + } + + auto eigen_out = framework::EigenVector::Flatten(*out); auto eigen_in = framework::EigenVector::Flatten(*in); - auto& dev = - *context.template device_context().eigen_device(); + auto& dev = *ctx.template device_context().eigen_device(); eigen_out.device(dev) = scale * eigen_in; } }; diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc index 14b07649c416ff1b671fc9b5ee4eb956b44570c5..40404295266899c6ac2f7b1e08fdf7db40958794 100644 --- a/paddle/fluid/operators/send_barrier_op.cc +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -56,6 +56,10 @@ class SendBarrierOp : public framework::OperatorBase { class SendBarrierOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() { + AddInput("X", "(Any) Dummy inputs, used for control dependency") + .AsDuplicable(); + AddOutput("Out", "(Any) Dummy outputs, used for control dependency") + .AsDuplicable(); AddComment(R"DOC( SendBarrier operator diff --git a/paddle/fluid/operators/sequence_pad_op.cc b/paddle/fluid/operators/sequence_pad_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..44d73aa4076abfe15c906478702ac7c4a55303d4 --- /dev/null +++ b/paddle/fluid/operators/sequence_pad_op.cc @@ -0,0 +1,194 @@ +/* Copyright (c) 2018 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. */ + +#include "paddle/fluid/operators/sequence_pad_op.h" + +namespace paddle { +namespace operators { + +class SequencePadOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequencePadOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("PadValue"), + "Input(PadValue) of SequencePadOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SequencePadOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE_GE(x_dims.size(), 2, + "The rank of Input(x) can't be less than 2."); + auto time_step_dims = framework::slice_ddim(x_dims, 1, x_dims.size()); + auto pad_value_dims = ctx->GetInputDim("PadValue"); + PADDLE_ENFORCE(pad_value_dims == framework::make_ddim({1}) || + pad_value_dims == time_step_dims, + "The Input(PadValue) must be a scalar or a tensor whose " + "shape equals to time steps in sequences"); + + int out_dim_0 = -1; + int out_dim_1 = -1; + + if (ctx->IsRuntime()) { + // run time + framework::Variable* x_var = + boost::get(ctx->GetInputVarPtrs("X")[0]); + const auto& x_lod = x_var->Get().lod(); + PADDLE_ENFORCE(!x_lod.empty(), "The Input(X) must hold lod info."); + const auto& x_lod_0 = x_lod[0]; + PADDLE_ENFORCE_GE(x_lod_0.size(), 2, + "The Input(X)'s lod info is corrupted."); + PADDLE_ENFORCE_EQ( + x_dims[0], static_cast(x_lod_0.back()), + "The Input(X)'s lod info mismatches the actual tensor shape."); + + int seq_num = x_lod_0.size() - 1; + int max_seq_len = math::MaximumSequenceLength(x_lod_0); + int padded_length = ctx->Attrs().Get("padded_length"); + if (padded_length == -1) { + padded_length = max_seq_len; + } + PADDLE_ENFORCE_GE(padded_length, max_seq_len, + "The Attr(padded_length) must be -1 or an int greater " + "than the length of the longest original sequence."); + out_dim_0 = seq_num; + out_dim_1 = padded_length; + } else { + // compile time + framework::VarDesc* x_desc = + boost::get(ctx->GetInputVarPtrs("X")[0]); + PADDLE_ENFORCE_GE(x_desc->GetLoDLevel(), 1); + } + + std::vector out_dims_vec{out_dim_0, out_dim_1}; + auto time_step_dims_vec = framework::vectorize2int(time_step_dims); + out_dims_vec.insert(out_dims_vec.end(), time_step_dims_vec.begin(), + time_step_dims_vec.end()); + ctx->SetOutputDim("Out", framework::make_ddim(out_dims_vec)); + } +}; + +class SequencePadOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(LoDTensor, default LoDTensor) Input variable which " + "should contain lod information."); + AddInput("PadValue", + "(LoDTensor), this Tensor holds values that will be fill into " + "padded steps. It can be a scalar or a tensor whose shape equals " + "to time steps in sequences. If it's a scalar, it will be " + "automatically broadcasted to the shape of time step."); + AddOutput( + "Out", + "(LoDTensor) The output vairable, which contains padded sequences."); + AddAttr( + "padded_length", + "The length of padded sequences. It can be setted to -1 or " + "any positive int. When it is -1, all sequences will be padded up to " + "the length of the longest one among them; when it a certain positive " + "value, it must be greater than the length of the longest original " + "sequence.") + .SetDefault(-1); + AddComment(R"DOC( + Sequence Pad Operator + + This operator pads sequences in a same batch to a consistent length. + The length is specified by attribute 'padded_length'. New elements, + whose values are specified by input 'PadValue', will be appended to + the end of each sequence, to make their final lengths consistent. + + Following are cases to better explain how this works: + + Case 1: + + Given a 1-level LoDTensor input(X): + X.lod = [[0, 2, 5]] + X.data = [a, b, c, d, e] + and Input(PadValue): + PadValue.data = [0] + and attribite 'padded_length' = 4, + then we get LoDTensor: + Out.data = [[a, b, 0, 0], + [c, d, e, 0]] + + Case 2: + + Given a 1-level LoDTensor input(X): + X.lod = [[0, 2, 5]] + X.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]] + and Input(PadValue): + PadValue.data = [0] + and attribite 'padded_length' = -1, which mean using the length + of longest input sequence(3 in this case), + then we get LoDTensor: + Out.data = [[[a1, a2], [b1, b2], [0, 0]], + [[c1, c2], [d1, d2], [e1, e2]]] + + Case 3: + + Given a 1-level LoDTensor input(X): + X.lod = [[0, 2, 5]] + X.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]] + and Input(PadValue): + PadValue.data = [p1, p2] + and attribite 'padded_length' = -1, which mean using the length + of longest input sequence(3 in this case), + then we get LoDTensor: + Out.data = [[[a1, a2], [b1, b2], [p1, p2]], + [[c1, c2], [d1, d2], [e1, e2]]] + + )DOC"); + } +}; + +class SequencePadGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequencePadGradOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) of SequencePadGradOp should not be null."); + + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ framework::GradVarName("X")); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(sequence_pad, ops::SequencePadOp, ops::SequencePadOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(sequence_pad_grad, ops::SequencePadGradOp); +REGISTER_OP_CPU_KERNEL( + sequence_pad, + ops::SequencePadOpKernel, + ops::SequencePadOpKernel, + ops::SequencePadOpKernel, + ops::SequencePadOpKernel); +REGISTER_OP_CPU_KERNEL( + sequence_pad_grad, + ops::SequencePadGradOpKernel, + ops::SequencePadGradOpKernel, + ops::SequencePadGradOpKernel, + ops::SequencePadGradOpKernel); diff --git a/paddle/fluid/operators/sequence_pad_op.cu b/paddle/fluid/operators/sequence_pad_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..ff8f81a2f0ec4a72befc3be2a5fc48c3a586c824 --- /dev/null +++ b/paddle/fluid/operators/sequence_pad_op.cu @@ -0,0 +1,29 @@ +/* Copyright (c) 2018 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. */ + +#include "paddle/fluid/operators/sequence_pad_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + sequence_pad, + ops::SequencePadOpKernel, + ops::SequencePadOpKernel, + ops::SequencePadOpKernel, + ops::SequencePadOpKernel); +REGISTER_OP_CUDA_KERNEL( + sequence_pad_grad, + ops::SequencePadGradOpKernel, + ops::SequencePadGradOpKernel, + ops::SequencePadGradOpKernel, + ops::SequencePadGradOpKernel); diff --git a/paddle/fluid/operators/sequence_pad_op.h b/paddle/fluid/operators/sequence_pad_op.h new file mode 100644 index 0000000000000000000000000000000000000000..5fc9da69d787ff3aeffa716689d44772ad8f7bd2 --- /dev/null +++ b/paddle/fluid/operators/sequence_pad_op.h @@ -0,0 +1,66 @@ +/* Copyright (c) 2018 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. */ + +#pragma once + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/memory/memcpy.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/sequence_padding.h" + +namespace paddle { +namespace operators { + +using LoDTensor = framework::LoDTensor; +using LoD = framework::LoD; + +template +class SequencePadOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + const auto* x = ctx.Input("X"); + auto* out = ctx.Output("Out"); + out->mutable_data(ctx.GetPlace()); + + const auto* pad_value = ctx.Input("PadValue"); + + int padded_length = ctx.Attr("padded_length"); + + math::PaddingLoDTensorFunctor()( + ctx.template device_context(), *x, out, *pad_value, + padded_length, 0, false, math::kBatchLengthWidth); + } +}; + +template +class SequencePadGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* d_x = ctx.Output(framework::GradVarName("X")); + if (d_x) { + const auto* d_out = ctx.Input(framework::GradVarName("Out")); + d_x->mutable_data(ctx.GetPlace()); + + int padded_length = ctx.Attr("padded_length"); + + math::UnpaddingLoDTensorFunctor()( + ctx.template device_context(), *d_out, d_x, + padded_length, 0, false, math::kBatchLengthWidth); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/shrink_rnn_memory_op.cc b/paddle/fluid/operators/shrink_rnn_memory_op.cc index 8146c5f56104b7dec86b1c4491ed10fc2e94b58b..29d2fb989754f5621222768a279a1c898ea1c355 100644 --- a/paddle/fluid/operators/shrink_rnn_memory_op.cc +++ b/paddle/fluid/operators/shrink_rnn_memory_op.cc @@ -62,7 +62,10 @@ class ShrinkRNNMemoryOp : public ArrayOp { } if (dst_num_rows != 0) { - out_tensor.ShareDataWith(x_tensor.Slice(0, height)); + out_tensor.mutable_data(place, x_tensor.type()); + auto dev_ctx = platform::DeviceContextPool::Instance().Get(place); + framework::TensorCopy(x_tensor.Slice(0, height), place, *dev_ctx, + &out_tensor); } } }; diff --git a/paddle/fluid/operators/tensorrt_engine_op.cc b/paddle/fluid/operators/tensorrt_engine_op.cc index ee3078876c15b06a887064f08dc0c05d450b5f77..1048d3017140c9e31426a1580b2862667116a024 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.cc +++ b/paddle/fluid/operators/tensorrt_engine_op.cc @@ -17,112 +17,16 @@ #include #include -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" -#include "paddle/fluid/inference/tensorrt/engine.h" -#include "paddle/fluid/inference/utils/singleton.h" #include "paddle/fluid/operators/tensorrt_engine_op.h" namespace paddle { DEFINE_int32(tensorrt_engine_batch_size, 1, "the batch_size of TensorRT"); +DEFINE_int32(tensorrt_max_batch_size, 1, "TensorRT maximum batch size"); +DEFINE_int32(tensorrt_workspace_size, 16 << 20, "TensorRT workspace size"); namespace operators { -using inference::Singleton; -using inference::tensorrt::TRT_EngineManager; - -using FluidDT = framework::proto::VarType_Type; -using TRT_DT = nvinfer1::DataType; - -namespace { - -TRT_DT FluidDataType2TRT(FluidDT type) { - switch (type) { - case FluidDT::VarType_Type_FP32: - return TRT_DT::kFLOAT; - case FluidDT::VarType_Type_INT32: - return TRT_DT::kINT32; - default: - return TRT_DT::kINT32; - } - PADDLE_THROW("unkown type"); - return TRT_DT::kINT32; -} - -nvinfer1::Dims Vec2TRT_Dims(const std::vector &shape) { - PADDLE_ENFORCE_GT(shape.size(), 1UL, - "TensorRT' tensor input requires at least 2 dimensions"); - PADDLE_ENFORCE_LE(shape.size(), 4UL, - "TensorRT' tensor input requires at most 4 dimensions"); - PADDLE_ENFORCE_EQ(shape.size(), 4UL); - return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]); -} - -} // namespace - -template -void TensorRTEngineKernel::Prepare( - const framework::ExecutionContext &context) const { - VLOG(4) << "Prepare engine"; - // Get the ProgramDesc and pass to convert. - framework::proto::BlockDesc block_desc; - block_desc.ParseFromString(context.Attr("subgraph")); - int max_batch = context.Attr("max_batch"); - auto max_workspace = context.Attr("max_workspace"); - auto params = context.Attr>("parameters"); - std::unordered_set parameters; - for (const auto ¶m : params) { - parameters.insert(param); - } - - std::vector output_maps = - context.Attr>("output_name_mapping"); - - // TODO(Superjomn) replace this with a different stream - auto *engine = Singleton::Global().Create( - max_batch, max_workspace, nullptr /*engine hold its own stream*/, - context.Attr("engine_uniq_key")); - engine->InitNetwork(); - - framework::BlockDesc block(nullptr /*programdesc*/, &block_desc); - VLOG(4) << "parsed var size " << block.AllVars().size(); - // Add inputs - VLOG(4) << "declare inputs"; - for (auto &input : context.Inputs("Xs")) { - if (parameters.count(input)) continue; - VLOG(4) << "declare input " << input; - auto *var = block.FindVar(input); - // TensorRT engine need to create parameters. The parameter's description - // should be set in - PADDLE_ENFORCE(var, "no variable called %s", input); - PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR, - "TensorRT engine only takes LoDTensor as input"); - auto shape = var->GetShape(); - // For the special batch_size placeholder -1, drop it and pass the real - // shape of data. - // TODO(Superjomn) fix this with batch broadcast, or it can't handle - // variational batch size. - if (shape[0] == -1) { - shape[0] = FLAGS_tensorrt_engine_batch_size; - } - engine->DeclareInput( - input, FluidDataType2TRT( - var->Proto()->type().lod_tensor().tensor().data_type()), - Vec2TRT_Dims(shape)); - } - - inference::Singleton::Global().ConvertBlock( - block_desc, parameters, context.scope(), engine); - - // Add outputs - for (auto &output : output_maps) { - engine->DeclareOutput(output); - } - - engine->FreezeNetwork(); -} - class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { @@ -130,8 +34,6 @@ class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("Ys", "A list of outputs").AsDuplicable(); AddAttr("subgraph", "the subgraph."); AddAttr("engine_uniq_key", "unique key for the TRT engine."); - AddAttr("max_batch", "the maximum batch size."); - AddAttr("max_workspace", "the maximum batch size."); AddComment("TensorRT engine operator."); } }; @@ -150,11 +52,4 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(tensorrt_engine, ops::TensorRTEngineOp, ops::TensorRTEngineOpMaker, ops::TensorRTEngineOpMaker); -REGISTER_OP_CPU_KERNEL( - tensorrt_engine, - ops::TensorRTEngineKernel, - ops::TensorRTEngineKernel, - ops::TensorRTEngineKernel, - ops::TensorRTEngineKernel); - #endif // PADDLE_WITH_CUDA diff --git a/paddle/fluid/operators/tensorrt_engine_op.cu.cc b/paddle/fluid/operators/tensorrt_engine_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..e1ddfde6d51ef719ca0b89cf286b176195ee682a --- /dev/null +++ b/paddle/fluid/operators/tensorrt_engine_op.cu.cc @@ -0,0 +1,24 @@ +/* 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. */ + +#include "paddle/fluid/operators/tensorrt_engine_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_CUDA_KERNEL( + tensorrt_engine, + ops::TensorRTEngineKernel, + ops::TensorRTEngineKernel, + ops::TensorRTEngineKernel, + ops::TensorRTEngineKernel); diff --git a/paddle/fluid/operators/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt_engine_op.h index 2cbe1213a2f428a3ce56b06f97636baeb4b66c26..bc556ab3643cefa3e45d2a8a3835937753af723f 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.h +++ b/paddle/fluid/operators/tensorrt_engine_op.h @@ -19,16 +19,51 @@ #include #include +#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/inference/analysis/helper.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/tensorrt/engine.h" namespace paddle { DECLARE_int32(tensorrt_engine_batch_size); +DECLARE_int32(tensorrt_max_batch_size); +DECLARE_int32(tensorrt_workspace_size); namespace operators { +using FluidDT = framework::proto::VarType_Type; +using TRT_DT = nvinfer1::DataType; + +namespace { + +TRT_DT FluidDataType2TRT(FluidDT type) { + switch (type) { + case FluidDT::VarType_Type_FP32: + return TRT_DT::kFLOAT; + case FluidDT::VarType_Type_INT32: + return TRT_DT::kINT32; + default: + return TRT_DT::kINT32; + } + PADDLE_THROW("unkown type"); + return TRT_DT::kINT32; +} + +nvinfer1::Dims Vec2TRT_Dims(const std::vector& shape) { + PADDLE_ENFORCE_GT(shape.size(), 1UL, + "TensorRT' tensor input requires at least 2 dimensions"); + PADDLE_ENFORCE_LE(shape.size(), 4UL, + "TensorRT' tensor input requires at most 4 dimensions"); + PADDLE_ENFORCE(shape.size() == 4UL || shape.size() == 2UL); + if (shape.size() == 4UL) + return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]); + return nvinfer1::DimsCHW(shape[1], 1, 1); +} + +} // namespace + using inference::Singleton; using inference::tensorrt::TRT_EngineManager; @@ -47,7 +82,7 @@ class TensorRTEngineOp : public framework::OperatorWithKernel { .FindVar(input0) ->GetMutable() ->type()), - platform::CPUPlace()); + ctx.GetPlace()); return kt; } }; @@ -64,7 +99,7 @@ class TensorRTEngineKernel : public framework::OpKernel { auto input_names = context.op().Inputs("Xs"); PADDLE_ENFORCE(!input_names.empty(), "should pass more than one inputs"); PADDLE_ENFORCE_LE(FLAGS_tensorrt_engine_batch_size, - context.Attr("max_batch")); + FLAGS_tensorrt_max_batch_size); std::vector output_maps = context.Attr>("output_name_mapping"); @@ -94,12 +129,19 @@ class TensorRTEngineKernel : public framework::OpKernel { // Convert output tensor from engine to fluid int output_index = 0; + VLOG(4) << "TensorRT Engine Op Outputs:"; for (const auto& y : context.Outputs("Ys")) { + VLOG(4) << y; // convert output and copy to fluid. nvinfer1::ITensor* trt_t = engine->GetITensor(output_maps[output_index]); auto dims = trt_t->getDimensions(); // Use the output ITensor's dims to reshape the Fluid Tensor. - std::vector ddim(dims.d, dims.d + dims.nbDims); + // The ITensor doesn't contain the batch size dim. + std::vector ddim; + ddim.push_back(FLAGS_tensorrt_engine_batch_size); + for (int i = 0; i < dims.nbDims; i++) { + ddim.push_back(dims.d[i]); + } auto* fluid_v = context.scope().FindVar(y); PADDLE_ENFORCE_NOT_NULL(fluid_v, "no output variable called %s", y); @@ -113,9 +155,11 @@ class TensorRTEngineKernel : public framework::OpKernel { // TODO(Superjomn) change this float to dtype size. auto size = inference::analysis::AccuDims(dims.d, dims.nbDims) * FLAGS_tensorrt_engine_batch_size; - engine->GetOutputInCPU(output_maps[output_index], - fluid_t->mutable_data(platform::CPUPlace()), - size * sizeof(float)); + engine->GetOutputInGPU( + output_maps[output_index], + fluid_t->mutable_data(platform::CUDAPlace( + boost::get(context.GetPlace()).device)), + size * sizeof(float)); //} else { // engine->GetOutputInGPU( // y, fluid_t->mutable_data(platform::CUDAPlace()), @@ -128,8 +172,67 @@ class TensorRTEngineKernel : public framework::OpKernel { } protected: - // Build the engine. - void Prepare(const framework::ExecutionContext& context) const; + void Prepare(const framework::ExecutionContext& context) const { + VLOG(4) << "Prepare engine"; + // Get the ProgramDesc and pass to convert. + framework::proto::BlockDesc block_desc; + block_desc.ParseFromString(context.Attr("subgraph")); + int max_batch = FLAGS_tensorrt_max_batch_size; + auto max_workspace = FLAGS_tensorrt_workspace_size; + auto params = context.Attr>("parameters"); + std::unordered_set parameters; + for (const auto& param : params) { + parameters.insert(param); + } + + std::vector output_maps = + context.Attr>("output_name_mapping"); + + // TODO(Superjomn) replace this with a different stream + auto* engine = Singleton::Global().Create( + max_batch, max_workspace, nullptr /*engine hold its own stream*/, + context.Attr("engine_uniq_key"), + boost::get(context.GetPlace()).device); + + engine->InitNetwork(); + + framework::BlockDesc block(nullptr /*programdesc*/, &block_desc); + VLOG(4) << "parsed var size " << block.AllVars().size(); + // Add inputs + VLOG(4) << "declare inputs"; + for (auto& input : context.Inputs("Xs")) { + if (parameters.count(input)) continue; + VLOG(4) << "declare input " << input; + auto* var = block.FindVar(input); + // TensorRT engine need to create parameters. The parameter's description + // should be set in + PADDLE_ENFORCE(var, "no variable called %s", input); + PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR, + "TensorRT engine only takes LoDTensor as input"); + auto shape = var->GetShape(); + // For the special batch_size placeholder -1, drop it and pass the real + // shape of data. + // TODO(Superjomn) fix this with batch broadcast, or it can't handle + // variational batch size. + if (shape[0] == -1) { + shape[0] = FLAGS_tensorrt_engine_batch_size; + } + engine->DeclareInput( + input, FluidDataType2TRT( + var->Proto()->type().lod_tensor().tensor().data_type()), + Vec2TRT_Dims(shape)); + } + + inference::Singleton::Global() + .ConvertBlock(block_desc, parameters, context.scope(), engine); + + // Add outputs + for (auto& output : output_maps) { + engine->DeclareOutput(output); + } + + engine->FreezeNetwork(); + } }; } // namespace operators diff --git a/paddle/fluid/operators/tensorrt_engine_op_test.cc b/paddle/fluid/operators/tensorrt_engine_op_test.cc index 37657fa0b0498986fe67027415279af1775e58b9..27c1d29762b3de5e57f877b271aae52e71eb7cf9 100644 --- a/paddle/fluid/operators/tensorrt_engine_op_test.cc +++ b/paddle/fluid/operators/tensorrt_engine_op_test.cc @@ -12,6 +12,7 @@ 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. */ +#include "paddle/fluid/operators/tensorrt_engine_op.h" #include #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/framework/lod_tensor.h" @@ -23,20 +24,20 @@ limitations under the License. */ #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" -USE_CPU_ONLY_OP(tensorrt_engine); +USE_CUDA_ONLY_OP(tensorrt_engine); namespace paddle { namespace operators { namespace { -void CreateCPUTensor(framework::Scope* scope, const std::string& name, - const std::vector& shape) { +void CreateCUDATensor(framework::Scope* scope, const std::string& name, + const std::vector& shape) { auto* var = scope->Var(name); auto* tensor = var->GetMutable(); auto dims = framework::make_ddim(shape); tensor->Resize(dims); - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); inference::tensorrt::RandomizeTensor(tensor, place, ctx); } @@ -57,6 +58,8 @@ void AddTensorToBlockDesc(framework::proto::BlockDesc* block, using inference::analysis::SetAttr; TEST(TensorRTEngineOp, manual) { + FLAGS_tensorrt_engine_batch_size = 2; + FLAGS_tensorrt_max_batch_size = 2; framework::ProgramDesc program; auto* block_ = program.Proto()->add_blocks(); block_->set_idx(0); @@ -98,8 +101,6 @@ TEST(TensorRTEngineOp, manual) { engine_op_desc.SetOutput("Ys", std::vector({"z0"})); SetAttr(engine_op_desc.Proto(), "subgraph", block_->SerializeAsString()); - SetAttr(engine_op_desc.Proto(), "max_batch", 100); - SetAttr(engine_op_desc.Proto(), "max_workspace", 1 << 10); SetAttr(engine_op_desc.Proto(), "engine_uniq_key", "a_engine"); SetAttr>(engine_op_desc.Proto(), "parameters", std::vector({})); @@ -112,15 +113,15 @@ TEST(TensorRTEngineOp, manual) { LOG(INFO) << "engine_op " << engine_op.get(); framework::Scope scope; - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); // Prepare variables. - CreateCPUTensor(&scope, "x", std::vector({2, 4})); - CreateCPUTensor(&scope, "y", std::vector({4, 6})); - CreateCPUTensor(&scope, "z", std::vector({2, 6})); + CreateCUDATensor(&scope, "x", std::vector({2, 4})); + CreateCUDATensor(&scope, "y", std::vector({4, 6})); + CreateCUDATensor(&scope, "z", std::vector({2, 6})); - CreateCPUTensor(&scope, "y0", std::vector({6, 8})); - CreateCPUTensor(&scope, "z0", std::vector({2, 8})); + CreateCUDATensor(&scope, "y0", std::vector({6, 8})); + CreateCUDATensor(&scope, "z0", std::vector({2, 8})); // Execute them. LOG(INFO) << "engine_op run"; @@ -128,10 +129,12 @@ TEST(TensorRTEngineOp, manual) { } void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) { + FLAGS_tensorrt_engine_batch_size = batch_size; + FLAGS_tensorrt_max_batch_size = batch_size; framework::ProgramDesc program; framework::Scope scope; - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); auto* block_ = program.Proto()->add_blocks(); block_->set_idx(0); @@ -165,10 +168,10 @@ void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) { // Prepare variables. if (!x_created) { - CreateCPUTensor(&scope, x_name, std::vector(x_shape)); + CreateCUDATensor(&scope, x_name, std::vector(x_shape)); } - CreateCPUTensor(&scope, y_name, std::vector(y_shape)); - CreateCPUTensor(&scope, z_name, std::vector(z_shape)); + CreateCUDATensor(&scope, y_name, std::vector(y_shape)); + CreateCUDATensor(&scope, z_name, std::vector(z_shape)); // It is wired, need to copy manually. *block_->add_ops() = *fc->Proto(); diff --git a/paddle/fluid/operators/warpctc_op.h b/paddle/fluid/operators/warpctc_op.h index ab70c1f0592d122ba248a101db487e64c0bdae6f..444265f58de732f07c5db2abd87811a063016866 100644 --- a/paddle/fluid/operators/warpctc_op.h +++ b/paddle/fluid/operators/warpctc_op.h @@ -153,17 +153,29 @@ class WarpCTCKernel : public framework::OpKernel { framework::make_ddim({static_cast(num_sequences), 1}); // warpctc needs sequences data stored in transposed padding format - Tensor warpctc_logits; + LoDTensor warpctc_logits; const size_t max_sequence_length = - math::MaximumSequenceLength(logits_lod, level); + math::MaximumSequenceLength(logits_lod[level]); auto warpctc_logits_dims = framework::make_ddim({static_cast(max_sequence_length), static_cast(num_sequences), static_cast(sequence_width)}); warpctc_logits.mutable_data(warpctc_logits_dims, ctx.GetPlace()); + + LoDTensor cpu_pad_value; + T* pad_value_data = + cpu_pad_value.mutable_data({1}, platform::CPUPlace()); + *pad_value_data = static_cast(0); + LoDTensor pad_value; + if (platform::is_cpu_place(ctx.GetPlace())) { + pad_value = cpu_pad_value; + } else { + TensorCopySync(cpu_pad_value, ctx.GetPlace(), &pad_value); + } + math::PaddingLoDTensorFunctor()( ctx.template device_context(), *logits, &warpctc_logits, - false); + pad_value, -1, 0, false /* norm_by_times */, math::kLengthBatchWidth); const T* warpctc_logits_data = warpctc_logits.data(); std::vector warpctc_label_lengths(num_sequences); @@ -209,15 +221,15 @@ template class WarpCTCGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - auto* warpctc_grad = ctx.Input("WarpCTCGrad"); + auto* warpctc_grad = ctx.Input("WarpCTCGrad"); auto* logits_grad = ctx.Output(framework::GradVarName("Logits")); const Tensor* loss_grad = ctx.Input(framework::GradVarName("Loss")); logits_grad->mutable_data(ctx.GetPlace()); bool norm_by_times = ctx.Attr("norm_by_times"); math::UnpaddingLoDTensorFunctor()( - ctx.template device_context(), logits_grad, - *warpctc_grad, norm_by_times); + ctx.template device_context(), *warpctc_grad, + logits_grad, -1, 0, norm_by_times, math::kLengthBatchWidth); const T* loss_grad_data = loss_grad->data(); math::ScaleLoDTensorFunctor()( diff --git a/paddle/fluid/operators/while_op.cc b/paddle/fluid/operators/while_op.cc index 733157ea05ed39434b9a750e3a94ea548f512ce6..147b5b1d1db61b534a94f65c55fe205840547920 100644 --- a/paddle/fluid/operators/while_op.cc +++ b/paddle/fluid/operators/while_op.cc @@ -55,6 +55,7 @@ class WhileOp : public framework::OperatorBase { auto step_scopes = scope.FindVar(Output(kStepScopes))->GetMutable(); + bool is_test = Attr("is_test"); PADDLE_ENFORCE(platform::is_cpu_place(cond.place()), "Condition of while op must in CPU memory."); while (cond.data()[0]) { @@ -63,6 +64,10 @@ class WhileOp : public framework::OperatorBase { executor.Run(*program, ¤t_scope, block->ID(), false /*create_local_scope*/); + + if (is_test) { + scope.DeleteScope(¤t_scope); + } } } }; @@ -88,6 +93,7 @@ class WhileOpMaker : public framework::OpProtoAndCheckerMaker { "variables generated in the i'th step."); AddAttr(kStepBlock, "The step block inside WhileOp"); + AddAttr("is_test", "True if in test phase.").SetDefault(false); AddComment(R"DOC( )DOC"); } @@ -103,6 +109,8 @@ class WhileGradOp : public framework::OperatorBase { private: void RunImpl(const framework::Scope &scope, const platform::Place &dev_place) const override { + PADDLE_ENFORCE(!Attr("is_test"), + "GradOp is only callable when is_test is false"); // get device context from pool platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); auto &dev_ctx = *pool.Get(dev_place); diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h index 88e0383146c1adf2752a362091996bad9cfcce5e..b97dad20db0b003b4886b7c7cfd1c8de8bf44ab9 100644 --- a/paddle/fluid/platform/device_context.h +++ b/paddle/fluid/platform/device_context.h @@ -24,7 +24,7 @@ limitations under the License. */ #endif #ifdef PADDLE_WITH_MKLDNN -#include +#include "mkldnn.hpp" #endif #include diff --git a/python/paddle/dataset/image.py b/python/paddle/dataset/image.py index b32736ee7c265e3a94207afc04673eec4fcf1c6e..920dbf3b4ebb0bc3d98c9ea986d7d039deed4a4c 100644 --- a/python/paddle/dataset/image.py +++ b/python/paddle/dataset/image.py @@ -203,7 +203,7 @@ def resize_short(im, size): h_new = size * h // w else: w_new = size * w // h - im = cv2.resize(im, (h_new, w_new), interpolation=cv2.INTER_CUBIC) + im = cv2.resize(im, (w_new, h_new), interpolation=cv2.INTER_CUBIC) return im @@ -345,7 +345,6 @@ def simple_transform(im, if np.random.randint(2) == 0: im = left_right_flip(im, is_color) else: - im = center_crop(im, crop_size, is_color) im = center_crop(im, crop_size, is_color=is_color) if len(im.shape) == 3: im = to_chw(im) diff --git a/python/paddle/fluid/inferencer.py b/python/paddle/fluid/inferencer.py index 3d2ef566173f81b29a6d8ea79cff00991a4ef3c4..a9b94a20720615dbfca97749463f27dbc88ac64f 100644 --- a/python/paddle/fluid/inferencer.py +++ b/python/paddle/fluid/inferencer.py @@ -98,10 +98,9 @@ class Inferencer(object): raise ValueError( "inputs should be a map of {'input_name': input_var}") - with executor.scope_guard(self.scope): - results = self.exe.run(self.inference_program, - feed=inputs, - fetch_list=[self.predict_var], + with self._prog_and_scope_guard(): + results = self.exe.run(feed=inputs, + fetch_list=[self.predict_var.name], return_numpy=return_numpy) return results diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 173567a0a374ac4453025b67b047950936df2055..3d9183e7cf92f68362fc05aa2e19b526270fc1ee 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -661,6 +661,7 @@ class While(object): Args: cond (Variable): condition used to compare. + is_test(bool): A flag indicating whether execution is in test phase. name (str): The name of this layer. Examples: @@ -683,7 +684,7 @@ class While(object): IN_WHILE_BLOCK = 1 AFTER_WHILE_BLOCK = 2 - def __init__(self, cond, name=None): + def __init__(self, cond, is_test=False, name=None): self.helper = LayerHelper("while", name=name) self.status = While.BEFORE_WHILE_BLOCK if not isinstance(cond, Variable): @@ -694,6 +695,7 @@ class While(object): if reduce(lambda a, b: a * b, cond.shape, 1) != 1: raise TypeError("condition should be a bool scalar") self.cond_var = cond + self.is_test = is_test def block(self): return WhileGuard(self) @@ -735,7 +737,8 @@ class While(object): }, outputs={'Out': out_vars, 'StepScopes': [step_scope]}, - attrs={'sub_block': while_block}) + attrs={'sub_block': while_block, + "is_test": self.is_test}) def lod_rank_table(x, level=0): diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index b03ee514f50f9a8c1425bd5b1d409b58ed62351a..0cf7aaef4ab75ca6976465d1b404004a9f2f64c5 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -246,7 +246,11 @@ def Send(endpoints, send_vars, dummy_output=None, sync=True): rpc_op_role_name: core.op_proto_and_checker_maker.OpRole.RPC }) if sync: - helper.append_op(type="send_barrier", attrs={"endpoints": endpoints}) + helper.append_op( + type="send_barrier", + inputs={"X": dummy_output}, + outputs={"Out": []}, + attrs={"endpoints": endpoints}) def Recv(endpoints, get_vars, dummy_input=None, sync=True): @@ -282,7 +286,10 @@ def Recv(endpoints, get_vars, dummy_input=None, sync=True): attrs={"endpoints": endpoints, "epmap": epmap}) if sync: - helper.append_op(type="fetch_barrier", attrs={"endpoints": endpoints}) + helper.append_op( + type="fetch_barrier", + outputs={"Out": get_vars}, + attrs={"endpoints": endpoints}) return get_vars diff --git a/python/paddle/fluid/layers/metric_op.py b/python/paddle/fluid/layers/metric_op.py index 2c3bdd77e1fa1c86baa3a288caab4ad4324e2ef2..0182bbeb637ec7b6a341a4822a1cc5fb5aef077d 100644 --- a/python/paddle/fluid/layers/metric_op.py +++ b/python/paddle/fluid/layers/metric_op.py @@ -119,10 +119,14 @@ def auc(input, label, curve='ROC', num_thresholds=200, topk=1): helper = LayerHelper("auc", **locals()) auc_out = helper.create_tmp_variable(dtype="float64") # make tp, tn, fp, fn persistable, so that can accumulate all batches. - tp = helper.create_global_variable(persistable=True, dtype='int64') - tn = helper.create_global_variable(persistable=True, dtype='int64') - fp = helper.create_global_variable(persistable=True, dtype='int64') - fn = helper.create_global_variable(persistable=True, dtype='int64') + tp = helper.create_global_variable( + persistable=True, dtype='int64', shape=[num_thresholds]) + tn = helper.create_global_variable( + persistable=True, dtype='int64', shape=[num_thresholds]) + fp = helper.create_global_variable( + persistable=True, dtype='int64', shape=[num_thresholds]) + fn = helper.create_global_variable( + persistable=True, dtype='int64', shape=[num_thresholds]) for var in [tp, tn, fp, fn]: helper.set_variable_initializer( var, Constant( diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index c5c92619ed6d61b17a06d4c0f671eeb1d8c04d17..2b51c80413a660f6922b971e9f4c623635b698dc 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -55,6 +55,7 @@ __all__ = [ 'conv2d_transpose', 'conv3d_transpose', 'sequence_expand', + 'sequence_pad', 'lstm_unit', 'reduce_sum', 'reduce_mean', @@ -2657,6 +2658,51 @@ def sequence_expand(x, y, ref_level=-1, name=None): return tmp +@templatedoc() +def sequence_pad(x, pad_value, maxlen=None): + """ + ${comment} + + Args: + x(Variable): Input variable which should contain lod information. + pad_value(Variable): The Variable that holds values that will be fill + into padded steps. It can be a scalar or a tensor whose shape + equals to time steps in sequences. If it's a scalar, it will be + automatically broadcasted to the shape of time step. + maxlen(int, default None): The length of padded sequences. It can be + None or any positive int. When it is None, all sequences will be + padded up to the length of the longest one among them; when it a + certain positive value, it must be greater than the length of the + longest original sequence." + + Returns: + Variable: The padded sequence batch. All sequences has the same length. + + Examples: + .. code-block:: python + + import numpy + + x = fluid.layers.data(name='y', shape=[10, 5], + dtype='float32', lod_level=1) + pad_value = fluid.layers.assign(input=numpy.array([0])) + out = fluid.layers.sequence_pad(x=x, pad_value=pad_value) + """ + + helper = LayerHelper('sequence_pad', input=x, **locals()) + dtype = helper.input_dtype() + out = helper.create_tmp_variable(dtype) + if maxlen is None: + maxlen = -1 + helper.append_op( + type='sequence_pad', + inputs={'X': x, + 'PadValue': pad_value}, + outputs={'Out': out}, + attrs={'padded_length': maxlen}) + return out + + def beam_search(pre_ids, pre_scores, ids, diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py index be494a0d340c62fb35afbf97fba38eff08a965e6..2e15c224f662171bf0fee228bdd9d36189fbe499 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py @@ -16,7 +16,9 @@ from __future__ import print_function import paddle import paddle.fluid as fluid +import paddle.fluid.core as core import numpy +import os import cifar10_small_test_set @@ -89,7 +91,7 @@ def optimizer_func(): return fluid.optimizer.Adam(learning_rate=0.001) -def train(use_cuda, train_program, params_dirname): +def train(use_cuda, train_program, parallel, params_dirname): BATCH_SIZE = 128 EPOCH_NUM = 1 @@ -116,7 +118,10 @@ def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_func=train_program, optimizer_func=optimizer_func, place=place) + train_func=train_program, + optimizer_func=optimizer_func, + place=place, + parallel=parallel) trainer.train( reader=train_reader, @@ -125,10 +130,13 @@ def train(use_cuda, train_program, params_dirname): feed_order=['pixel', 'label']) -def infer(use_cuda, inference_program, params_dirname=None): +def infer(use_cuda, inference_program, parallel, params_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() inferencer = fluid.Inferencer( - infer_func=inference_program, param_path=params_dirname, place=place) + infer_func=inference_program, + param_path=params_dirname, + place=place, + parallel=parallel) # The input's dimension of conv should be 4-D or 5-D. # Use normilized image pixels as input data, which should be in the range @@ -139,22 +147,34 @@ def infer(use_cuda, inference_program, params_dirname=None): print("infer results: ", results) -def main(use_cuda): +def main(use_cuda, parallel): if use_cuda and not fluid.core.is_compiled_with_cuda(): return save_path = "image_classification_resnet.inference.model" + os.environ['CPU_NUM'] = str(4) train( use_cuda=use_cuda, train_program=train_network, - params_dirname=save_path) + params_dirname=save_path, + parallel=parallel) + # FIXME(zcd): in the inference stage, the number of + # input data is one, it is not appropriate to use parallel. + if parallel and use_cuda: + return + + os.environ['CPU_NUM'] = str(1) infer( use_cuda=use_cuda, inference_program=inference_network, - params_dirname=save_path) + params_dirname=save_path, + parallel=parallel) if __name__ == '__main__': for use_cuda in (False, True): - main(use_cuda=use_cuda) + for parallel in (False, True): + if use_cuda and not core.is_compiled_with_cuda(): + continue + main(use_cuda=use_cuda, parallel=parallel) diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py index dbc7bc06c93157f271c79e85b6925468e861e57f..2f205de1c011cd714439d4896adc8862ce68d99b 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py @@ -16,7 +16,9 @@ from __future__ import print_function import paddle import paddle.fluid as fluid +import paddle.fluid.core as core import numpy +import os import cifar10_small_test_set @@ -68,7 +70,7 @@ def optimizer_func(): return fluid.optimizer.Adam(learning_rate=0.001) -def train(use_cuda, train_program, params_dirname): +def train(use_cuda, train_program, parallel, params_dirname): BATCH_SIZE = 128 train_reader = paddle.batch( paddle.reader.shuffle( @@ -93,7 +95,10 @@ def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_func=train_program, place=place, optimizer_func=optimizer_func) + train_func=train_program, + place=place, + optimizer_func=optimizer_func, + parallel=parallel) trainer.train( reader=train_reader, @@ -102,10 +107,13 @@ def train(use_cuda, train_program, params_dirname): feed_order=['pixel', 'label']) -def infer(use_cuda, inference_program, params_dirname=None): +def infer(use_cuda, inference_program, parallel, params_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() inferencer = fluid.Inferencer( - infer_func=inference_program, param_path=params_dirname, place=place) + infer_func=inference_program, + param_path=params_dirname, + place=place, + parallel=parallel) # The input's dimension of conv should be 4-D or 5-D. # Use normilized image pixels as input data, which should be in the range @@ -116,22 +124,31 @@ def infer(use_cuda, inference_program, params_dirname=None): print("infer results: ", results) -def main(use_cuda): - if use_cuda and not fluid.core.is_compiled_with_cuda(): - return +def main(use_cuda, parallel): save_path = "image_classification_vgg.inference.model" + os.environ['CPU_NUM'] = str(4) train( use_cuda=use_cuda, train_program=train_network, - params_dirname=save_path) + params_dirname=save_path, + parallel=parallel) + # FIXME(zcd): in the inference stage, the number of + # input data is one, it is not appropriate to use parallel. + if parallel and use_cuda: + return + os.environ['CPU_NUM'] = str(1) infer( use_cuda=use_cuda, inference_program=inference_network, - params_dirname=save_path) + params_dirname=save_path, + parallel=parallel) if __name__ == '__main__': for use_cuda in (False, True): - main(use_cuda=use_cuda) + for parallel in (False, True): + if use_cuda and not core.is_compiled_with_cuda(): + continue + main(use_cuda=use_cuda, parallel=parallel) diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py index 187bef1b0c1a614fbca88ef22097831d7bd5cd7f..a5adf68158526b628deba3fc7ca6856eb7c9cded 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py @@ -64,14 +64,14 @@ def optimizer_func(): return fluid.optimizer.Adam(learning_rate=0.001) -def train(use_cuda, train_program, params_dirname): +def train(use_cuda, train_program, parallel, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( train_func=train_program, place=place, optimizer_func=optimizer_func, - parallel=True) + parallel=parallel) def event_handler(event): if isinstance(event, fluid.EndEpochEvent): @@ -108,11 +108,14 @@ def train(use_cuda, train_program, params_dirname): feed_order=['img', 'label']) -def infer(use_cuda, inference_program, params_dirname=None): +def infer(use_cuda, inference_program, parallel, params_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() inferencer = fluid.Inferencer( - infer_func=inference_program, param_path=params_dirname, place=place) + infer_func=inference_program, + param_path=params_dirname, + place=place, + parallel=parallel) batch_size = 1 tensor_img = numpy.random.uniform(-1.0, 1.0, @@ -123,20 +126,32 @@ def infer(use_cuda, inference_program, params_dirname=None): print("infer results: ", results[0]) -def main(use_cuda): +def main(use_cuda, parallel): params_dirname = "recognize_digits_conv.inference.model" # call train() with is_local argument to run distributed train + os.environ['CPU_NUM'] = str(4) train( use_cuda=use_cuda, train_program=train_program, - params_dirname=params_dirname) + params_dirname=params_dirname, + parallel=parallel) + + # FIXME(zcd): in the inference stage, the number of + # input data is one, it is not appropriate to use parallel. + if parallel and use_cuda: + return + os.environ['CPU_NUM'] = str(1) infer( use_cuda=use_cuda, inference_program=inference_program, - params_dirname=params_dirname) + params_dirname=params_dirname, + parallel=parallel) if __name__ == '__main__': - # for use_cuda in (False, True): - main(use_cuda=core.is_compiled_with_cuda()) + for use_cuda in (False, True): + for parallel in (False, True): + if use_cuda and not core.is_compiled_with_cuda(): + continue + main(use_cuda=use_cuda, parallel=parallel) diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py index b95e7db122adbb1414da1691926c920b963fd6fe..e7d8b23b3253d368210c08be4e53c06ba0c5d618 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py @@ -16,6 +16,7 @@ from __future__ import print_function import argparse import paddle.fluid as fluid +import paddle.fluid.core as core import paddle import sys import numpy @@ -50,11 +51,14 @@ def optimizer_func(): return fluid.optimizer.Adam(learning_rate=0.001) -def train(use_cuda, train_program, params_dirname): +def train(use_cuda, train_program, params_dirname, parallel): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_func=train_program, place=place, optimizer_func=optimizer_func) + train_func=train_program, + place=place, + optimizer_func=optimizer_func, + parallel=parallel) def event_handler(event): if isinstance(event, fluid.EndEpochEvent): @@ -86,11 +90,14 @@ def train(use_cuda, train_program, params_dirname): feed_order=['img', 'label']) -def infer(use_cuda, inference_program, params_dirname=None): +def infer(use_cuda, inference_program, parallel, params_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() inferencer = fluid.Inferencer( - infer_func=inference_program, param_path=params_dirname, place=place) + infer_func=inference_program, + param_path=params_dirname, + place=place, + parallel=parallel) batch_size = 1 tensor_img = numpy.random.uniform(-1.0, 1.0, @@ -101,20 +108,32 @@ def infer(use_cuda, inference_program, params_dirname=None): print("infer results: ", results[0]) -def main(use_cuda): +def main(use_cuda, parallel): params_dirname = "recognize_digits_mlp.inference.model" # call train() with is_local argument to run distributed train + os.environ['CPU_NUM'] = str(4) train( use_cuda=use_cuda, train_program=train_program, - params_dirname=params_dirname) + params_dirname=params_dirname, + parallel=parallel) + + # FIXME(zcd): in the inference stage, the number of + # input data is one, it is not appropriate to use parallel. + if parallel and use_cuda: + return + os.environ['CPU_NUM'] = str(1) infer( use_cuda=use_cuda, inference_program=inference_program, - params_dirname=params_dirname) + params_dirname=params_dirname, + parallel=parallel) if __name__ == '__main__': - # for use_cuda in (False, True): - main(use_cuda=False) + for use_cuda in (False, True): + for parallel in (False, True): + if use_cuda and not core.is_compiled_with_cuda(): + continue + main(use_cuda=use_cuda, parallel=parallel) diff --git a/python/paddle/fluid/tests/unittests/dist_se_resnext.py b/python/paddle/fluid/tests/unittests/dist_se_resnext.py index b0ee6ff9f5941b090e663dac0122cfb575f2f442..5440a8c7fe7a5411c17bb6a9ea9f4e50c85d6796 100644 --- a/python/paddle/fluid/tests/unittests/dist_se_resnext.py +++ b/python/paddle/fluid/tests/unittests/dist_se_resnext.py @@ -130,7 +130,12 @@ class SE_ResNeXt(): input=conv, pool_size=7, pool_type='avg', global_pooling=True) drop = fluid.layers.dropout(x=pool, dropout_prob=0.2) stdv = 1.0 / math.sqrt(drop.shape[1] * 1.0) - out = fluid.layers.fc(input=drop, size=class_dim, act='softmax') + out = fluid.layers.fc( + input=drop, + size=class_dim, + act='softmax', + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.05))) return out def shortcut(self, input, ch_out, stride): @@ -180,7 +185,7 @@ class SE_ResNeXt(): act=None, # avoid pserver CPU init differs from GPU param_attr=fluid.ParamAttr( - initializer=fluid.initializer.Constant()), + initializer=fluid.initializer.Constant(value=0.05)), bias_attr=False) return fluid.layers.batch_norm(input=conv, act=act) @@ -188,13 +193,19 @@ class SE_ResNeXt(): pool = fluid.layers.pool2d( input=input, pool_size=0, pool_type='avg', global_pooling=True) stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0) - squeeze = fluid.layers.fc(input=pool, - size=num_channels // reduction_ratio, - act='relu') + squeeze = fluid.layers.fc( + input=pool, + size=num_channels // reduction_ratio, + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.05)), + act='relu') stdv = 1.0 / math.sqrt(squeeze.shape[1] * 1.0) - excitation = fluid.layers.fc(input=squeeze, - size=num_channels, - act='sigmoid') + excitation = fluid.layers.fc( + input=squeeze, + size=num_channels, + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.05)), + act='sigmoid') scale = fluid.layers.elementwise_mul(x=input, y=excitation, axis=0) return scale diff --git a/python/paddle/fluid/tests/unittests/dist_transformer.py b/python/paddle/fluid/tests/unittests/dist_transformer.py index ab4c5c3f368333ac42781aebd495579b5c26f388..7abfa0a4be0dec9fe251704e22dfef1f932e7c5b 100644 --- a/python/paddle/fluid/tests/unittests/dist_transformer.py +++ b/python/paddle/fluid/tests/unittests/dist_transformer.py @@ -18,54 +18,129 @@ import numpy as np import argparse import time import math +import os +import sys +import six +import argparse +import ast +import multiprocessing +import time +from functools import partial +from os.path import expanduser +import glob +import random +import tarfile import paddle import paddle.fluid as fluid +import paddle.fluid.layers as layers from paddle.fluid import core -import os -import sys -import six -import transformer_model -import paddle.dataset.wmt16 as wmt16 +from test_dist_base import TestDistRunnerBase, runtime_main +from paddle.compat import long_type + +import hashlib + +from paddle.fluid.transpiler.details import program_to_code + +const_para_attr = fluid.ParamAttr(initializer=fluid.initializer.Constant(0.001)) +const_bias_attr = const_para_attr # Fix seed for test fluid.default_startup_program().random_seed = 1 fluid.default_main_program().random_seed = 1 -WMT16_RECORDIO_FILE = "/tmp/wmt16.recordio" +#from transformer_config import ModelHyperParams, TrainTaskConfig, merge_cfg_from_list +class TrainTaskConfig(object): + # only support GPU currently + use_gpu = True + # the epoch number to train. + pass_num = 1 + # the number of sequences contained in a mini-batch. + # deprecated, set batch_size in args. + batch_size = 20 + # the hyper parameters for Adam optimizer. + # This static learning_rate will be multiplied to the LearningRateScheduler + # derived learning rate the to get the final learning rate. + learning_rate = 1 + beta1 = 0.9 + beta2 = 0.98 + eps = 1e-9 + # the parameters for learning rate scheduling. + warmup_steps = 4000 + # the weight used to mix up the ground-truth distribution and the fixed + # uniform distribution in label smoothing when training. + # Set this as zero if label smoothing is not wanted. + label_smooth_eps = 0.1 + # the directory for saving trained models. + model_dir = "trained_models" + # the directory for saving checkpoints. + ckpt_dir = "trained_ckpts" + # the directory for loading checkpoint. + # If provided, continue training from the checkpoint. + ckpt_path = None + # the parameter to initialize the learning rate scheduler. + # It should be provided if use checkpoints, since the checkpoint doesn't + # include the training step counter currently. + start_step = 0 -class ModelHyperParams(object): - # Dictionary size for source and target language. This model directly uses - # paddle.dataset.wmt16 in which , and token has - # alreay been added, but the token is not added. Transformer requires - # sequences in a mini-batch are padded to have the same length. A token is - # added into the original dictionary in paddle.dateset.wmt16. + check_acc = True - # size of source word dictionary. - src_vocab_size = 10000 - # index for token in source language. - src_pad_idx = src_vocab_size + data_path = expanduser("~") + ( + "/.cache/paddle/dataset/test_dist_transformer/") + src_vocab_fpath = data_path + "vocab.bpe.32000" + trg_vocab_fpath = data_path + "vocab.bpe.32000" + train_file_pattern = data_path + "train.tok.clean.bpe.32000.en-de" + val_file_pattern = data_path + "newstest2013.tok.bpe.32000.en-de" + pool_size = 2000 + sort_type = None + local = True + shuffle = False + shuffle_batch = False + special_token = ['', '', ''] + token_delimiter = ' ' + use_token_batch = False - # size of target word dictionay - trg_vocab_size = 10000 - # index for token in target language. - trg_pad_idx = trg_vocab_size - # position value corresponding to the token. - pos_pad_idx = 0 +class InferTaskConfig(object): + use_gpu = True + # the number of examples in one run for sequence generation. + batch_size = 10 + # the parameters for beam search. + beam_size = 5 + max_out_len = 256 + # the number of decoded sentences to output. + n_best = 1 + # the flags indicating whether to output the special tokens. + output_bos = False + output_eos = False + output_unk = True + # the directory for loading the trained model. + model_path = "trained_models/pass_1.infer.model" - # max length of sequences. It should plus 1 to include position - # padding token for position encoding. - max_length = 50 +class ModelHyperParams(object): + # These following five vocabularies related configurations will be set + # automatically according to the passed vocabulary path and special tokens. + # size of source word dictionary. + src_vocab_size = 10000 + # size of target word dictionay + trg_vocab_size = 10000 + # index for token + bos_idx = 0 + # index for token + eos_idx = 1 + # index for token + unk_idx = 2 + # max length of sequences deciding the size of position encoding table. + # Start from 1 and count start and end tokens in. + max_length = 256 # the dimension for word embeddings, which is also the last dimension of # the input and output of multi-head attention, position-wise feed-forward # networks, encoder and decoder. - d_model = 512 # size of the hidden layer in position-wise feed-forward networks. - d_inner_hid = 1024 + d_inner_hid = 2048 # the dimension that keys are projected to for dot-product attention. d_key = 64 # the dimension that values are projected to for dot-product attention. @@ -75,210 +150,1577 @@ class ModelHyperParams(object): # number of sub-layers to be stacked in the encoder and decoder. n_layer = 6 # dropout rate used by all dropout layers. - dropout = 0.1 + dropout = 0.0 # no random + # random seed used in dropout for CE. + dropout_seed = None + # the flag indicating whether to share embedding and softmax weights. + # vocabularies in source and target should be same for weight sharing. + weight_sharing = True + + +def merge_cfg_from_list(cfg_list, g_cfgs): + """ + Set the above global configurations using the cfg_list. + """ + assert len(cfg_list) % 2 == 0 + for key, value in zip(cfg_list[0::2], cfg_list[1::2]): + for g_cfg in g_cfgs: + if hasattr(g_cfg, key): + try: + value = eval(value) + except Exception: # for file path + pass + setattr(g_cfg, key, value) + break + + +# The placeholder for batch_size in compile time. Must be -1 currently to be +# consistent with some ops' infer-shape output in compile time, such as the +# sequence_expand op used in beamsearch decoder. +batch_size = -1 +# The placeholder for squence length in compile time. +seq_len = ModelHyperParams.max_length +# Here list the data shapes and data types of all inputs. +# The shapes here act as placeholder and are set to pass the infer-shape in +# compile time. +input_descs = { + # The actual data shape of src_word is: + # [batch_size * max_src_len_in_batch, 1] + "src_word": [(batch_size, seq_len, long_type(1)), "int64", 2], + # The actual data shape of src_pos is: + # [batch_size * max_src_len_in_batch, 1] + "src_pos": [(batch_size, seq_len, long_type(1)), "int64"], + # This input is used to remove attention weights on paddings in the + # encoder. + # The actual data shape of src_slf_attn_bias is: + # [batch_size, n_head, max_src_len_in_batch, max_src_len_in_batch] + "src_slf_attn_bias": [(batch_size, ModelHyperParams.n_head, seq_len, + seq_len), "float32"], + # The actual data shape of trg_word is: + # [batch_size * max_trg_len_in_batch, 1] + "trg_word": [(batch_size, seq_len, long_type(1)), "int64", + 2], # lod_level is only used in fast decoder. + # The actual data shape of trg_pos is: + # [batch_size * max_trg_len_in_batch, 1] + "trg_pos": [(batch_size, seq_len, long_type(1)), "int64"], + # This input is used to remove attention weights on paddings and + # subsequent words in the decoder. + # The actual data shape of trg_slf_attn_bias is: + # [batch_size, n_head, max_trg_len_in_batch, max_trg_len_in_batch] + "trg_slf_attn_bias": [(batch_size, ModelHyperParams.n_head, seq_len, + seq_len), "float32"], + # This input is used to remove attention weights on paddings of the source + # input in the encoder-decoder attention. + # The actual data shape of trg_src_attn_bias is: + # [batch_size, n_head, max_trg_len_in_batch, max_src_len_in_batch] + "trg_src_attn_bias": [(batch_size, ModelHyperParams.n_head, seq_len, + seq_len), "float32"], + # This input is used in independent decoder program for inference. + # The actual data shape of enc_output is: + # [batch_size, max_src_len_in_batch, d_model] + "enc_output": [(batch_size, seq_len, ModelHyperParams.d_model), "float32"], + # The actual data shape of label_word is: + # [batch_size * max_trg_len_in_batch, 1] + "lbl_word": [(batch_size * seq_len, long_type(1)), "int64"], + # This input is used to mask out the loss of paddding tokens. + # The actual data shape of label_weight is: + # [batch_size * max_trg_len_in_batch, 1] + "lbl_weight": [(batch_size * seq_len, long_type(1)), "float32"], + # These inputs are used to change the shape tensor in beam-search decoder. + "trg_slf_attn_pre_softmax_shape_delta": [(long_type(2), ), "int32"], + "trg_slf_attn_post_softmax_shape_delta": [(long_type(4), ), "int32"], + "init_score": [(batch_size, long_type(1)), "float32"], +} + +# Names of word embedding table which might be reused for weight sharing. +word_emb_param_names = ( + "src_word_emb_table", + "trg_word_emb_table", ) +# Names of position encoding table which will be initialized externally. +pos_enc_param_names = ( + "src_pos_enc_table", + "trg_pos_enc_table", ) +# separated inputs for different usages. +encoder_data_input_fields = ( + "src_word", + "src_pos", + "src_slf_attn_bias", ) +decoder_data_input_fields = ( + "trg_word", + "trg_pos", + "trg_slf_attn_bias", + "trg_src_attn_bias", + "enc_output", ) +label_data_input_fields = ( + "lbl_word", + "lbl_weight", ) +# In fast decoder, trg_pos (only containing the current time step) is generated +# by ops and trg_slf_attn_bias is not needed. +fast_decoder_data_input_fields = ( + "trg_word", + "init_score", + "trg_src_attn_bias", ) + +# fast_decoder_util_input_fields = ( +# "trg_slf_attn_pre_softmax_shape_delta", +# "trg_slf_attn_post_softmax_shape_delta", ) + +#from optim import LearningRateScheduler +class LearningRateScheduler(object): + """ + Wrapper for learning rate scheduling as described in the Transformer paper. + LearningRateScheduler adapts the learning rate externally and the adapted + learning rate will be feeded into the main_program as input data. + """ + + def __init__(self, + d_model, + warmup_steps, + learning_rate=0.001, + current_steps=0, + name="learning_rate"): + self.current_steps = current_steps + self.warmup_steps = warmup_steps + self.d_model = d_model + self.static_lr = learning_rate + self.learning_rate = layers.create_global_var( + name=name, + shape=[1], + value=float(learning_rate), + dtype="float32", + persistable=True) + + def update_learning_rate(self): + self.current_steps += 1 + lr_value = np.power(self.d_model, -0.5) * np.min([ + np.power(self.current_steps, -0.5), + np.power(self.warmup_steps, -1.5) * self.current_steps + ]) * self.static_lr + return np.array([lr_value], dtype="float32") -def prepare_batch_input(insts, src_pad_idx, trg_pad_idx, n_head): + +#from transformer_train import train_loop +def pad_batch_data(insts, + pad_idx, + n_head, + is_target=False, + is_label=False, + return_attn_bias=True, + return_max_len=True, + return_num_token=False): """ Pad the instances to the max sequence length in batch, and generate the - corresponding position data and attention bias. Then, convert the numpy - data to tensors and return a dict mapping names to tensors. + corresponding position data and attention bias. """ + return_list = [] + max_len = max(len(inst) for inst in insts) + num_token = reduce(lambda x, y: x + y, + [len(inst) for inst in insts]) if return_num_token else 0 + # Any token included in dict can be used to pad, since the paddings' loss + # will be masked out by weights and make no effect on parameter gradients. + inst_data = np.array( + [inst + [pad_idx] * (max_len - len(inst)) for inst in insts]) + return_list += [inst_data.astype("int64").reshape([-1, 1])] + if is_label: # label weight + inst_weight = np.array( + [[1.] * len(inst) + [0.] * (max_len - len(inst)) for inst in insts]) + return_list += [inst_weight.astype("float32").reshape([-1, 1])] + else: # position data + inst_pos = np.array([ + range(1, len(inst) + 1) + [0] * (max_len - len(inst)) + for inst in insts + ]) + return_list += [inst_pos.astype("int64").reshape([-1, 1])] + if return_attn_bias: + if is_target: + # This is used to avoid attention on paddings and subsequent + # words. + slf_attn_bias_data = np.ones((inst_data.shape[0], max_len, max_len)) + slf_attn_bias_data = np.triu(slf_attn_bias_data, + 1).reshape([-1, 1, max_len, max_len]) + slf_attn_bias_data = np.tile(slf_attn_bias_data, + [1, n_head, 1, 1]) * [-1e9] + else: + # This is used to avoid attention on paddings. + slf_attn_bias_data = np.array([[0] * len(inst) + [-1e9] * + (max_len - len(inst)) + for inst in insts]) + slf_attn_bias_data = np.tile( + slf_attn_bias_data.reshape([-1, 1, 1, max_len]), + [1, n_head, max_len, 1]) + return_list += [slf_attn_bias_data.astype("float32")] + if return_max_len: + return_list += [max_len] + if return_num_token: + return_list += [num_token] + return return_list if len(return_list) > 1 else return_list[0] + + +def prepare_batch_input(insts, data_input_names, src_pad_idx, trg_pad_idx, + n_head, d_model): + """ + Put all padded data needed by training into a dict. + """ + src_word, src_pos, src_slf_attn_bias, src_max_len = pad_batch_data( + [inst[0] for inst in insts], src_pad_idx, n_head, is_target=False) + src_word = src_word.reshape(-1, src_max_len, 1) + src_pos = src_pos.reshape(-1, src_max_len, 1) + trg_word, trg_pos, trg_slf_attn_bias, trg_max_len = pad_batch_data( + [inst[1] for inst in insts], trg_pad_idx, n_head, is_target=True) + trg_word = trg_word.reshape(-1, trg_max_len, 1) + trg_pos = trg_pos.reshape(-1, trg_max_len, 1) - def __pad_batch_data(insts, - pad_idx, - is_target=False, - return_pos=True, - return_attn_bias=True, - return_max_len=True): - """ - Pad the instances to the max sequence length in batch, and generate the - corresponding position data and attention bias. - """ - return_list = [] - max_len = max(len(inst) for inst in insts) - inst_data = np.array( - [inst + [pad_idx] * (max_len - len(inst)) for inst in insts]) - return_list += [inst_data.astype("int64").reshape([-1, 1])] - if return_pos: - inst_pos = np.array([[ - pos_i + 1 if w_i != pad_idx else 0 - for pos_i, w_i in enumerate(inst) - ] for inst in inst_data]) - - return_list += [inst_pos.astype("int64").reshape([-1, 1])] - if return_attn_bias: - if is_target: - # This is used to avoid attention on paddings and subsequent - # words. - slf_attn_bias_data = np.ones((inst_data.shape[0], max_len, - max_len)) - slf_attn_bias_data = np.triu(slf_attn_bias_data, 1).reshape( - [-1, 1, max_len, max_len]) - slf_attn_bias_data = np.tile(slf_attn_bias_data, - [1, n_head, 1, 1]) * [-1e9] - else: - # This is used to avoid attention on paddings. - slf_attn_bias_data = np.array([[0] * len(inst) + [-1e9] * - (max_len - len(inst)) - for inst in insts]) - slf_attn_bias_data = np.tile( - slf_attn_bias_data.reshape([-1, 1, 1, max_len]), - [1, n_head, max_len, 1]) - return_list += [slf_attn_bias_data.astype("float32")] - if return_max_len: - return_list += [max_len] - return return_list if len(return_list) > 1 else return_list[0] - - src_word, src_pos, src_slf_attn_bias, src_max_len = __pad_batch_data( - [inst[0] for inst in insts], src_pad_idx, is_target=False) - trg_word, trg_pos, trg_slf_attn_bias, trg_max_len = __pad_batch_data( - [inst[1] for inst in insts], trg_pad_idx, is_target=True) trg_src_attn_bias = np.tile(src_slf_attn_bias[:, :, ::src_max_len, :], [1, 1, trg_max_len, 1]).astype("float32") - lbl_word = __pad_batch_data([inst[2] for inst in insts], trg_pad_idx, False, - False, False, False) - lbl_weight = (lbl_word != trg_pad_idx).astype("float32").reshape([-1, 1]) + lbl_word, lbl_weight, num_token = pad_batch_data( + [inst[2] for inst in insts], + trg_pad_idx, + n_head, + is_target=False, + is_label=True, + return_attn_bias=False, + return_max_len=False, + return_num_token=True) + + data_input_dict = dict( + zip(data_input_names, [ + src_word, src_pos, src_slf_attn_bias, trg_word, trg_pos, + trg_slf_attn_bias, trg_src_attn_bias, lbl_word, lbl_weight + ])) + return data_input_dict, np.asarray([num_token], dtype="float32") + + +def read_multiple(reader, count, clip_last=True): + """ + Stack data from reader for multi-devices. + """ + + def __impl__(): + res = [] + for item in reader(): + res.append(item) + if len(res) == count: + yield res + res = [] + if len(res) == count: + yield res + elif not clip_last: + data = [] + for item in res: + data += item + if len(data) > count: + inst_num_per_part = len(data) // count + yield [ + data[inst_num_per_part * i:inst_num_per_part * (i + 1)] + for i in range(count) + ] + + return __impl__ + + +def split_data(data, num_part): + """ + Split data for each device. + """ + if len(data) == num_part: + return data + data = data[0] + inst_num_per_part = len(data) // num_part return [ - src_word, src_pos, trg_word, trg_pos, src_slf_attn_bias, - trg_slf_attn_bias, trg_src_attn_bias, lbl_word, lbl_weight + data[inst_num_per_part * i:inst_num_per_part * (i + 1)] + for i in range(num_part) ] -def transformer(use_feed): - assert not use_feed, "transfomer doesn't support feed yet" - return transformer_model.transformer( - ModelHyperParams.src_vocab_size + 1, - ModelHyperParams.trg_vocab_size + 1, ModelHyperParams.max_length + 1, - ModelHyperParams.n_layer, ModelHyperParams.n_head, - ModelHyperParams.d_key, ModelHyperParams.d_value, - ModelHyperParams.d_model, ModelHyperParams.d_inner_hid, - ModelHyperParams.dropout, ModelHyperParams.src_pad_idx, - ModelHyperParams.trg_pad_idx, ModelHyperParams.pos_pad_idx) - - -def get_model(): - avg_cost = transformer(use_feed=False) - optimizer = fluid.optimizer.Adam() - optimizer.minimize(avg_cost) - fluid.memory_optimize(fluid.default_main_program()) - return avg_cost - - -def get_transpiler(trainer_id, main_program, pserver_endpoints, trainers): - t = fluid.DistributeTranspiler() - t.transpile( - trainer_id=trainer_id, - program=main_program, - pservers=pserver_endpoints, - trainers=trainers) - return t - - -class DistTransformer2x2(object): - def run_pserver(self, pserver_endpoints, trainers, current_endpoint, - trainer_id): - get_model() - t = get_transpiler(trainer_id, - fluid.default_main_program(), pserver_endpoints, - trainers) - pserver_prog = t.get_pserver_program(current_endpoint) - startup_prog = t.get_startup_program(current_endpoint, pserver_prog) +def test_context(train_progm, avg_cost, train_exe, dev_count, data_input_names, + sum_cost, token_num): + # Context to do validation. + test_program = train_progm.clone() + with fluid.program_guard(test_program): + test_program = fluid.io.get_inference_program([avg_cost]) + + val_data = DataReader( + src_vocab_fpath=TrainTaskConfig.src_vocab_fpath, + trg_vocab_fpath=TrainTaskConfig.trg_vocab_fpath, + fpattern=TrainTaskConfig.val_file_pattern, + token_delimiter=TrainTaskConfig.token_delimiter, + use_token_batch=TrainTaskConfig.use_token_batch, + batch_size=TrainTaskConfig.batch_size * + (1 if TrainTaskConfig.use_token_batch else dev_count), + pool_size=TrainTaskConfig.pool_size, + sort_type=TrainTaskConfig.sort_type, + start_mark=TrainTaskConfig.special_token[0], + end_mark=TrainTaskConfig.special_token[1], + unk_mark=TrainTaskConfig.special_token[2], + # count start and end tokens out + max_length=ModelHyperParams.max_length - 2, + clip_last_batch=False, + shuffle=False, + shuffle_batch=False) + + build_strategy = fluid.BuildStrategy() + + strategy = fluid.ExecutionStrategy() + strategy.num_threads = 1 + + test_exe = fluid.ParallelExecutor( + use_cuda=TrainTaskConfig.use_gpu, + main_program=test_program, + share_vars_from=train_exe, + build_strategy=build_strategy, + exec_strategy=strategy) + + def test(exe=test_exe): + test_total_cost = 0 + test_total_token = 0 + test_data = read_multiple( + reader=val_data.batch_generator, + count=dev_count if TrainTaskConfig.use_token_batch else 1) + for batch_id, data in enumerate(test_data()): + feed_list = [] + for place_id, data_buffer in enumerate( + split_data( + data, num_part=dev_count)): + data_input_dict, _ = prepare_batch_input( + data_buffer, data_input_names, ModelHyperParams.eos_idx, + ModelHyperParams.eos_idx, ModelHyperParams.n_head, + ModelHyperParams.d_model) + feed_list.append(data_input_dict) + + outs = exe.run(feed=feed_list, + fetch_list=[sum_cost.name, token_num.name]) + sum_cost_val, token_num_val = np.array(outs[0]), np.array(outs[1]) + test_total_cost += sum_cost_val.sum() + test_total_token += token_num_val.sum() + test_avg_cost = test_total_cost / test_total_token + test_ppl = np.exp([min(test_avg_cost, 100)]) + return test_avg_cost, test_ppl + + return test + + +def train_loop(exe, train_progm, dev_count, sum_cost, avg_cost, lr_scheduler, + token_num, predict): + # Initialize the parameters. + if TrainTaskConfig.ckpt_path: + lr_scheduler.current_steps = TrainTaskConfig.start_step + else: + exe.run(fluid.framework.default_startup_program()) + + train_data = DataReader( + src_vocab_fpath=TrainTaskConfig.src_vocab_fpath, + trg_vocab_fpath=TrainTaskConfig.trg_vocab_fpath, + fpattern=TrainTaskConfig.train_file_pattern, + token_delimiter=TrainTaskConfig.token_delimiter, + use_token_batch=TrainTaskConfig.use_token_batch, + batch_size=TrainTaskConfig.batch_size * + (1 if TrainTaskConfig.use_token_batch else dev_count), + pool_size=TrainTaskConfig.pool_size, + sort_type=TrainTaskConfig.sort_type, + shuffle=TrainTaskConfig.shuffle, + shuffle_batch=TrainTaskConfig.shuffle_batch, + start_mark=TrainTaskConfig.special_token[0], + end_mark=TrainTaskConfig.special_token[1], + unk_mark=TrainTaskConfig.special_token[2], + # count start and end tokens out + max_length=ModelHyperParams.max_length - 2, + clip_last_batch=False) + train_data = read_multiple( + reader=train_data.batch_generator, + count=dev_count if TrainTaskConfig.use_token_batch else 1) + + build_strategy = fluid.BuildStrategy() + # Since the token number differs among devices, customize gradient scale to + # use token average cost among multi-devices. and the gradient scale is + # `1 / token_number` for average cost. + build_strategy.gradient_scale_strategy = fluid.BuildStrategy.GradientScaleStrategy.Customized + + strategy = fluid.ExecutionStrategy() + strategy.num_threads = 1 + + train_exe = fluid.ParallelExecutor( + use_cuda=TrainTaskConfig.use_gpu, + loss_name=sum_cost.name, + main_program=train_progm, + build_strategy=build_strategy, + exec_strategy=strategy) + + data_input_names = encoder_data_input_fields + decoder_data_input_fields[: + -1] + label_data_input_fields + + if TrainTaskConfig.val_file_pattern is not None: + test = test_context(train_progm, avg_cost, train_exe, dev_count, + data_input_names, sum_cost, token_num) + + # the best cross-entropy value with label smoothing + loss_normalizer = -((1. - TrainTaskConfig.label_smooth_eps) * np.log( + (1. - TrainTaskConfig.label_smooth_eps + )) + TrainTaskConfig.label_smooth_eps * + np.log(TrainTaskConfig.label_smooth_eps / ( + ModelHyperParams.trg_vocab_size - 1) + 1e-20)) + init = False + for pass_id in xrange(TrainTaskConfig.pass_num): + pass_start_time = time.time() + for batch_id, data in enumerate(train_data()): + if batch_id >= 5: + break + + feed_list = [] + total_num_token = 0 + + #if TrainTaskConfig.local: + # lr_rate = lr_scheduler.update_learning_rate() + #for place_id, data_buffer in enumerate( + # split_data( + # data, num_part=dev_count)): + + if TrainTaskConfig.local: + lr_rate = lr_scheduler.update_learning_rate() + + for place_id, data_buffer in enumerate( + split_data( + data, num_part=dev_count)): + data_input_dict, num_token = prepare_batch_input( + data_buffer, data_input_names, ModelHyperParams.eos_idx, + ModelHyperParams.eos_idx, ModelHyperParams.n_head, + ModelHyperParams.d_model) + total_num_token += num_token + feed_kv_pairs = data_input_dict.items() + if TrainTaskConfig.local: + feed_kv_pairs += { + lr_scheduler.learning_rate.name: lr_rate + }.items() + feed_list.append(dict(feed_kv_pairs)) + + if not init: + for pos_enc_param_name in pos_enc_param_names: + pos_enc = position_encoding_init( + ModelHyperParams.max_length + 1, + ModelHyperParams.d_model) + feed_list[place_id][pos_enc_param_name] = pos_enc + + if not TrainTaskConfig.check_acc: + for feed_dict in feed_list: + feed_dict[sum_cost.name + "@GRAD"] = 1. / total_num_token + else: + b = 100 * TrainTaskConfig.batch_size + a = np.asarray([b], dtype="float32") + for feed_dict in feed_list: + feed_dict[sum_cost.name + "@GRAD"] = 1. / a + + outs = train_exe.run(fetch_list=[sum_cost.name, token_num.name], + feed=feed_list) + + sum_cost_val, token_num_val = np.array(outs[0]), np.array(outs[1]) + total_sum_cost = sum_cost_val.sum() + total_token_num = token_num_val.sum() + total_avg_cost = total_sum_cost / total_token_num + + init = True + + # Validate and save the model for inference. + if TrainTaskConfig.val_file_pattern is not None: + val_avg_cost, val_ppl = test() + print("[%f]" % val_avg_cost) + else: + assert (False) + + +#import transformer_reader as reader +class SortType(object): + GLOBAL = 'global' + POOL = 'pool' + NONE = "none" + + +class Converter(object): + def __init__(self, vocab, beg, end, unk, delimiter): + self._vocab = vocab + self._beg = beg + self._end = end + self._unk = unk + self._delimiter = delimiter + + def __call__(self, sentence): + return [self._beg] + [ + self._vocab.get(w, self._unk) + for w in sentence.split(self._delimiter) + ] + [self._end] + + +class ComposedConverter(object): + def __init__(self, converters): + self._converters = converters + + def __call__(self, parallel_sentence): + return [ + self._converters[i](parallel_sentence[i]) + for i in range(len(self._converters)) + ] + + +class SentenceBatchCreator(object): + def __init__(self, batch_size): + self.batch = [] + self._batch_size = batch_size + + def append(self, info): + self.batch.append(info) + if len(self.batch) == self._batch_size: + tmp = self.batch + self.batch = [] + return tmp + + +class TokenBatchCreator(object): + def __init__(self, batch_size): + self.batch = [] + self.max_len = -1 + self._batch_size = batch_size + + def append(self, info): + cur_len = info.max_len + max_len = max(self.max_len, cur_len) + if max_len * (len(self.batch) + 1) > self._batch_size: + result = self.batch + self.batch = [info] + self.max_len = cur_len + return result + else: + self.max_len = max_len + self.batch.append(info) + + +class SampleInfo(object): + def __init__(self, i, max_len, min_len): + self.i = i + self.min_len = min_len + self.max_len = max_len + + +class MinMaxFilter(object): + def __init__(self, max_len, min_len, underlying_creator): + self._min_len = min_len + self._max_len = max_len + self._creator = underlying_creator + + def append(self, info): + if info.max_len > self._max_len or info.min_len < self._min_len: + return + else: + return self._creator.append(info) + + @property + def batch(self): + return self._creator.batch + + +class DataReader(object): + """ + The data reader loads all data from files and produces batches of data + in the way corresponding to settings. + + An example of returning a generator producing data batches whose data + is shuffled in each pass and sorted in each pool: + + ``` + train_data = DataReader( + src_vocab_fpath='data/src_vocab_file', + trg_vocab_fpath='data/trg_vocab_file', + fpattern='data/part-*', + use_token_batch=True, + batch_size=2000, + pool_size=10000, + sort_type=SortType.POOL, + shuffle=True, + shuffle_batch=True, + start_mark='', + end_mark='', + unk_mark='', + clip_last_batch=False).batch_generator + ``` + + :param src_vocab_fpath: The path of vocabulary file of source language. + :type src_vocab_fpath: basestring + :param trg_vocab_fpath: The path of vocabulary file of target language. + :type trg_vocab_fpath: basestring + :param fpattern: The pattern to match data files. + :type fpattern: basestring + :param batch_size: The number of sequences contained in a mini-batch. + or the maximum number of tokens (include paddings) contained in a + mini-batch. + :type batch_size: int + :param pool_size: The size of pool buffer. + :type pool_size: int + :param sort_type: The grain to sort by length: 'global' for all + instances; 'pool' for instances in pool; 'none' for no sort. + :type sort_type: basestring + :param clip_last_batch: Whether to clip the last uncompleted batch. + :type clip_last_batch: bool + :param tar_fname: The data file in tar if fpattern matches a tar file. + :type tar_fname: basestring + :param min_length: The minimum length used to filt sequences. + :type min_length: int + :param max_length: The maximum length used to filt sequences. + :type max_length: int + :param shuffle: Whether to shuffle all instances. + :type shuffle: bool + :param shuffle_batch: Whether to shuffle the generated batches. + :type shuffle_batch: bool + :param use_token_batch: Whether to produce batch data according to + token number. + :type use_token_batch: bool + :param field_delimiter: The delimiter used to split source and target in + each line of data file. + :type field_delimiter: basestring + :param token_delimiter: The delimiter used to split tokens in source or + target sentences. + :type token_delimiter: basestring + :param start_mark: The token representing for the beginning of + sentences in dictionary. + :type start_mark: basestring + :param end_mark: The token representing for the end of sentences + in dictionary. + :type end_mark: basestring + :param unk_mark: The token representing for unknown word in dictionary. + :type unk_mark: basestring + :param seed: The seed for random. + :type seed: int + """ + + def __init__(self, + src_vocab_fpath, + trg_vocab_fpath, + fpattern, + batch_size, + pool_size, + sort_type=SortType.GLOBAL, + clip_last_batch=True, + tar_fname=None, + min_length=0, + max_length=100, + shuffle=True, + shuffle_batch=False, + use_token_batch=False, + field_delimiter="\t", + token_delimiter=" ", + start_mark="", + end_mark="", + unk_mark="", + seed=0): + self._src_vocab = self.load_dict(src_vocab_fpath) + self._only_src = True + if trg_vocab_fpath is not None: + self._trg_vocab = self.load_dict(trg_vocab_fpath) + self._only_src = False + self._pool_size = pool_size + self._batch_size = batch_size + self._use_token_batch = use_token_batch + self._sort_type = sort_type + self._clip_last_batch = clip_last_batch + self._shuffle = shuffle + self._shuffle_batch = shuffle_batch + self._min_length = min_length + self._max_length = max_length + self._field_delimiter = field_delimiter + self._token_delimiter = token_delimiter + self.load_src_trg_ids(end_mark, fpattern, start_mark, tar_fname, + unk_mark) + self._random = random.Random(x=seed) + + def load_src_trg_ids(self, end_mark, fpattern, start_mark, tar_fname, + unk_mark): + converters = [ + Converter( + vocab=self._src_vocab, + beg=self._src_vocab[start_mark], + end=self._src_vocab[end_mark], + unk=self._src_vocab[unk_mark], + delimiter=self._token_delimiter) + ] + if not self._only_src: + converters.append( + Converter( + vocab=self._trg_vocab, + beg=self._trg_vocab[start_mark], + end=self._trg_vocab[end_mark], + unk=self._trg_vocab[unk_mark], + delimiter=self._token_delimiter)) + + converters = ComposedConverter(converters) + + self._src_seq_ids = [] + self._trg_seq_ids = None if self._only_src else [] + self._sample_infos = [] + + for i, line in enumerate(self._load_lines(fpattern, tar_fname)): + src_trg_ids = converters(line) + self._src_seq_ids.append(src_trg_ids[0]) + lens = [len(src_trg_ids[0])] + if not self._only_src: + self._trg_seq_ids.append(src_trg_ids[1]) + lens.append(len(src_trg_ids[1])) + self._sample_infos.append(SampleInfo(i, max(lens), min(lens))) + + def _load_lines(self, fpattern, tar_fname): + fpaths = glob.glob(fpattern) + + if len(fpaths) == 1 and tarfile.is_tarfile(fpaths[0]): + if tar_fname is None: + raise Exception("If tar file provided, please set tar_fname.") + + f = tarfile.open(fpaths[0], "r") + for line in f.extractfile(tar_fname): + fields = line.strip("\n").split(self._field_delimiter) + if (not self._only_src and len(fields) == 2) or ( + self._only_src and len(fields) == 1): + yield fields + else: + for fpath in fpaths: + if not os.path.isfile(fpath): + raise IOError("Invalid file: %s" % fpath) + + with open(fpath, "r") as f: + for line in f: + fields = line.strip("\n").split(self._field_delimiter) + if (not self._only_src and len(fields) == 2) or ( + self._only_src and len(fields) == 1): + yield fields + + @staticmethod + def load_dict(dict_path, reverse=False): + word_dict = {} + with open(dict_path, "r") as fdict: + for idx, line in enumerate(fdict): + if reverse: + word_dict[idx] = line.strip("\n") + else: + word_dict[line.strip("\n")] = idx + return word_dict + + def batch_generator(self): + # global sort or global shuffle + if self._sort_type == SortType.GLOBAL: + infos = sorted( + self._sample_infos, key=lambda x: x.max_len, reverse=True) + else: + if self._shuffle: + infos = self._sample_infos + self._random.shuffle(infos) + else: + infos = self._sample_infos + + if self._sort_type == SortType.POOL: + for i in range(0, len(infos), self._pool_size): + infos[i:i + self._pool_size] = sorted( + infos[i:i + self._pool_size], key=lambda x: x.max_len) + + # concat batch + batches = [] + batch_creator = TokenBatchCreator( + self._batch_size + ) if self._use_token_batch else SentenceBatchCreator(self._batch_size) + batch_creator = MinMaxFilter(self._max_length, self._min_length, + batch_creator) + + for info in infos: + batch = batch_creator.append(info) + if batch is not None: + batches.append(batch) + + if not self._clip_last_batch and len(batch_creator.batch) != 0: + batches.append(batch_creator.batch) + + if self._shuffle_batch: + self._random.shuffle(batches) + + for batch in batches: + batch_ids = [info.i for info in batch] + + if self._only_src: + yield [[self._src_seq_ids[idx]] for idx in batch_ids] + else: + yield [(self._src_seq_ids[idx], self._trg_seq_ids[idx][:-1], + self._trg_seq_ids[idx][1:]) for idx in batch_ids] + + +#from transformer_model import transformer +def position_encoding_init(n_position, d_pos_vec): + """ + Generate the initial values for the sinusoid position encoding table. + """ + position_enc = np.array([[ + pos / np.power(10000, 2 * (j // 2) / d_pos_vec) + for j in range(d_pos_vec) + ] if pos != 0 else np.zeros(d_pos_vec) for pos in range(n_position)]) + position_enc[1:, 0::2] = np.sin(position_enc[1:, 0::2]) # dim 2i + position_enc[1:, 1::2] = np.cos(position_enc[1:, 1::2]) # dim 2i+1 + return position_enc.astype("float32") + + +def multi_head_attention(queries, + keys, + values, + attn_bias, + d_key, + d_value, + d_model, + n_head=1, + dropout_rate=0., + cache=None): + """ + Multi-Head Attention. Note that attn_bias is added to the logit before + computing softmax activiation to mask certain selected positions so that + they will not considered in attention weights. + """ + if not (len(queries.shape) == len(keys.shape) == len(values.shape) == 3): + raise ValueError( + "Inputs: quries, keys and values should all be 3-D tensors.") + + def __compute_qkv(queries, keys, values, n_head, d_key, d_value): + """ + Add linear projection to queries, keys, and values. + """ + q = layers.fc(input=queries, + size=d_key * n_head, + num_flatten_dims=2, + param_attr=const_para_attr, + bias_attr=const_bias_attr) + k = layers.fc(input=keys, + size=d_key * n_head, + num_flatten_dims=2, + param_attr=const_para_attr, + bias_attr=const_bias_attr) + v = layers.fc(input=values, + size=d_value * n_head, + num_flatten_dims=2, + param_attr=const_para_attr, + bias_attr=const_bias_attr) + return q, k, v + + def __split_heads(x, n_head): + """ + Reshape the last dimension of inpunt tensor x so that it becomes two + dimensions and then transpose. Specifically, input a tensor with shape + [bs, max_sequence_length, n_head * hidden_dim] then output a tensor + with shape [bs, n_head, max_sequence_length, hidden_dim]. + """ + if n_head == 1: + return x + + hidden_size = x.shape[-1] + # The value 0 in shape attr means copying the corresponding dimension + # size of the input as the output dimension size. + reshaped = layers.reshape( + x=x, shape=[0, 0, n_head, hidden_size // n_head]) + + # permuate the dimensions into: + # [batch_size, n_head, max_sequence_len, hidden_size_per_head] + return layers.transpose(x=reshaped, perm=[0, 2, 1, 3]) + + def __combine_heads(x): + """ + Transpose and then reshape the last two dimensions of inpunt tensor x + so that it becomes one dimension, which is reverse to __split_heads. + """ + if len(x.shape) == 3: return x + if len(x.shape) != 4: + raise ValueError("Input(x) should be a 4-D Tensor.") + + trans_x = layers.transpose(x, perm=[0, 2, 1, 3]) + # The value 0 in shape attr means copying the corresponding dimension + # size of the input as the output dimension size. + return layers.reshape( + x=trans_x, + shape=map(int, [0, 0, trans_x.shape[2] * trans_x.shape[3]])) + + def scaled_dot_product_attention(q, k, v, attn_bias, d_model, dropout_rate): + """ + Scaled Dot-Product Attention + """ + scaled_q = layers.scale(x=q, scale=d_model**-0.5) + product = layers.matmul(x=scaled_q, y=k, transpose_y=True) + if attn_bias: + product += attn_bias + weights = layers.softmax(product) + if dropout_rate: + weights = layers.dropout( + weights, + dropout_prob=dropout_rate, + seed=ModelHyperParams.dropout_seed, + is_test=False) + out = layers.matmul(weights, v) + return out + + q, k, v = __compute_qkv(queries, keys, values, n_head, d_key, d_value) + + if cache is not None: # use cache and concat time steps + k = cache["k"] = layers.concat([cache["k"], k], axis=1) + v = cache["v"] = layers.concat([cache["v"], v], axis=1) + + q = __split_heads(q, n_head) + k = __split_heads(k, n_head) + v = __split_heads(v, n_head) + + ctx_multiheads = scaled_dot_product_attention(q, k, v, attn_bias, d_model, + dropout_rate) + + out = __combine_heads(ctx_multiheads) + + # Project back to the model size. + proj_out = layers.fc(input=out, + size=d_model, + num_flatten_dims=2, + param_attr=const_para_attr, + bias_attr=const_bias_attr) + return proj_out + + +def positionwise_feed_forward(x, d_inner_hid, d_hid): + """ + Position-wise Feed-Forward Networks. + This module consists of two linear transformations with a ReLU activation + in between, which is applied to each position separately and identically. + """ + hidden = layers.fc(input=x, + size=d_inner_hid, + num_flatten_dims=2, + act="relu", + param_attr=const_para_attr, + bias_attr=const_bias_attr) + out = layers.fc(input=hidden, + size=d_hid, + num_flatten_dims=2, + param_attr=const_para_attr, + bias_attr=const_bias_attr) + return out + + +def pre_post_process_layer(prev_out, out, process_cmd, dropout_rate=0.): + """ + Add residual connection, layer normalization and droput to the out tensor + optionally according to the value of process_cmd. + This will be used before or after multi-head attention and position-wise + feed-forward networks. + """ + for cmd in process_cmd: + if cmd == "a": # add residual connection + out = out + prev_out if prev_out else out + elif cmd == "n": # add layer normalization + out = layers.layer_norm( + out, + begin_norm_axis=len(out.shape) - 1, + param_attr=fluid.initializer.Constant(1.), + bias_attr=fluid.initializer.Constant(0.)) + elif cmd == "d": # add dropout + if dropout_rate: + out = layers.dropout( + out, + dropout_prob=dropout_rate, + seed=ModelHyperParams.dropout_seed, + is_test=False) + return out + + +pre_process_layer = partial(pre_post_process_layer, None) +post_process_layer = pre_post_process_layer + + +def prepare_encoder(src_word, + src_pos, + src_vocab_size, + src_emb_dim, + src_max_len, + dropout_rate=0., + word_emb_param_name=None, + pos_enc_param_name=None): + """Add word embeddings and position encodings. + The output tensor has a shape of: + [batch_size, max_src_length_in_batch, d_model]. + This module is used at the bottom of the encoder stacks. + """ + if TrainTaskConfig.check_acc: + src_word_emb = layers.embedding( + src_word, + size=[src_vocab_size, src_emb_dim], + param_attr=fluid.ParamAttr( + name=word_emb_param_name, + initializer=fluid.initializer.ConstantInitializer(0.001))) + else: + src_word_emb = layers.embedding( + src_word, + size=[src_vocab_size, src_emb_dim], + param_attr=fluid.ParamAttr( + name=word_emb_param_name, + initializer=fluid.initializer.Normal(0., src_emb_dim**-0.5))) + + src_word_emb = layers.scale(x=src_word_emb, scale=src_emb_dim**0.5) + src_pos_enc = layers.embedding( + src_pos, + size=[src_max_len, src_emb_dim], + param_attr=fluid.ParamAttr( + name=pos_enc_param_name, + trainable=False, + initializer=fluid.initializer.ConstantInitializer(0.001))) + enc_input = src_word_emb + src_pos_enc + return layers.dropout( + enc_input, + dropout_prob=dropout_rate, + seed=ModelHyperParams.dropout_seed, + is_test=False) if dropout_rate else enc_input + + +prepare_encoder = partial( + prepare_encoder, pos_enc_param_name=pos_enc_param_names[0]) +prepare_decoder = partial( + prepare_encoder, pos_enc_param_name=pos_enc_param_names[1]) + + +def encoder_layer(enc_input, + attn_bias, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate=0.): + """The encoder layers that can be stacked to form a deep encoder. + This module consits of a multi-head (self) attention followed by + position-wise feed-forward networks and both the two components companied + with the post_process_layer to add residual connection, layer normalization + and droput. + """ + attn_output = multi_head_attention(enc_input, enc_input, enc_input, + attn_bias, d_key, d_value, d_model, + n_head, dropout_rate) + attn_output = post_process_layer(enc_input, attn_output, "dan", + dropout_rate) + ffd_output = positionwise_feed_forward(attn_output, d_inner_hid, d_model) + return post_process_layer(attn_output, ffd_output, "dan", dropout_rate) + + +def encoder(enc_input, + attn_bias, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate=0.): + """ + The encoder is composed of a stack of identical layers returned by calling + encoder_layer. + """ + for i in range(n_layer): + enc_output = encoder_layer(enc_input, attn_bias, n_head, d_key, d_value, + d_model, d_inner_hid, dropout_rate) + enc_input = enc_output + return enc_output + + +def decoder_layer(dec_input, + enc_output, + slf_attn_bias, + dec_enc_attn_bias, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate=0., + cache=None): + """ The layer to be stacked in decoder part. + The structure of this module is similar to that in the encoder part except + a multi-head attention is added to implement encoder-decoder attention. + """ + slf_attn_output = multi_head_attention( + dec_input, + dec_input, + dec_input, + slf_attn_bias, + d_key, + d_value, + d_model, + n_head, + dropout_rate, + cache, ) + slf_attn_output = post_process_layer( + dec_input, + slf_attn_output, + "dan", # residual connection + dropout + layer normalization + dropout_rate, ) + enc_attn_output = multi_head_attention( + slf_attn_output, + enc_output, + enc_output, + dec_enc_attn_bias, + d_key, + d_value, + d_model, + n_head, + dropout_rate, ) + enc_attn_output = post_process_layer( + slf_attn_output, + enc_attn_output, + "dan", # residual connection + dropout + layer normalization + dropout_rate, ) + ffd_output = positionwise_feed_forward( + enc_attn_output, + d_inner_hid, + d_model, ) + dec_output = post_process_layer( + enc_attn_output, + ffd_output, + "dan", # residual connection + dropout + layer normalization + dropout_rate, ) + return dec_output + + +def decoder(dec_input, + enc_output, + dec_slf_attn_bias, + dec_enc_attn_bias, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate=0., + caches=None): + """ + The decoder is composed of a stack of identical decoder_layer layers. + """ + for i in range(n_layer): + cache = None + if caches is not None: + cache = caches[i] + + dec_output = decoder_layer( + dec_input, + enc_output, + dec_slf_attn_bias, + dec_enc_attn_bias, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, + cache=cache) + dec_input = dec_output + return dec_output + + +def make_all_inputs(input_fields): + """ + Define the input data layers for the transformer model. + """ + inputs = [] + for input_field in input_fields: + input_var = layers.data( + name=input_field, + shape=input_descs[input_field][0], + dtype=input_descs[input_field][1], + lod_level=input_descs[input_field][2] + if len(input_descs[input_field]) == 3 else 0, + append_batch_size=False) + inputs.append(input_var) + return inputs + + +def transformer( + src_vocab_size, + trg_vocab_size, + max_length, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, + weight_sharing, + label_smooth_eps, ): + if weight_sharing: + assert src_vocab_size == src_vocab_size, ( + "Vocabularies in source and target should be same for weight sharing." + ) + enc_inputs = make_all_inputs(encoder_data_input_fields) + + enc_output = wrap_encoder( + src_vocab_size, + max_length, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, + weight_sharing, + enc_inputs, ) + + dec_inputs = make_all_inputs(decoder_data_input_fields[:-1]) + + predict = wrap_decoder( + trg_vocab_size, + max_length, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, + weight_sharing, + dec_inputs, + enc_output, ) + + # Padding index do not contribute to the total loss. The weights is used to + # cancel padding index in calculating the loss. + label, weights = make_all_inputs(label_data_input_fields) + if label_smooth_eps: + label = layers.label_smooth( + label=layers.one_hot( + input=label, depth=trg_vocab_size), + epsilon=label_smooth_eps) + + cost = layers.softmax_with_cross_entropy( + logits=layers.reshape( + predict, shape=[-1, trg_vocab_size]), + label=label, + soft_label=True if label_smooth_eps else False) + weighted_cost = cost * weights + sum_cost = layers.reduce_sum(weighted_cost) + token_num = layers.reduce_sum(weights) + avg_cost = sum_cost / token_num + avg_cost.stop_gradient = True + return sum_cost, avg_cost, predict, token_num + + +def wrap_encoder(src_vocab_size, + max_length, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, + weight_sharing, + enc_inputs=None): + """ + The wrapper assembles together all needed layers for the encoder. + """ + if enc_inputs is None: + # This is used to implement independent encoder program in inference. + src_word, src_pos, src_slf_attn_bias = \ + make_all_inputs(encoder_data_input_fields) + else: + src_word, src_pos, src_slf_attn_bias = \ + enc_inputs + enc_input = prepare_encoder( + src_word, + src_pos, + src_vocab_size, + d_model, + max_length, + dropout_rate, + word_emb_param_name=word_emb_param_names[0]) + enc_output = encoder(enc_input, src_slf_attn_bias, n_layer, n_head, d_key, + d_value, d_model, d_inner_hid, dropout_rate) + return enc_output + + +def wrap_decoder(trg_vocab_size, + max_length, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, + weight_sharing, + dec_inputs=None, + enc_output=None, + caches=None): + """ + The wrapper assembles together all needed layers for the decoder. + """ + if dec_inputs is None: + # This is used to implement independent decoder program in inference. + trg_word, trg_pos, trg_slf_attn_bias, trg_src_attn_bias, \ + enc_output = make_all_inputs( + decoder_data_input_fields + decoder_util_input_fields) + else: + trg_word, trg_pos, trg_slf_attn_bias, trg_src_attn_bias = dec_inputs + + dec_input = prepare_decoder( + trg_word, + trg_pos, + trg_vocab_size, + d_model, + max_length, + dropout_rate, + word_emb_param_name=word_emb_param_names[0] + if weight_sharing else word_emb_param_names[1]) + dec_output = decoder( + dec_input, + enc_output, + trg_slf_attn_bias, + trg_src_attn_bias, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, + caches=caches) + # Return logits for training and probs for inference. + if weight_sharing: + predict = layers.matmul( + x=dec_output, + y=fluid.get_var(word_emb_param_names[0]), + transpose_y=True) + else: + predict = layers.fc(input=dec_output, + size=trg_vocab_size, + num_flatten_dims=2, + param_attr=const_para_attr, + bias_attr=const_bias_attr) + if dec_inputs is None: + predict = layers.softmax(predict) + return predict + + +def fast_decode( + src_vocab_size, + trg_vocab_size, + max_in_len, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, + weight_sharing, + beam_size, + max_out_len, + eos_idx, ): + """ + Use beam search to decode. Caches will be used to store states of history + steps which can make the decoding faster. + """ + enc_output = wrap_encoder(src_vocab_size, max_in_len, n_layer, n_head, + d_key, d_value, d_model, d_inner_hid, + dropout_rate, weight_sharing) + start_tokens, init_scores, trg_src_attn_bias = \ + make_all_inputs(fast_decoder_data_input_fields ) + + def beam_search(): + max_len = layers.fill_constant( + shape=[1], dtype=start_tokens.dtype, value=max_out_len) + step_idx = layers.fill_constant( + shape=[1], dtype=start_tokens.dtype, value=0) + cond = layers.less_than(x=step_idx, y=max_len) + while_op = layers.While(cond) + # array states will be stored for each step. + ids = layers.array_write( + layers.reshape(start_tokens, (-1, 1)), step_idx) + scores = layers.array_write(init_scores, step_idx) + # cell states will be overwrited at each step. + # caches contains states of history steps to reduce redundant + # computation in decoder. + caches = [{ + "k": layers.fill_constant_batch_size_like( + input=start_tokens, + shape=[-1, 0, d_model], + dtype=enc_output.dtype, + value=0), + "v": layers.fill_constant_batch_size_like( + input=start_tokens, + shape=[-1, 0, d_model], + dtype=enc_output.dtype, + value=0) + } for i in range(n_layer)] + with while_op.block(): + pre_ids = layers.array_read(array=ids, i=step_idx) + pre_ids = layers.reshape(pre_ids, (-1, 1, 1)) + pre_scores = layers.array_read(array=scores, i=step_idx) + # sequence_expand can gather sequences according to lod thus can be + # used in beam search to sift states corresponding to selected ids. + pre_src_attn_bias = layers.sequence_expand( + x=trg_src_attn_bias, y=pre_scores) + pre_enc_output = layers.sequence_expand(x=enc_output, y=pre_scores) + pre_caches = [{ + "k": layers.sequence_expand( + x=cache["k"], y=pre_scores), + "v": layers.sequence_expand( + x=cache["v"], y=pre_scores), + } for cache in caches] + pre_pos = layers.elementwise_mul( + x=layers.fill_constant_batch_size_like( + input=pre_enc_output, # cann't use pre_ids here since it has lod + value=1, + shape=[-1, 1, 1], + dtype=pre_ids.dtype), + y=layers.increment( + x=step_idx, value=1.0, in_place=False), + axis=0) + logits = wrap_decoder( + trg_vocab_size, + max_in_len, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, + weight_sharing, + dec_inputs=(pre_ids, pre_pos, None, pre_src_attn_bias), + enc_output=pre_enc_output, + caches=pre_caches) + logits = layers.reshape(logits, (-1, trg_vocab_size)) + + topk_scores, topk_indices = layers.topk( + input=layers.softmax(logits), k=beam_size) + accu_scores = layers.elementwise_add( + x=layers.log(topk_scores), + y=layers.reshape( + pre_scores, shape=[-1]), + axis=0) + # beam_search op uses lod to distinguish branches. + topk_indices = layers.lod_reset(topk_indices, pre_ids) + selected_ids, selected_scores = layers.beam_search( + pre_ids=pre_ids, + pre_scores=pre_scores, + ids=topk_indices, + scores=accu_scores, + beam_size=beam_size, + end_id=eos_idx) + + layers.increment(x=step_idx, value=1.0, in_place=True) + # update states + layers.array_write(selected_ids, i=step_idx, array=ids) + layers.array_write(selected_scores, i=step_idx, array=scores) + layers.assign(pre_src_attn_bias, trg_src_attn_bias) + layers.assign(pre_enc_output, enc_output) + for i in range(n_layer): + layers.assign(pre_caches[i]["k"], caches[i]["k"]) + layers.assign(pre_caches[i]["v"], caches[i]["v"]) + length_cond = layers.less_than(x=step_idx, y=max_len) + finish_cond = layers.logical_not(layers.is_empty(x=selected_ids)) + layers.logical_and(x=length_cond, y=finish_cond, out=cond) + + finished_ids, finished_scores = layers.beam_search_decode( + ids, scores, beam_size=beam_size, end_id=eos_idx) + return finished_ids, finished_scores + + finished_ids, finished_scores = beam_search() + return finished_ids, finished_scores + + +def get_model(is_dist, is_async): + sum_cost, avg_cost, predict, token_num = transformer( + ModelHyperParams.src_vocab_size, ModelHyperParams.trg_vocab_size, + ModelHyperParams.max_length + 1, ModelHyperParams.n_layer, + ModelHyperParams.n_head, ModelHyperParams.d_key, + ModelHyperParams.d_value, ModelHyperParams.d_model, + ModelHyperParams.d_inner_hid, ModelHyperParams.dropout, + ModelHyperParams.weight_sharing, TrainTaskConfig.label_smooth_eps) + + local_lr_scheduler = LearningRateScheduler(ModelHyperParams.d_model, + TrainTaskConfig.warmup_steps, + TrainTaskConfig.learning_rate) + + if not is_dist: + optimizer = fluid.optimizer.Adam( + learning_rate=local_lr_scheduler.learning_rate, + beta1=TrainTaskConfig.beta1, + beta2=TrainTaskConfig.beta2, + epsilon=TrainTaskConfig.eps) + optimizer.minimize(sum_cost) + elif is_async: + optimizer = fluid.optimizer.SGD(0.003) + optimizer.minimize(sum_cost) + else: + lr_decay = fluid.layers\ + .learning_rate_scheduler\ + .noam_decay(ModelHyperParams.d_model, + TrainTaskConfig.warmup_steps) + + optimizer = fluid.optimizer.Adam( + learning_rate=lr_decay, + beta1=TrainTaskConfig.beta1, + beta2=TrainTaskConfig.beta2, + epsilon=TrainTaskConfig.eps) + optimizer.minimize(sum_cost) + + return sum_cost, avg_cost, predict, token_num, local_lr_scheduler + + +def update_args(): + src_dict = DataReader.load_dict(TrainTaskConfig.src_vocab_fpath) + trg_dict = DataReader.load_dict(TrainTaskConfig.trg_vocab_fpath) + dict_args = [ + "src_vocab_size", str(len(src_dict)), "trg_vocab_size", + str(len(trg_dict)), "bos_idx", + str(src_dict[TrainTaskConfig.special_token[0]]), "eos_idx", + str(src_dict[TrainTaskConfig.special_token[1]]), "unk_idx", + str(src_dict[TrainTaskConfig.special_token[2]]) + ] + merge_cfg_from_list(dict_args, [TrainTaskConfig, ModelHyperParams]) + + +class DistTransformer2x2(TestDistRunnerBase): + def run_pserver(self, args): + get_model(True, not args.sync_mode) + t = self.get_transpiler(args.trainer_id, + fluid.default_main_program(), args.endpoints, + args.trainers, args.sync_mode) + pserver_prog = t.get_pserver_program(args.current_endpoint) + startup_prog = t.get_startup_program(args.current_endpoint, + pserver_prog) place = fluid.CPUPlace() exe = fluid.Executor(place) exe.run(startup_prog) exe.run(pserver_prog) - def _wait_ps_ready(self, pid): - retry_times = 20 - while True: - assert retry_times >= 0, "wait ps ready failed" - time.sleep(3) - print("waiting ps ready: ", pid) - try: - # the listen_and_serv_op would touch a file which contains the listen port - # on the /tmp directory until it was ready to process all the RPC call. - os.stat("/tmp/paddle.%d.port" % pid) - return - except os.error: - retry_times -= 1 - - def run_trainer(self, place, endpoints, trainer_id, trainers, is_dist=True): - avg_cost = get_model() - if is_dist: - t = get_transpiler(trainer_id, - fluid.default_main_program(), endpoints, - trainers) + def run_trainer(self, place, args): + + sum_cost, avg_cost, predict, token_num, local_lr_scheduler = get_model( + args.is_dist, not args.sync_mode) + + if args.is_dist: + t = self.get_transpiler(args.trainer_id, + fluid.default_main_program(), + args.endpoints, args.trainers, + args.sync_mode) trainer_prog = t.get_trainer_program() + TrainTaskConfig.batch_size = 10 + TrainTaskConfig.train_file_pattern = TrainTaskConfig.data_path + "train.tok.clean.bpe.32000.en-de.train_{}".format( + args.trainer_id) else: + TrainTaskConfig.batch_size = 20 trainer_prog = fluid.default_main_program() startup_exe = fluid.Executor(place) - startup_exe.run(fluid.default_startup_program()) - - strategy = fluid.ExecutionStrategy() - strategy.num_threads = 1 - strategy.allow_op_delay = False - exe = fluid.ParallelExecutor( - True, loss_name=avg_cost.name, exec_strategy=strategy) - - first_loss, = exe.run(fetch_list=[avg_cost.name]) - print(first_loss) - for i in six.moves.xrange(5): - _ = exe.run(fetch_list=[avg_cost.name]) - last_loss, = exe.run(fetch_list=[avg_cost.name]) - print(last_loss) - - -def main(role="pserver", - endpoints="127.0.0.1:9123", - trainer_id=0, - current_endpoint="127.0.0.1:9123", - trainers=1, - is_dist=True): - - reader = paddle.batch( - wmt16.train(ModelHyperParams.src_vocab_size, - ModelHyperParams.trg_vocab_size), - batch_size=transformer_model.batch_size) - - with fluid.recordio_writer.create_recordio_writer( - WMT16_RECORDIO_FILE) as writer: - for batch in reader(): - for tensor in prepare_batch_input( - batch, ModelHyperParams.src_pad_idx, - ModelHyperParams.trg_pad_idx, ModelHyperParams.n_head): - t = fluid.LoDTensor() - t.set(tensor, fluid.CPUPlace()) - writer.append_tensor(t) - writer.complete_append_tensor() - - model = DistTransformer2x2() - if role == "pserver": - model.run_pserver(endpoints, trainers, current_endpoint, trainer_id) - else: - p = fluid.CUDAPlace(0) if core.is_compiled_with_cuda( - ) else fluid.CPUPlace() - model.run_trainer(p, endpoints, trainer_id, trainers, is_dist) + + TrainTaskConfig.local = not args.is_dist + + train_loop(startup_exe, trainer_prog, 1, sum_cost, avg_cost, + local_lr_scheduler, token_num, predict) if __name__ == "__main__": - if len(sys.argv) != 7: - print( - "Usage: python dist_transformer.py [pserver/trainer] [endpoints] [trainer_id] [current_endpoint] [trainers] [is_dist]" - ) - role = sys.argv[1] - endpoints = sys.argv[2] - trainer_id = int(sys.argv[3]) - current_endpoint = sys.argv[4] - trainers = int(sys.argv[5]) - is_dist = True if sys.argv[6] == "TRUE" else False - main( - role=role, - endpoints=endpoints, - trainer_id=trainer_id, - current_endpoint=current_endpoint, - trainers=trainers, - is_dist=is_dist) + update_args() + runtime_main(DistTransformer2x2) diff --git a/python/paddle/fluid/tests/unittests/dist_word2vec.py b/python/paddle/fluid/tests/unittests/dist_word2vec.py index 0ad994a258c04cabc807823b7d2a8ae8bb62ab2c..f3e740fc7027a4a562b836c3113b87d55062c185 100644 --- a/python/paddle/fluid/tests/unittests/dist_word2vec.py +++ b/python/paddle/fluid/tests/unittests/dist_word2vec.py @@ -49,28 +49,32 @@ class TestDistWord2vec2x2(TestDistRunnerBase): dtype='float32', is_sparse=IS_SPARSE, param_attr=fluid.ParamAttr( - name='shared_w', initializer=fluid.initializer.Constant())) + name='shared_w', + initializer=fluid.initializer.Constant(value=0.1))) embed_second = fluid.layers.embedding( input=words[1], size=[dict_size, EMBED_SIZE], dtype='float32', is_sparse=IS_SPARSE, param_attr=fluid.ParamAttr( - name='shared_w', initializer=fluid.initializer.Constant())) + name='shared_w', + initializer=fluid.initializer.Constant(value=0.1))) embed_third = fluid.layers.embedding( input=words[2], size=[dict_size, EMBED_SIZE], dtype='float32', is_sparse=IS_SPARSE, param_attr=fluid.ParamAttr( - name='shared_w', initializer=fluid.initializer.Constant())) + name='shared_w', + initializer=fluid.initializer.Constant(value=0.1))) embed_forth = fluid.layers.embedding( input=words[3], size=[dict_size, EMBED_SIZE], dtype='float32', is_sparse=IS_SPARSE, param_attr=fluid.ParamAttr( - name='shared_w', initializer=fluid.initializer.Constant())) + name='shared_w', + initializer=fluid.initializer.Constant(value=0.1))) concat_embed = fluid.layers.concat( input=[embed_first, embed_second, embed_third, embed_forth], @@ -80,13 +84,13 @@ class TestDistWord2vec2x2(TestDistRunnerBase): size=HIDDEN_SIZE, act='sigmoid', param_attr=fluid.ParamAttr( - initializer=fluid.initializer.Constant())) + initializer=fluid.initializer.Constant(value=0.1))) predict_word = fluid.layers.fc( input=hidden1, size=dict_size, act='softmax', param_attr=fluid.ParamAttr( - initializer=fluid.initializer.Constant())) + initializer=fluid.initializer.Constant(value=0.1))) cost = fluid.layers.cross_entropy( input=predict_word, label=words[4]) avg_cost = fluid.layers.mean(cost) diff --git a/python/paddle/fluid/tests/unittests/test_dist_base.py b/python/paddle/fluid/tests/unittests/test_dist_base.py index 4c71181d0d736bd1a8796b2d38ed1667557e3db8..f730a17477d2c1d1edc7ee55becc632d96c0e260 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_base.py +++ b/python/paddle/fluid/tests/unittests/test_dist_base.py @@ -21,7 +21,7 @@ import sys import six import signal import subprocess -import six +import argparse class TestDistRunnerBase(object): @@ -30,7 +30,7 @@ class TestDistRunnerBase(object): "get_model should be implemented by child classes.") def get_transpiler(self, trainer_id, main_program, pserver_endpoints, - trainers): + trainers, sync_mode): # NOTE: import fluid until runtime, or else forking processes will cause error. import paddle import paddle.fluid as fluid @@ -39,33 +39,35 @@ class TestDistRunnerBase(object): trainer_id=trainer_id, program=main_program, pservers=pserver_endpoints, - trainers=trainers) + trainers=trainers, + sync_mode=sync_mode) return t - def run_pserver(self, pserver_endpoints, trainers, current_endpoint, - trainer_id): + def run_pserver(self, args): import paddle import paddle.fluid as fluid self.get_model(batch_size=2) - t = self.get_transpiler(trainer_id, - fluid.default_main_program(), pserver_endpoints, - trainers) - pserver_prog = t.get_pserver_program(current_endpoint) - startup_prog = t.get_startup_program(current_endpoint, pserver_prog) + t = self.get_transpiler(args.trainer_id, + fluid.default_main_program(), args.endpoints, + args.trainers, args.sync_mode) + pserver_prog = t.get_pserver_program(args.current_endpoint) + startup_prog = t.get_startup_program(args.current_endpoint, + pserver_prog) place = fluid.CPUPlace() exe = fluid.Executor(place) exe.run(startup_prog) exe.run(pserver_prog) - def run_trainer(self, place, endpoints, trainer_id, trainers, is_dist=True): + def run_trainer(self, place, args): import paddle import paddle.fluid as fluid test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \ self.get_model(batch_size=2) - if is_dist: - t = self.get_transpiler(trainer_id, - fluid.default_main_program(), endpoints, - trainers) + if args.is_dist: + t = self.get_transpiler(args.trainer_id, + fluid.default_main_program(), + args.endpoints, args.trainers, + args.sync_mode) trainer_prog = t.get_trainer_program() else: trainer_prog = fluid.default_main_program() @@ -76,8 +78,18 @@ class TestDistRunnerBase(object): strategy = fluid.ExecutionStrategy() strategy.num_threads = 1 strategy.allow_op_delay = False + build_stra = fluid.BuildStrategy() + + if args.use_reduce: + build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce + else: + build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce + exe = fluid.ParallelExecutor( - True, loss_name=avg_cost.name, exec_strategy=strategy) + True, + loss_name=avg_cost.name, + exec_strategy=strategy, + build_strategy=build_stra) feed_var_list = [ var for var in trainer_prog.global_block().vars.values() @@ -106,45 +118,64 @@ def runtime_main(test_class): import paddle.fluid as fluid import paddle.fluid.core as core - if len(sys.argv) != 7: - print( - "Usage: python dist_se_resnext.py [pserver/trainer] [endpoints] [trainer_id] [current_endpoint] [trainers] [is_dist]" - ) - role = sys.argv[1] - endpoints = sys.argv[2] - trainer_id = int(sys.argv[3]) - current_endpoint = sys.argv[4] - trainers = int(sys.argv[5]) - is_dist = True if sys.argv[6] == "TRUE" else False + parser = argparse.ArgumentParser(description='Run dist test.') + parser.add_argument( + '--role', type=str, required=True, choices=['pserver', 'trainer']) + parser.add_argument('--endpoints', type=str, required=False, default="") + parser.add_argument('--is_dist', action='store_true') + parser.add_argument('--trainer_id', type=int, required=False, default=0) + parser.add_argument('--trainers', type=int, required=False, default=1) + parser.add_argument( + '--current_endpoint', type=str, required=False, default="") + parser.add_argument('--sync_mode', action='store_true') + parser.add_argument('--mem_opt', action='store_true') + parser.add_argument('--use_reduce', action='store_true') + + args = parser.parse_args() model = test_class() - if role == "pserver": - model.run_pserver(endpoints, trainers, current_endpoint, trainer_id) + if args.role == "pserver": + model.run_pserver(args) else: p = fluid.CUDAPlace(0) if core.is_compiled_with_cuda( ) else fluid.CPUPlace() - model.run_trainer(p, endpoints, trainer_id, trainers, is_dist) + model.run_trainer(p, args) import paddle.compat as cpt class TestDistBase(unittest.TestCase): + def _setup_config(self): + raise NotImplementedError("tests should have _setup_config implemented") + def setUp(self): self._trainers = 2 self._pservers = 2 self._ps_endpoints = "127.0.0.1:9123,127.0.0.1:9124" self._python_interp = "python" + self._sync_mode = True + self._mem_opt = False + self._use_reduce = False + self._setup_config() def start_pserver(self, model_file, check_error_log): ps0_ep, ps1_ep = self._ps_endpoints.split(",") - ps0_cmd = "%s %s pserver %s 0 %s %d TRUE" % \ + ps_cmd = "%s %s --role pserver --endpoints %s --trainer_id 0 --current_endpoint %s --trainers %d --is_dist" + ps0_cmd = ps_cmd % \ (self._python_interp, model_file, self._ps_endpoints, ps0_ep, self._trainers) - ps1_cmd = "%s %s pserver %s 0 %s %d TRUE" % \ + ps1_cmd = ps_cmd % \ (self._python_interp, model_file, self._ps_endpoints, ps1_ep, self._trainers) + if self._sync_mode: + ps0_cmd += " --sync_mode" + ps1_cmd += " --sync_mode" + if self._mem_opt: + ps0_cmd += " --mem_opt" + ps1_cmd += " --mem_opt" + ps0_pipe = subprocess.PIPE ps1_pipe = subprocess.PIPE if check_error_log: @@ -195,9 +226,7 @@ class TestDistBase(unittest.TestCase): # Run local to get a base line env_local = {"CUDA_VISIBLE_DEVICES": "0"} env_local.update(required_envs) - local_cmd = "%s %s trainer %s 0 %s %d FLASE" % \ - (self._python_interp, model_file, - "127.0.0.1:1234", "127.0.0.1:1234", 1) + local_cmd = "%s %s --role trainer" % (self._python_interp, model_file) if not check_error_log: local_proc = subprocess.Popen( local_cmd.split(" "), @@ -226,12 +255,23 @@ class TestDistBase(unittest.TestCase): self._wait_ps_ready(ps1.pid) ps0_ep, ps1_ep = self._ps_endpoints.split(",") - tr0_cmd = "%s %s trainer %s 0 %s %d TRUE" % \ - (self._python_interp, model_file, self._ps_endpoints, ps0_ep, - self._trainers) - tr1_cmd = "%s %s trainer %s 1 %s %d TRUE" % \ - (self._python_interp, model_file, self._ps_endpoints, ps1_ep, - self._trainers) + tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --trainers %d --is_dist" + tr0_cmd = tr_cmd % \ + (self._python_interp, model_file, self._ps_endpoints, + 0, ps0_ep, self._trainers) + tr1_cmd = tr_cmd % \ + (self._python_interp, model_file, self._ps_endpoints, + 1, ps1_ep, self._trainers) + + if self._sync_mode: + tr0_cmd += " --sync_mode" + tr1_cmd += " --sync_mode" + if self._mem_opt: + tr0_cmd += " --mem_opt" + tr1_cmd += " --mem_opt" + if self._use_reduce: + tr0_cmd += " --use_reduce" + tr1_cmd += " --use_reduce" env0 = {"CUDA_VISIBLE_DEVICES": "0"} env1 = {"CUDA_VISIBLE_DEVICES": "1"} @@ -282,6 +322,10 @@ class TestDistBase(unittest.TestCase): # FIXME: use terminate() instead of sigkill. os.kill(ps0.pid, signal.SIGKILL) os.kill(ps1.pid, signal.SIGKILL) + ps0.terminate() + ps1.terminate() + ps0.wait() + ps1.wait() FNULL.close() self.assertAlmostEqual(local_first_loss, dist_first_loss, delta=delta) diff --git a/python/paddle/fluid/tests/unittests/test_dist_mnist.py b/python/paddle/fluid/tests/unittests/test_dist_mnist.py index 4ec68d411b0f0e9ae89b107914e8fd844a19228b..59a137c18c9435ef5c5772d0cc08f197c1d86603 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_mnist.py +++ b/python/paddle/fluid/tests/unittests/test_dist_mnist.py @@ -17,10 +17,51 @@ import unittest from test_dist_base import TestDistBase -class TestDistSeResneXt2x2(TestDistBase): +class TestDistMnist2x2(TestDistBase): + def _setup_config(self): + self._sync_mode = True + self._use_reduce = False + + def test_se_resnext(self): + self.check_with_place("dist_mnist.py", delta=1e-7) + + +class TestDistMnist2x2WithMemopt(TestDistBase): + def _setup_config(self): + self._sync_mode = True + self._mem_opt = True + def test_se_resnext(self): self.check_with_place("dist_mnist.py", delta=1e-7) +class TestDistMnistAsync(TestDistBase): + def _setup_config(self): + self._sync_mode = False + self._use_reduce = False + + def test_se_resnext(self): + self.check_with_place("dist_mnist.py", delta=200) + + +# FIXME(typhoonzero): enable these tests once we have 4 +# 4 GPUs on CI machine, and the base class should be updated. +# +# class TestDistMnist2x2ReduceMode(TestDistBase): +# def _setup_config(self): +# self._sync_mode = True +# self._use_reduce = True + +# def test_se_resnext(self): +# self.check_with_place("dist_mnist.py", delta=1e-7) + +# class TestDistMnistAsyncReduceMode(TestDistBase): +# def _setup_config(self): +# self._sync_mode = False +# self._use_reduce = True + +# def test_se_resnext(self): +# self.check_with_place("dist_mnist.py", delta=200) + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py b/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py index 16525f6fdb60a90a44a628fb0648f4130218c102..c0e9fa38e7d1eadd89eff9a8ba4442f888b8120e 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py +++ b/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py @@ -18,9 +18,20 @@ from test_dist_base import TestDistBase class TestDistSeResneXt2x2(TestDistBase): + def _setup_config(self): + self._sync_mode = True + def test_se_resnext(self): self.check_with_place("dist_se_resnext.py", delta=1e-7) +class TestDistSeResneXt2x2Async(TestDistBase): + def _setup_config(self): + self._sync_mode = False + + def test_se_resnext(self): + self.check_with_place("dist_se_resnext.py", delta=100) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_train.py b/python/paddle/fluid/tests/unittests/test_dist_train.py index 9581abdf394d738470d32ae609838832077ee519..083525ccf54d389b60c4aaa9f8c6223f07c773cd 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_train.py +++ b/python/paddle/fluid/tests/unittests/test_dist_train.py @@ -100,7 +100,7 @@ class TestSendOp(unittest.TestCase): main.global_block().append_op( type="fetch_barrier", inputs={}, - outputs={}, + outputs={"Out": []}, attrs={ "endpoints": ["127.0.0.1:{0}".format(port)], RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE diff --git a/python/paddle/fluid/tests/unittests/test_dist_transformer.py b/python/paddle/fluid/tests/unittests/test_dist_transformer.py index 313207ff9ce054f81322224cb6ceafaaf25bbedf..a8e6ce4cfe18384e405f1602429628914d2c2e00 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transformer.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transformer.py @@ -15,14 +15,55 @@ from __future__ import print_function import unittest +import paddle from test_dist_base import TestDistBase -class TestDistTransformer2x2(TestDistBase): +def download_files(): + url_prefix = 'http://paddle-unittest-data.cdn.bcebos.com/dist_transformer/' + vocab_url = url_prefix + 'vocab.bpe.32000' + vocab_md5 = 'a86d345ca6e27f6591d0dccb1b9be853' + paddle.dataset.common.download(vocab_url, 'test_dist_transformer', + vocab_md5) + + local_train_url = url_prefix + 'train.tok.clean.bpe.32000.en-de' + local_train_md5 = '033eb02b9449e6dd823f050782ac8914' + paddle.dataset.common.download(local_train_url, 'test_dist_transformer', + local_train_md5) + + train0_url = url_prefix + 'train.tok.clean.bpe.32000.en-de.train_0' + train0_md5 = 'ddce7f602f352a0405267285379a38b1' + paddle.dataset.common.download(train0_url, 'test_dist_transformer', + train0_md5) + + train1_url = url_prefix + 'train.tok.clean.bpe.32000.en-de.train_1' + train1_md5 = '8757798200180285b1a619cd7f408747' + paddle.dataset.common.download(train1_url, 'test_dist_transformer', + train1_md5) + + test_url = url_prefix + 'newstest2013.tok.bpe.32000.en-de' + test_md5 = '9dd74a266dbdb25314183899f269b4a2' + paddle.dataset.common.download(test_url, 'test_dist_transformer', test_md5) + + +class TestDistTransformer2x2Sync(TestDistBase): + def _setup_config(self): + self._sync_mode = True + + def test_transformer(self): + download_files() + #Note: loss on test dataset of the first 5 batch are: + # 10.518872, 10.518871, 10.518868, 10.518862, 10.518855 + self.check_with_place("dist_transformer.py", delta=1e-7) + + +class TestDistTransformer2x2Async(TestDistBase): + def _setup_config(self): + self._sync_mode = False + def test_transformer(self): - # TODO(paddle-dev): check if the delta is OK. - # Usually start around ~8000 and converge to ~5000 - self.check_with_place("dist_transformer.py", delta=400) + download_files() + self.check_with_place("dist_transformer.py", delta=1.0) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py index abac533658fc20deda4a37448966282423e938e0..9be3ba405896c8f0f0edca9e379cd24a335a4f93 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py @@ -441,7 +441,7 @@ class TestLocalLookupTable(TestDistLookupTableBase): # 2 optimize for table adam # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num self.assertEqual([op.type for op in pserver1.blocks[2].ops], - ["sum", "adam", "scale", "scale"]) + ["sum", "scale", "adam", "scale", "scale"]) trainer, _ = self.get_trainer() self.assertEqual(len(trainer.blocks), 1) diff --git a/python/paddle/fluid/tests/unittests/test_dist_word2vec.py b/python/paddle/fluid/tests/unittests/test_dist_word2vec.py index e43992c488d35d1b3f670e13650d420b0498eeec..9a3e92e8d775a37e0c24ee1bcc5435628d61bb91 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_word2vec.py +++ b/python/paddle/fluid/tests/unittests/test_dist_word2vec.py @@ -18,8 +18,19 @@ from test_dist_base import TestDistBase class TestDistSeResneXt2x2(TestDistBase): + def _setup_config(self): + self._sync_mode = True + + def test_se_resnext(self): + self.check_with_place("dist_word2vec.py", delta=1e-4) + + +class TestDistSeResneXt2x2Async(TestDistBase): + def _setup_config(self): + self._sync_mode = False + def test_se_resnext(self): - self.check_with_place("dist_word2vec.py", delta=1e-7) + self.check_with_place("dist_word2vec.py", delta=1) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_scale_op.py b/python/paddle/fluid/tests/unittests/test_scale_op.py index 0a8a43253d79ba21c7333dd19af05d8adf410289..032af6ed5ce9e1007d6775306ef4c0aefb9dcc41 100644 --- a/python/paddle/fluid/tests/unittests/test_scale_op.py +++ b/python/paddle/fluid/tests/unittests/test_scale_op.py @@ -17,6 +17,8 @@ from __future__ import print_function import unittest import numpy as np from op_test import OpTest +import paddle.fluid.core as core +from paddle.fluid.op import Operator class TestScaleOp(OpTest): @@ -33,5 +35,57 @@ class TestScaleOp(OpTest): self.check_grad(['X'], 'Out') +class TestScaleOpSelectedRows(unittest.TestCase): + def check_with_place(self, place, in_name, out_name): + scope = core.Scope() + + # create and initialize Grad Variable + in_height = 10 + in_rows = [0, 4, 7] + in_row_numel = 12 + scale = 2.0 + + in_selected_rows = scope.var(in_name).get_selected_rows() + in_selected_rows.set_height(in_height) + in_selected_rows.set_rows(in_rows) + in_array = np.random.random( + (len(in_rows), in_row_numel)).astype("float32") + + in_tensor = in_selected_rows.get_tensor() + in_tensor.set(in_array, place) + + # create and initialize Param Variable + out_selected_rows = scope.var(out_name).get_selected_rows() + out_tensor = out_selected_rows.get_tensor() + out_tensor._set_dims(in_tensor._get_dims()) + + # create and run sgd operator + scale_op = Operator("scale", X=in_name, Out=out_name, scale=scale) + scale_op.run(scope, place) + + # get and compare result + out_height = out_selected_rows.height() + out_rows = out_selected_rows.rows() + result_array = np.array(out_tensor) + + assert (in_array * scale == result_array).all() + assert in_height == out_height + assert in_rows == out_rows + + def test_scale_selected_rows(self): + places = [core.CPUPlace()] + if core.is_compiled_with_cuda(): + places.append(core.CUDAPlace(0)) + for place in places: + self.check_with_place(place, 'in', 'out') + + def test_scale_selected_rows_inplace(self): + places = [core.CPUPlace()] + if core.is_compiled_with_cuda(): + places.append(core.CUDAPlace(0)) + for place in places: + self.check_with_place(place, 'in', 'in') + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_sequence_pad_op.py b/python/paddle/fluid/tests/unittests/test_sequence_pad_op.py new file mode 100644 index 0000000000000000000000000000000000000000..471515c817541976a06eb024fa3d4f77b78f920d --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_sequence_pad_op.py @@ -0,0 +1,131 @@ +# Copyright (c) 2018 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 unittest +import numpy as np +from op_test import OpTest + + +class TestSequencePadOp(OpTest): + def set_attr(self): + self.x_shape = [12, 4] + self.x_len_lod = [[2, 3, 4, 3]] + self.pad_value = [1.0] + self.padded_length = -1 + self.dtype = 'float32' + + def set_data(self): + x_data = np.random.uniform(0.1, 0.5, self.x_shape).astype(self.dtype) + pad_value_data = np.array(self.pad_value).astype(self.dtype) + self.inputs = { + 'X': (x_data, self.x_len_lod), + 'PadValue': pad_value_data + } + self.attrs = {'padded_length': self.padded_length} + + def compute(self): + # get padded length + padded_length = self.padded_length + x_len_lod_0 = self.x_len_lod[0] + if padded_length == -1: + max_seq_len = 0 + for l in x_len_lod_0: + max_seq_len = max(max_seq_len, l) + padded_length = max_seq_len + + # do padding + x_data = self.inputs['X'][0] + pad_value_data = self.inputs['PadValue'] + if pad_value_data.shape == (1, ): + pad_value_data = np.broadcast_to( + pad_value_data, shape=x_data.shape[1:]) + padded_sequences = [] + start_idx = 0 + for l in x_len_lod_0: + end_idx = start_idx + l + seq = x_data[start_idx:end_idx] + to_pad_len = padded_length - l + for _ in range(to_pad_len): + seq = np.append(seq, pad_value_data[np.newaxis, :], axis=0) + padded_sequences.append(seq) + start_idx = end_idx + + out_data = np.array(padded_sequences) + self.outputs = {'Out': out_data} + + def setUp(self): + self.op_type = 'sequence_pad' + self.set_attr() + self.set_data() + self.compute() + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out") + + +class TestSequencePadOp2(TestSequencePadOp): + def set_attr(self): + self.x_shape = [12, 4] + self.x_len_lod = [[2, 3, 4, 3]] + self.pad_value = [1.0, 2.0, 3.0, 4.0] + self.padded_length = -1 + self.dtype = 'float32' + + +class TestSequencePadOp3(TestSequencePadOp): + def set_attr(self): + self.x_shape = [12, 4] + self.x_len_lod = [[2, 3, 4, 3]] + self.pad_value = [1.0] + self.padded_length = 7 + self.dtype = 'float32' + + +class TestSequencePadOp4(TestSequencePadOp): + def set_attr(self): + self.x_shape = [12, 4] + self.x_len_lod = [[2, 3, 4, 3]] + self.pad_value = [1.0, 2.0, 3.0, 4.0] + self.padded_length = 7 + self.dtype = 'float32' + + +class TestSequencePadOp5(TestSequencePadOp): + def set_attr(self): + self.x_shape = [12, 2, 2] + self.x_len_lod = [[2, 3, 4, 3]] + self.pad_value = [1.0] + self.padded_length = -1 + self.dtype = 'float32' + + +class TestSequencePadOp6(TestSequencePadOp): + def set_attr(self): + self.x_shape = [12, 2, 2] + self.x_len_lod = [[2, 3, 4, 3]] + self.pad_value = [[1.0, 2.0], [3.0, 4.0]] + self.padded_length = -1 + self.dtype = 'float32' + + +class TestSequencePadOp7(TestSequencePadOp): + def set_attr(self): + self.x_shape = [12, 2, 2] + self.x_len_lod = [[2, 3, 4, 3]] + self.pad_value = [1.0] + self.padded_length = 7 + self.dtype = 'float32' diff --git a/python/paddle/fluid/transpiler/details/program_utils.py b/python/paddle/fluid/transpiler/details/program_utils.py index 640dbf4bbed58edf746456419af18c75241fa03c..4261cccdc2e9baaac21b83c22e2090854378c67d 100644 --- a/python/paddle/fluid/transpiler/details/program_utils.py +++ b/python/paddle/fluid/transpiler/details/program_utils.py @@ -39,3 +39,136 @@ def find_op_by_output_arg(block, arg_name): if arg_name in op.output_arg_names: return index return -1 + + +def get_indent_space(indent, space_num=4): + ret = "" + for i in range(0, indent * space_num): + ret += " " + + return ret + + +def variable_to_code(var): + """ + Get readable codes of fluid variable. + Args: + var: A fluid operator. + Returns: + string: The formatted string. + """ + if var.type == core.VarDesc.VarType.SELECTED_ROWS or var.type == core.VarDesc.VarType.LOD_TENSOR: + var_str = "{name} : fluid.{type}.shape{shape}.astype({dtype})".\ + format(i="{", e="}", name=var.name, type=var.type, shape=var.shape, dtype=var.dtype) + else: + var_str = "{name} : fluid.{type})".\ + format(i="{", e="}", name=var.name, type=var.type) + + if type(var) == paddle.fluid.framework.Parameter: + if var.trainable: + var_str = "trainable parameter " + var_str + else: + var_str = "parameter " + var_str + else: + var_str = "var " + var_str + + if var.persistable: + var_str = "persist " + var_str + + return var_str + + +def op_to_code(op): + """ + Get readable codes of fluid operator. + Args: + op: A fluid operator. + Returns: + string: The foramtted string. + """ + + outputs_str = "{" + for i in range(0, len(op.output_names)): + outputs_str += "{name}=".format(name=op.output_names[i]) + o = op.output(op.output_names[i]) + outputs_str += "{value}".format(value=o) + if i != len(op.output_names) - 1: + outputs_str += ", " + outputs_str += "}" + + inputs_str = "{" + for i in range(0, len(op.input_names)): + inputs_str += "{name}=".format(name=op.input_names[i]) + o = op.input(op.input_names[i]) + inputs_str += "{value}".format(value=o) + + if i != len(op.input_names) - 1: + inputs_str += ", " + inputs_str += "}" + + attrs_str = "" + for i in range(0, len(op.attr_names)): + name = op.attr_names[i] + + attr_type = op.desc.attr_type(name) + if attr_type == core.AttrType.BLOCK: + a = "{name} = block[{value}]".format( + name=name, type=attr_type, value=op.block_attr_id(name)) + attrs_str += a + continue + + if attr_type == core.AttrType.BLOCKS: + a = "{name} = blocks{value}".format( + name=name, type=attr_type, value=op.blocks_attr_ids(name)) + attrs_str += a + continue + + a = "{name} = {value}".format( + name=name, type=attr_type, value=op.desc.attr(name)) + attrs_str += a + if i != len(op.attr_names) - 1: + attrs_str += ", " + + if outputs_str != "{}": + op_str = "{outputs} = {op_type}(inputs={inputs}, {attrs})".\ + format(outputs = outputs_str, op_type=op.type, inputs=inputs_str, attrs=attrs_str) + else: + op_str = "{op_type}(inputs={inputs}, {attrs})".\ + format(op_type=op.type, inputs=inputs_str, attrs=attrs_str) + return op_str + + +def block_to_code(block, block_idx): + indent = 0 + + print("{0}{1} // block {2}".format( + get_indent_space(indent), '{', block_idx)) + + indent += 1 + # sort all vars + all_vars = sorted(block.vars.iteritems(), key=lambda x: x[0]) + for var in all_vars: + print("{}{}".format(get_indent_space(indent), variable_to_code(var[1]))) + + if len(all_vars) > 0: + print("") + + for op in block.ops: + print("{}{}".format(get_indent_space(indent), op_to_code(op))) + indent -= 1 + + print("{0}{1}".format(get_indent_space(indent), '}')) + + +def program_to_code(prog): + """ + Print readable codes of fluid program. + Args: + prog : A fluid program. + An example result like bellow: + https://github.com/PaddlePaddle/Paddle/pull/12673 + """ + block_idx = 0 + for block in prog.blocks: + block_to_code(block, block_idx) + block_idx += 1 diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index a11b157c39f2c29a17db99e877969b87b03e170a..50ffe80d03f36a3a24b863a8ebcab1e66d98df7a 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -31,9 +31,10 @@ Steps to transpile pserver: """ import math -import random +import sys import numpy as np import collections +import random from .ps_dispatcher import RoundRobin, HashName, PSDispatcher from .. import core, framework @@ -181,7 +182,8 @@ class DistributeTranspiler(object): program=None, pservers="127.0.0.1:6174", trainers=1, - sync_mode=True): + sync_mode=True, + startup_program=None): """ Run the transpiler. @@ -194,13 +196,17 @@ class DistributeTranspiler(object): list. trainers (int): number of trainers in the distributed job. sync_mode (bool): Do sync training or not, default is True. + startup_program (Program|None): startup_program to transpile, + default is fluid.default_main_program(). """ if program is None: program = default_main_program() + if startup_program is None: + startup_program = default_startup_program() self.origin_program = program - self.origin_startup_program = default_startup_program().clone() + self.startup_program = startup_program + self.origin_startup_program = self.startup_program.clone() - self.startup_program = default_startup_program() self.trainer_num = trainers self.sync_mode = sync_mode self.trainer_id = trainer_id @@ -260,6 +266,10 @@ class DistributeTranspiler(object): name=framework.generate_control_dev_var_name()) grad_name_to_send_dummy_out[grad_varname] = dummy_output + # get send op_role_var, if not splited, the grad should have .trainer suffix + # if splited, grad should be the original grad var name (split_by_ref and send + # will be on the same place). ParallelExecutor + # will use op_role_var to get expected device place to run this op. program.global_block()._insert_op( index=index + 1, type="send", @@ -268,18 +278,23 @@ class DistributeTranspiler(object): attrs={ "epmap": eplist, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE, - OP_ROLE_VAR_ATTR_NAME: - [self.grad_name_to_param_name[grad_varname], grad_varname], + OP_ROLE_VAR_ATTR_NAME: [ + self.grad_name_to_param_name[grad_varname], + splited_grad_varname + ], "sync_mode": not self.sync_mode, }) for _, var in enumerate(splited_vars): send_vars.append(var) if self.sync_mode: + send_barrier_out = program.global_block().create_var( + name=framework.generate_control_dev_var_name()) + input_deps = grad_name_to_send_dummy_out.values() program.global_block().append_op( type="send_barrier", - inputs={}, - outputs={}, + inputs={"X": input_deps}, + outputs={"Out": send_barrier_out}, attrs={ "endpoints": pserver_endpoints, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE @@ -297,32 +312,46 @@ class DistributeTranspiler(object): self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i]) # step4: Concat the parameters splits together after recv. + all_recv_outputs = [] for param_varname, splited_var in six.iteritems(self.param_var_mapping): eps = [] for var in splited_var: index = [v.name for v in recv_vars].index(var.name) eps.append(eplist[index]) - grad_send_dummy_out = grad_name_to_send_dummy_out[ - self.param_name_to_grad_name[param_varname]] + if self.sync_mode: + recv_dep_in = send_barrier_out + else: + # connect deps to send op in async mode + recv_dep_in = grad_name_to_send_dummy_out[ + self.param_name_to_grad_name[param_varname]] + all_recv_outputs.extend(splited_var) + # get recv op_role_var, if not splited, the grad should have .trainer suffix + # if splited, grad should be the original grad var name. ParallelExecutor + # will use op_role_var to get expected device place to run this op. + orig_grad_name = self.param_name_to_grad_name[param_varname] + recv_op_role_var_name = orig_grad_name + splited_trainer_grad = self.grad_var_mapping[orig_grad_name] + if len(splited_trainer_grad) == 1: + recv_op_role_var_name = splited_trainer_grad[0].name + program.global_block().append_op( type="recv", - inputs={"X": [grad_send_dummy_out]}, + inputs={"X": [recv_dep_in]}, outputs={"Out": splited_var}, attrs={ "epmap": eps, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE, - OP_ROLE_VAR_ATTR_NAME: [ - param_varname, - self.param_name_to_grad_name[param_varname] - ], + OP_ROLE_VAR_ATTR_NAME: + [param_varname, recv_op_role_var_name], "sync_mode": not self.sync_mode }) if self.sync_mode: + # form a WAW dependency program.global_block().append_op( type="fetch_barrier", inputs={}, - outputs={}, + outputs={"Out": all_recv_outputs}, attrs={ "endpoints": pserver_endpoints, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE @@ -359,21 +388,18 @@ class DistributeTranspiler(object): return self.origin_program - def _get_trainer_startup_program(self, - recv_vars, - eplist, - startup_program=None): + def _get_trainer_startup_program(self, recv_vars, eplist): """ Get transpiled trainer side startup program. Args: - startup_program(Program): Startup program. + recv_vars (list): Variable list to recv for current trainer_id + eplist (list): A list of strings indicating Returns: Program: trainer side startup program. """ - if startup_program is None: - startup_program = self.startup_program + startup_program = self.startup_program # FIXME(gongwb): delete not need ops. # note that: some parameter is not trainable and those ops can't be deleted. @@ -406,10 +432,12 @@ class DistributeTranspiler(object): RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) + fetch_barrier_out = startup_program.global_block().create_var( + name=framework.generate_control_dev_var_name()) startup_program.global_block().append_op( type="fetch_barrier", inputs={}, - outputs={}, + outputs={"Out": fetch_barrier_out}, attrs={ "endpoints": self.pserver_endpoints, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE @@ -419,7 +447,18 @@ class DistributeTranspiler(object): #add concat ops to merge splited parameters received from parameter servers. if len(splited_var) <= 1: continue - orig_param = startup_program.global_block().vars[varname] + # NOTE: if enable memory optimization, origin vars maybe removed. + if startup_program.global_block().vars.has_key(varname): + orig_param = startup_program.global_block().vars[varname] + else: + origin_param_var = self.origin_program.global_block().vars[ + varname] + orig_param = startup_program.global_block().create_var( + name=varname, + persistable=origin_param_var.persistable, + type=origin_param_var.type, + dtype=origin_param_var.dtype, + shape=origin_param_var.shape) startup_program.global_block().append_op( type="concat", inputs={"X": splited_var}, @@ -442,7 +481,9 @@ class DistributeTranspiler(object): # NOTE: assume blocks of the same variable is not distributed # on the same pserver, only change param/grad varnames for # trainers to fetch. - + sys.stderr.write("get_pserver_program() is deprecated, call\ + get_pserver_programs() to get pserver main and startup\ + in a single call.") # step1 pserver_program = Program() pserver_program.random_seed = self.origin_program.random_seed @@ -626,32 +667,58 @@ class DistributeTranspiler(object): attrs=attrs) pserver_program._sync_with_cpp() + # save pserver program to generate pserver side startup relatively. + self.pserver_program = pserver_program return pserver_program + def get_pserver_programs(self, endpoint): + """ + Get pserver side main program and startup program for distributed training. + + Args: + endpoint (str): current pserver endpoint. + + Returns: + tuple: (main_program, startup_program), of type "Program" + """ + pserver_prog = self.get_pserver_program(endpoint) + pserver_startup = self.get_startup_program(endpoint) + return pserver_prog, pserver_startup + def get_startup_program(self, endpoint, - pserver_program, + pserver_program=None, startup_program=None): """ + **Deprecated** + Get startup program for current parameter server. Modify operator input variables if there are variables that were split to several blocks. Args: endpoint (str): current pserver endpoint. - pserver_program (Program): call get_pserver_program first and - pass the result here. - startup_program (Program): if pass None, will use - default_startup_program + pserver_program (Program): deprecated, call get_pserver_program first. + startup_program (Program): deprecated, should pass startup_program + when initalizing Returns: Program: parameter server side startup program. """ + sys.stderr.write("get_startup_program() is deprecated, call\ + get_pserver_programs() to get pserver main and startup\ + in a single call.") + if pserver_program != None: + sys.stderr.write("passing pserver_program to get_startup_program()\ + is deprecated, you can use new API get_pserver_programs() to\ + get both pserver main program and startup program.") + if startup_program != None: + sys.stderr.write("passing startup_program to get_startup_program()\ + is deprecated, use fluid.program_guard() or pass this argument\ + to transpile() call.") + s_prog = Program() - if not startup_program: - orig_s_prog = default_startup_program() - else: - orig_s_prog = startup_program + orig_s_prog = self.startup_program s_prog.random_seed = orig_s_prog.random_seed params = self.param_grad_ep_mapping[endpoint]["params"] @@ -1287,13 +1354,11 @@ class DistributeTranspiler(object): inputs={"X": vars2merge}, outputs={"Out": merged_var}, attrs={"use_mkldnn": False}) - # TODO(panyx0718): What if it's SELECTED_ROWS. - if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: - optimize_block.append_op( - type="scale", - inputs={"X": merged_var}, - outputs={"Out": merged_var}, - attrs={"scale": 1.0 / float(self.trainer_num)}) + optimize_block.append_op( + type="scale", + inputs={"X": merged_var}, + outputs={"Out": merged_var}, + attrs={"scale": 1.0 / float(self.trainer_num)}) return merged_var def _append_pserver_ops(self, optimize_block, opt_op, endpoint,