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。
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<divid="markdown"style='display:none'>
# Linear Regression
Let us begin the tutorial with a classical problem called Linear Regression \[[1](#References)\]. In this chapter, we will train a model from a realistic dataset to predict home prices. Some important concepts in Machine Learning will be covered through this example.
The source code for this tutorial lives on [book/fit_a_line](https://github.com/PaddlePaddle/book/tree/develop/01.fit_a_line). For instructions on getting started with PaddlePaddle, see [PaddlePaddle installation guide](https://github.com/PaddlePaddle/book/blob/develop/README.en.md#running-the-book).
## Problem Setup
Suppose we have a dataset of $n$ real estate properties. These real estate properties will be referred to as *homes* in this chapter for clarity.
Each home is associated with $d$ attributes. The attributes describe characteristics such the number of rooms in the home, the number of schools or hospitals in the neighborhood, and the traffic condition nearby.
In our problem setup, the attribute $x_{i,j}$ denotes the $j$th characteristic of the $i$th home. In addition, $y_i$ denotes the price of the $i$th home. Our task is to predict $y_i$ given a set of attributes $\{x_{i,1}, ..., x_{i,d}\}$. We assume that the price of a home is a linear combination of all of its attributes, namely,
where $\vec{\omega}$ and $b$ are the model parameters we want to estimate. Once they are learned, we will be able to predict the price of a home, given the attributes associated with it. We call this model **Linear Regression**. In other words, we want to regress a value against several values linearly. In practice, a linear model is often too simplistic to capture the real relationships between the variables. Yet, because Linear Regression is easy to train and analyze, it has been applied to a large number of real problems. As a result, it is an important topic in many classic Statistical Learning and Machine Learning textbooks \[[2,3,4](#References)\].
## Results Demonstration
We first show the result of our model. The dataset [UCI Housing Data Set](https://archive.ics.uci.edu/ml/datasets/Housing) is used to train a linear model to predict the home prices in Boston. The figure below shows the predictions the model makes for some home prices. The $X$-axis represents the median value of the prices of simlilar homes within a bin, while the $Y$-axis represents the home value our linear model predicts. The dotted line represents points where $X=Y$. When reading the diagram, the more precise the model predicts, the closer the point is to the dotted line.
In the UCI Housing Data Set, there are 13 home attributes $\{x_{i,j}\}$ that are related to the median home price $y_i$, which we aim to predict. Thus, our model can be written as:
where $\hat{Y}$ is the predicted value used to differentiate from actual value $Y$. The model learns parameters $\omega_1, \ldots, \omega_{13}, b$, where the entries of $\vec{\omega}$ are **weights** and $b$ is **bias**.
Now we need an objective to optimize, so that the learned parameters can make $\hat{Y}$ as close to $Y$ as possible. Let's refer to the concept of [Loss Function (Cost Function)](https://en.wikipedia.org/wiki/Loss_function). A loss function must output a non-negative value, given any pair of the actual value $y_i$ and the predicted value $\hat{y_i}$. This value reflects the magnitutude of the model error.
For Linear Regression, the most common loss function is [Mean Square Error (MSE)](https://en.wikipedia.org/wiki/Mean_squared_error) which has the following form:
That is, for a dataset of size $n$, MSE is the average value of the the prediction sqaure errors.
### Training
After setting up our model, there are several major steps to go through to train it:
1. Initialize the parameters including the weights $\vec{\omega}$ and the bias $b$. For example, we can set their mean values as $0$s, and their standard deviations as $1$s.
2. Feedforward. Evaluate the network output and compute the corresponding loss.
3. [Backpropagate](https://en.wikipedia.org/wiki/Backpropagation) the errors. The errors will be propagated from the output layer back to the input layer, during which the model parameters will be updated with the corresponding errors.
4. Repeat steps 2~3, until the loss is below a predefined threshold or the maximum number of repeats is reached.
## Dataset
### Python Dataset Modules
Our program starts with importing necessary packages:
```python
import paddle.v2 as paddle
import paddle.v2.dataset.uci_housing as uci_housing
```
We encapsulated the [UCI Housing Data Set](https://archive.ics.uci.edu/ml/datasets/Housing) in our Python module `uci_housing`. This module can
1. download the dataset to `~/.cache/paddle/dataset/uci_housing/housing.data`, if not yet, and
2. [preprocesses](#preprocessing) the dataset.
### An Introduction of the Dataset
The UCI housing dataset has 506 instances. Each instance describes the attributes of a house in surburban Boston. The attributes are explained below:
| Attribute Name | Characteristic | Data Type |
| ------| ------ | ------ |
| CRIM | per capita crime rate by town | Continuous|
| ZN | proportion of residential land zoned for lots over 25,000 sq.ft. | Continuous |
| INDUS | proportion of non-retail business acres per town | Continuous |
| CHAS | Charles River dummy variable | Discrete, 1 if tract bounds river; 0 otherwise|
| PTRATIO | pupil-teacher ratio by town | Continuous |
| B | 1000(Bk - 0.63)^2 where Bk is the proportion of blacks by town | Continuous |
| LSTAT | % lower status of the population | Continuous |
| MEDV | Median value of owner-occupied homes in $1000's | Continuous |
The last entry is the median home price.
### Preprocessing
#### Continuous and Discrete Data
We define a feature vector of length 13 for each home, where each entry corresponds to an attribute. Our first observation is that, among the 13 dimensions, there are 12 continuous dimensions and 1 discrete dimension.
Note that although a discrete value is also written as numeric values such as 0, 1, or 2, its meaning differs from a continuous value drastically. The linear difference between two discrete values has no meaning. For example, suppose $0$, $1$, and $2$ are used to represent colors *Red*, *Green*, and *Blue* respectively. Judging from the numeric representation of these colors, *Red* differs more from *Blue* than it does from *Green*. Yet in actuality, it is not true that extent to which the color *Blue* is different from *Red* is greater than the extent to which *Green* is different from *Red*. Therefore, when handling a discrete feature that has $d$ possible values, we usually convert it to $d$ new features where each feature takes a binary value, $0$ or $1$, indicating whether the original value is absent or present. Alternatively, the discrete features can be mapped onto a continuous multi-dimensional vector through an embedding table. For our problem here, because CHAS itself is a binary discrete value, we do not need to do any preprocessing.
#### Feature Normalization
We also observe a huge difference among the value ranges of the 13 features (Figure 2). For instance, the values of feature *B* fall in $[0.32, 396.90]$, whereas those of feature *NOX* has a range of $[0.3850, 0.8170]$. An effective optimization would require data normalization. The goal of data normalization is to scale te values of each feature into roughly the same range, perhaps $[-0.5, 0.5]$. Here, we adopt a popular normalization technique where we substract the mean value from the feature value and divide the result by the width of the original range.
There are at least three reasons for [Feature Normalization](https://en.wikipedia.org/wiki/Feature_scaling) (Feature Scaling):
- A value range that is too large or too small might cause floating number overflow or underflow during computation.
- Different value ranges might result in varying *importances* of different features to the model (at least in the beginning of the training process). This assumption about the data is often unreasonable, making the optimization difficult, which in turn results in increased training time.
- Many machine learning techniques or models (e.g., *L1/L2 regularization* and *Vector Space Model*) assumes that all the features have roughly zero means and their value ranges are similar.
<palign="center">
<imgsrc = "image/ranges_en.png"width=550><br/>
Figure 2. The value ranges of the features
</p>
#### Prepare Training and Test Sets
We split the dataset in two, one for adjusting the model parameters, namely, for model training, and the other for model testing. The model error on the former is called the **training error**, and the error on the latter is called the **test error**. Our goal in training a model is to find the statistical dependency between the outputs and the inputs, so that we can predict new outputs given new inputs. As a result, the test error reflects the performance of the model better than the training error does. We consider two things when deciding the ratio of the training set to the test set: 1) More training data will decrease the variance of the parameter estimation, yielding more reliable models; 2) More test data will decrease the variance of the test error, yielding more reliable test errors. One standard split ratio is $8:2$.
When training complex models, we usually have one more split: the validation set. Complex models usually have [Hyperparameters](https://en.wikipedia.org/wiki/Hyperparameter_optimization) that need to be set before the training process, such as the number of layers in the network. Because hyperparameters are not part of the model parameters, they cannot be trained using the same loss function. Thus we will try several sets of hyperparameters to train several models and cross-validate them on the validation set to pick the best one; finally, the selected trained model is tested on the test set. Because our model is relatively simple, we will omit this validation process.
## Training
`fit_a_line/trainer.py` demonstrates the training using [PaddlePaddle](http://paddlepaddle.org).
### Initialize PaddlePaddle
```python
paddle.init(use_gpu=False, trainer_count=1)
```
### Model Configuration
Logistic regression is essentially a fully-connected layer with linear activation:
```python
x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13))
y_predict = paddle.layer.fc(input=x,
size=1,
act=paddle.activation.Linear())
y = paddle.layer.data(name='y', type=paddle.data_type.dense_vector(1))
for loadinng training data. A reader may return multiple columns, and we need a Python dictionary to specify the mapping from column index to data layers.
```python
feeding={'x': 0, 'y': 1}
```
Moreover, an event handler is provided to print the training progress:
```python
# event_handler to print training and testing info
# event_handler to print training and testing info
from paddle.v2.plot import Ploter
train_title = "Train cost"
test_title = "Test cost"
plot_cost = Ploter(train_title, test_title)
step = 0
def event_handler_plot(event):
global step
if isinstance(event, paddle.event.EndIteration):
if step % 10 == 0: # every 10 batches, record a train cost
plot_cost.append(train_title, step, event.cost)
if step % 100 == 0: # every 100 batches, record a test cost
result = trainer.test(
reader=paddle.batch(
uci_housing.test(), batch_size=2),
feeding=feeding)
plot_cost.append(test_title, step, result.cost)
if step % 100 == 0: # every 100 batches, update cost plot
plot_cost.plot()
step += 1
```
### Start Training
```python
trainer.train(
reader=paddle.batch(
paddle.reader.shuffle(
uci_housing.train(), buf_size=500),
batch_size=2),
feeding=feeding,
event_handler=event_handler_plot,
num_passes=30)
```
## Summary
This chapter introduces *Linear Regression* and how to train and test this model with PaddlePaddle, using the UCI Housing Data Set. Because a large number of more complex models and techniques are derived from linear regression, it is important to understand its underlying theory and limitation.
2. Friedman J, Hastie T, Tibshirani R. The elements of statistical learning[M]. Springer, Berlin: Springer series in statistics, 2001.
3. Murphy K P. Machine learning: a probabilistic perspective[M]. MIT press, 2012.
4. Bishop C M. Pattern recognition[J]. Machine Learning, 2006, 128.
<br/>
This tutorial is contributed by <axmlns:cc="http://creativecommons.org/ns#"href="http://book.paddlepaddle.org"property="cc:attributionName"rel="cc:attributionURL">PaddlePaddle</a>, and licensed under a <arel="license"href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<divid="markdown"style='display:none'>
<divid="markdown"style='display:none'>
# Recognize Digits
# 识别数字
The source code for this tutorial is live at [book/recognize_digits](https://github.com/PaddlePaddle/book/tree/develop/02.recognize_digits). For instructions on getting started with Paddle, please refer to [installation instructions](https://github.com/PaddlePaddle/book/blob/develop/README.en.md#running-the-book).
When one learns to program, the first task is usually to write a program that prints "Hello World!". In Machine Learning or Deep Learning, the equivalent task is to train a model to recognize hand-written digits on the dataset [MNIST](http://yann.lecun.com/exdb/mnist/). Handwriting recognition is a classic image classification problem. The problem is relatively easy and MNIST is a complete dataset. As a simple Computer Vision dataset, MNIST contains images of handwritten digits and their corresponding labels (Fig. 1). The input image is a $28\times28$ matrix, and the label is one of the digits from $0$ to $9$. All images are normalized, meaning that they are both rescaled and centered.
The MNIST dataset is created from the [NIST](https://www.nist.gov/srd/nist-special-database-19) Special Database 3 (SD-3) and the Special Database 1 (SD-1). The SD-3 is labeled by the staff of the U.S. Census Bureau, while SD-1 is labeled by high school students the in U.S. Therefore the SD-3 is cleaner and easier to recognize than the SD-1 dataset. Yann LeCun et al. used half of the samples from each of SD-1 and SD-3 to create the MNIST training set (60,000 samples) and test set (10,000 samples), where training set was labeled by 250 different annotators, and it was guaranteed that there wasn't a complete overlap of annotators of training set and test set.
Yann LeCun, one of the founders of Deep Learning, have previously made tremendous contributions to handwritten character recognition and proposed the **Convolutional Neural Network** (CNN), which drastically improved recognition capability for handwritten characters. CNNs are now a critical concept in Deep Learning. From the LeNet proposal by Yann LeCun, to those winning models in ImageNet competitions, such as VGGNet, GoogLeNet, and ResNet (See [Image Classification](https://github.com/PaddlePaddle/book/tree/develop/03.image_classification) tutorial), CNNs have achieved a series of impressive results in Image Classification tasks.
Many algorithms are tested on MNIST. In 1998, LeCun experimented with single layer linear classifier, Multilayer Perceptron (MLP) and Multilayer CNN LeNet. These algorithms quickly reduced test error from 12% to 0.7% \[[1](#references)\]. Since then, researchers have worked on many algorithms such as **K-Nearest Neighbors** (k-NN) \[[2](#references)\], **Support Vector Machine** (SVM) \[[3](#references)\], **Neural Networks** \[[4-7](#references)\] and **Boosting** \[[8](#references)\]. Various preprocessing methods like distortion removal, noise removal, and blurring, have also been applied to increase recognition accuracy.
In this tutorial, we tackle the task of handwritten character recognition. We start with a simple **softmax** regression model and guide our readers step-by-step to improve this model's performance on the task of recognition.
本教程中，我们从简单的模型Softmax回归开始，带大家入门手写字符识别，并逐步进行模型优化。
## Model Overview
## 模型概览
Before introducing classification algorithms and training procedure, we define the following symbols:
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：
- $X$ is the input: Input is a $28\times 28$ MNIST image. It is flattened to a $784$ dimensional vector. $X=\left (x_0, x_1, \dots, x_{783} \right )$.
- $Y$ is the output: Output of the classifier is 1 of the 10 classes (digits from 0 to 9). $Y=\left (y_0, y_1, \dots, y_9 \right )$. Each dimension $y_i$ represents the probability that the input image belongs to class $i$.
- $L$ is the ground truth label: $L=\left ( l_0, l_1, \dots, l_9 \right )$. It is also 10 dimensional, but only one entry is $1$ and all others are $0$s.
In a simple softmax regression model, the input is first fed to fully connected layers. Then, a softmax function is applied to output probabilities of multiple output classes\[[9](#references)\].
For an $N$-class classification problem with $N$ output nodes, Softmax normalizes the resulting $N$ dimensional vector so that each of its entries falls in the range $[0,1]\in\math{R}$, representing the probability that the sample belongs to a certain class. Here $y_i$ denotes the predicted probability that an image is of digit $i$.
The softmax regression model described above uses the simplest two-layer neural network. That is, it only contains an input layer and an output layer, with limited regression capability. To achieve better recognition results, consider adding several hidden layers\[[10](#references)\] between the input layer and the output layer.
1. After the first hidden layer, we get $ H_1 = \phi(W_1X + b_1) $, where $\phi$ denotes the activation function. Some [common ones](###list-of-common-activation-functions) are sigmoid, tanh and ReLU.
The **convolutional layer** is the core of a Convolutional Neural Network. The parameters in this layer are composed of a set of filters, also called kernels. We could visualize the convolution step in the following fashion: Each kernel slides horizontally and vertically till it covers the whole image. At every window, we compute the dot product of the kernel and the input. Then, we add the bias and apply an activation function. The result is a two-dimensional activation map. For example, some kernel may recognize corners, and some may recognize circles. These convolution kernels may respond strongly to the corresponding features.
Fig. 4 illustrates the dynamic programming of a convolutional layer, where depths are flattened for simplicity. The input is $W_1=5$, $H_1=5$, $D_1=3$. In fact, this is a common representation for colored images. $W_1$ and $H_1$ correspond to the width and height in a colored image. $D_1$ corresponds to the 3 color channels for RGB. The parameters of the convolutional layer are $K=2$, $F=3$, $S=2$, $P=1$. $K$ denotes the number of kernels; specifically, $Filter$ $W_0$ and $Filter$ $W_1$ are the kernels. $F$ is kernel size while $W0$ and $W1$ are both $F\timesF = 3\times3$ matrices in all depths. $S$ is the stride, which is the width of the sliding window; here, kernels move leftwards or downwards by 2 units each time. $P$ is the width of the padding, which denotes an extension of the input; here, the gray area shows zero padding with size 1.
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。
A **pooling layer** performs downsampling. The main functionality of this layer is to reduce computation by reducing the network parameters. It also prevents over-fitting to some extent. Usually, a pooling layer is added after a convolutional layer. Pooling layer can use various techniques, such as max pooling and average pooling. As shown in Fig.5, max pooling uses rectangles to segment the input layer into several parts and computes the maximum value in each part as the output.
[**LeNet-5**](http://yann.lecun.com/exdb/lenet/) is one of the simplest Convolutional Neural Networks. Fig. 6. shows its architecture: A 2-dimensional input image is fed into two sets of convolutional layers and pooling layers. This output is then fed to a fully connected layer and a softmax classifier. Compared to multilayer, fully connected perceptrons, the LeNet-5 can recognize images better. This is due to the following three properties of the convolution:
- The 3D nature of the neurons: a convolutional layer is organized by width, height and depth. Neurons in each layer are connected to only a small region in the previous layer. This region is called the receptive field.
- Local connectivity: A CNN utilizes the local space correlation by connecting local neurons. This design guarantees that the learned filter has a strong response to local input features. Stacking many such layers generates a non-linear filter that is more global. This enables the network to first obtain good representation for small parts of input and then combine them to represent a larger region.
- Weight sharing: In a CNN, computation is iterated on shared parameters (weights and bias) to form a feature map. This means that all the neurons in the same depth of the output respond to the same feature. This allows the network to detect a feature regardless of its position in the input. In other words, it is shift invariant.
For more details on Convolutional Neural Networks, please refer to the tutorial on [Image Classification](https://github.com/PaddlePaddle/book/blob/develop/image_classification/README.md) and the [relevant lecture](http://cs231n.github.io/convolutional-networks/) from a Stanford open course.
In fact, tanh function is just a rescaled version of the sigmoid function. It is obtained by magnifying the value of the sigmoid function and moving it downwards by 1.
PaddlePaddle provides a Python module, `paddle.dataset.mnist`, which downloads and caches the [MNIST dataset](http://yann.lecun.com/exdb/mnist/). The cache is under `/home/username/.cache/paddle/dataset/mnist`:
PaddlePaddle provides a special layer `layer.data` for reading data. Let us create a data layer for reading images and connect it to a classification network created using one of above three functions. We also need a cost layer for training the model.
Now, it is time to specify training parameters. In the following `Momentum` optimizer, `momentum=0.9` means that 90% of the current momentum comes from that of the previous iteration. The learning rate relates to the speed at which the network training converges. Regularization is meant to prevent over-fitting; here we use the L2 regularization.
Then we specify the training data `paddle.dataset.movielens.train()` and testing data `paddle.dataset.movielens.test()`. These two methods are *reader creators*. Once called, a reader creator returns a *reader*. A reader is a Python method, which, once called, returns a Python generator, which yields instances of data.
`shuffle` is a reader decorator. It takes in a reader A as input and returns a new reader B. Under the hood, B calls A to read data in the following fashion: it copies in `buffer_size` instances at a time into a buffer, shuffles the data, and yields the shuffled instances one at a time. A large buffer size would yield very shuffled data.
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` is a special decorator, which takes in reader and outputs a *batch reader*, which doesn't yield an instance, but a minibatch at a time.
best = sorted(lists, key=lambda list: float(list[1]))[0]
print 'Best pass is %s, testing Avgcost is %s' % (best[0], best[1])
print 'The classification accuracy is %.2f%%' % (100 - float(best[2]) * 100)
```
Usually, with MNIST data, the softmax regression model achieves an accuracy around 92.34%, the MLP 97.66%, and the convolution network around 99.20%. Convolution layers have been widely considered a great invention for image processing.
After training is done, user can use the trained model to classify images. The following code shows how to inference MNIST images through `paddle.infer` interface.
```python
```python
from PIL import Image
from PIL import Image
...
@@ -381,17 +377,11 @@ lab = np.argsort(-probs) # probs and lab are the results of one batch data
...
@@ -381,17 +377,11 @@ lab = np.argsort(-probs) # probs and lab are the results of one batch data
print "Label of image/infer_3.png is: %d" % lab[0][0]
print "Label of image/infer_3.png is: %d" % lab[0][0]
This tutorial describes a few common deep learning models using **Softmax regression**, **Multilayer Perceptron Network**, and **Convolutional Neural Network**. Understanding these models is crucial for future learning; the subsequent tutorials derive more sophisticated networks by building on top of them.
When our model evolves from a simple softmax regression to a slightly complex Convolutional Neural Network, the recognition accuracy on the MNIST data set achieves a large improvement in accuracy. This is due to the Convolutional layers' local connections and parameter sharing. While learning new models in the future, we encourage the readers to understand the key ideas that lead a new model to improve the results of an old one.
Moreover, this tutorial introduces the basic flow of PaddlePaddle model design, which starts with a *dataprovider*, a model layer construction, and finally training and prediction. Motivated readers can leverage the flow used in this MNIST handwritten digit classification example and experiment with different data and network architectures to train models for classification tasks of their choice.
## References
## 参考文献
1. LeCun, Yann, Léon Bottou, Yoshua Bengio, and Patrick Haffner. ["Gradient-based learning applied to document recognition."](http://ieeexplore.ieee.org/abstract/document/726791/) Proceedings of the IEEE 86, no. 11 (1998): 2278-2324.
1. LeCun, Yann, Léon Bottou, Yoshua Bengio, and Patrick Haffner. ["Gradient-based learning applied to document recognition."](http://ieeexplore.ieee.org/abstract/document/726791/) Proceedings of the IEEE 86, no. 11 (1998): 2278-2324.
2. Wejéus, Samuel. ["A Neural Network Approach to Arbitrary SymbolRecognition on Modern Smartphones."](http://www.diva-portal.org/smash/record.jsf?pid=diva2%3A753279&dswid=-434) (2014).
2. Wejéus, Samuel. ["A Neural Network Approach to Arbitrary SymbolRecognition on Modern Smartphones."](http://www.diva-portal.org/smash/record.jsf?pid=diva2%3A753279&dswid=-434) (2014).
...
@@ -405,7 +395,7 @@ Moreover, this tutorial introduces the basic flow of PaddlePaddle model design,
...
@@ -405,7 +395,7 @@ Moreover, this tutorial introduces the basic flow of PaddlePaddle model design,
10. Bishop, Christopher M. ["Pattern recognition."](http://users.isr.ist.utl.pt/~wurmd/Livros/school/Bishop%20-%20Pattern%20Recognition%20And%20Machine%20Learning%20-%20Springer%20%202006.pdf) Machine Learning 128 (2006): 1-58.
10. Bishop, Christopher M. ["Pattern recognition."](http://users.isr.ist.utl.pt/~wurmd/Livros/school/Bishop%20-%20Pattern%20Recognition%20And%20Machine%20Learning%20-%20Springer%20%202006.pdf) Machine Learning 128 (2006): 1-58.
<br/>
<br/>
This tutorial is contributed by <axmlns:cc="http://creativecommons.org/ns#"href="http://book.paddlepaddle.org"property="cc:attributionName"rel="cc:attributionURL">PaddlePaddle</a>, and licensed under a <arel="license"href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<divid="markdown"style='display:none'>
<divid="markdown"style='display:none'>
Image Classification
# 图像分类
=======================
The source code for this chapter is at [book/image_classification](https://github.com/PaddlePaddle/book/tree/develop/03.image_classification). First-time users, please refer to PaddlePaddle [Installation Tutorial](https://github.com/PaddlePaddle/book/blob/develop/README.en.md#running-the-book) for installation instructions.
Compared to words, images provide much more vivid and easier to understand information with an artistic sense. They are an important source for people to express and exchange ideas. In this chapter, we focus on one of the essential problems in image recognition -- image classification.
Image classification is the task of distinguishing images in different categories based on their semantic meaning. It is a core problem in computer vision and is also the foundation of other higher level computer vision tasks such as object detection, image segmentation, object tracking, action recognition, etc. Image classification has applications in many areas such as face recognition, intelligent video analysis in security systems, traffic scene recognition in transportation systems, content-based image retrieval and automatic photo indexing in web services, image classification in medicine, etc.
To classify an image we first encode the entire image using handcrafted or learned features and then determine the category using a classifier. Thus, feature extraction plays an important role in image classification. Prior to deep learning the BoW(Bag of Words) model was the most widely used method for classifying an image as well as an object. The BoW technique was introduced in Natural Language Processing where a training sentence is represented as a bag of words. In the context of image classification, the BoW model requires constructing a dictionary. The simplest BoW framework can be designed with three steps: **feature extraction**, **feature encoding** and **classifier design**.
Using Deep learning, image classification can be framed as a supervised or unsupervised learning problem that uses hierarchical features automatically without any need for manually crafted features from the image. In recent years, Convolutional Neural Networks (CNNs) have made significant progress in image classification. CNNs use raw image pixels as input, extract low-level and high-level abstract features through convolution operations, and directly output the classification results from the model. This style of end-to-end learning has lead to not only increased performance but also wider adoption various applications.
一般来说，图像分类通过手工特征或特征学习方法对整个图像进行全部描述，然后使用分类器判别物体类别，因此如何提取图像的特征至关重要。在深度学习算法之前使用较多的是基于词袋(Bag of Words)模型的物体分类方法。词袋方法从自然语言处理中引入，即一句话可以用一个装了词的袋子表示其特征，袋子中的词为句子中的单词、短语或字。对于图像而言，词袋方法需要构建字典。最简单的词袋模型框架可以设计为**底层特征抽取**、**特征编码**、**分类器设计**三个过程。
In this chapter, we introduce deep-learning-based image classification methods and explain how to train a CNN model using PaddlePaddle.
Figure 1 shows the results of a general image classifier -- the trained model can correctly recognize the main objects in the images.
<palign="center">
<palign="center">
<imgsrc="image/dog_cat.png "width="350"><br/>
<imgsrc="image/dog_cat.png "width="350"><br/>
Figure 1. General image classification
图1. 通用图像分类展示
</p>
</p>
Figure 2 shows the results of a fine-grained image classifier. This task of flower recognition requires correctly recognizing of the flower's categories.
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。
<palign="center">
<palign="center">
<imgsrc="image/flowers.png"width="400"><br/>
<imgsrc="image/flowers.png"width="400"><br/>
Figure 2. Fine-grained image classification
图2. 细粒度图像分类展示
</p>
</p>
A good model should recognize objects of different categories correctly. The results of such a model should not vary due to viewpoint variation, illumination conditions, object distortion or occlusion.
A large amount of research in image classification is built upon public datasets such as [PASCAL VOC](http://host.robots.ox.ac.uk/pascal/VOC/), [ImageNet](http://image-net.org/) etc. Many image classification algorithms are usually evaluated and compared on these datasets. PASCAL VOC is a computer vision competition started in 2005, and ImageNet is a dataset for Large Scale Visual Recognition Challenge (ILSVRC) started in 2010. In this chapter, we introduce some image classification models from the submissions to these competitions.
Before 2012, traditional image classification was accomplished with the three steps described in the background section. A complete model construction usually involves the following stages: low-level feature extraction, feature encoding, spatial constraint or feature clustering, classifier design, model ensemble.
1). **Low-level feature extraction**: This step extracts large amounts of local features according to fixed strides and scales. Popular local features include Scale-Invariant Feature Transform (SIFT)[1], Histogram of Oriented Gradient(HOG)[2], Local Binary Pattern(LBP)[3], etc. A common practice is to employ multiple feature descriptors in order to avoid missing a lot of information.
2). **Feature encoding**: Low-level features contain a large amount of redundancy and noise. In order to improve the robustness of features, it is necessary to employ a feature transformation to encode low-level features. This is called feature encoding. Common feature encoding methods include vector quantization [4], sparse coding [5], locality-constrained linear coding [6], Fisher vector encoding [7], etc.
3). **Spatial constraint**: Spatial constraint or feature clustering is usually adopted after feature encoding for extracting the maximum or average of each dimension in the spatial domain. Pyramid feature matching--a popular feature clustering method--divides an image uniformly into patches and performs feature clustering in each patch.
4). **Classification**: In the above steps an image can be described by a vector of fixed dimension. Then a classifier can be used to classify the image into categories. Common classifiers include Support Vector Machine(SVM), random forest etc. Kernel SVM is the most popular classifier and has achieved very good performance in traditional image classification tasks.
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。
This method has been used widely as image classification algorithm in PASCAL VOC [18]. NEC Labs(http://www.nec-labs.com/) won the championship by employing SIFT and LBP features, two non-linear encoders and SVM in ILSVRC 2010 [8].
The CNN model--AlexNet proposed by Alex Krizhevsky et al.[9], made a breakthrough in ILSVRC 2012. It dramatically outperformed traditional methods and won the ILSVRC championship in 2012. This was also the first time that a deep learning method was used for large-scale image classification. Since AlexNet, a series of CNN models have been proposed that have advanced the state of the art steadily on Imagenet as shown in Figure 4. With deeper and more sophisticated architectures, Top-5 error rate is getting lower and lower (to around 3.5%). The error rate of human raters on the same Imagenet dataset is 5.1%, which means that the image classification capability of a deep learning model has surpassed human raters.
<palign="center">
<palign="center">
<imgsrc="image/ilsvrc.png"width="500"><br/>
<imgsrc="image/ilsvrc.png"width="500"><br/>
Figure 4. Top-5 error rates on ILSVRC image classification
图4. ILSVRC图像分类Top-5错误率
</p>
</p>
### CNN
### CNN
Traditional CNNs consist of convolutional and fully-connected layers and use the softmax multi-category classifier with the cross-entropy loss function. Figure 5 shows a typical CNN. We first introduce the common components of a CNN.
- convolutional layer: this layer uses the convolution operation to extract (low-level and high-level) features and to discover local correlation and spatial invariance.
- pooling layer: this layer down samples feature maps by extracting local max (max-pooling) or average (avg-pooling) value of each patch in the feature map. Down-sampling is a common operation in image processing and is used to filter out high-frequency information.
- non-linear activation: Convolutional and fully-connected layers are usually followed by some non-linear activation layers. Non-linearities enhance the expression capability of the network. Some examples of non-linear activation functions are Sigmoid, Tanh and ReLU. ReLU is the most commonly used activation function in CNN.
接下来我们主要介绍VGG，GoogleNet和ResNet网络结构。
- Dropout [10]: At each training stage, individual nodes are dropped out of the network with a certain probability. This improves the network's ability to generalize and avoids overfitting.
Parameter updates at each layer during training causes input layer distributions to change and in turn requires hyper-parameters to be careful tuned. In 2015, Sergey Ioffe and Christian Szegedy proposed a Batch Normalization (BN) algorithm [14], which normalizes the features of each batch in a layer, and enables relatively stable distribution in each layer. Not only does BN algorithm act as a regularizer, but also reduces the need for careful hyper-parameter design. Experiments demonstrate that BN algorithm accelerates the training convergence and has been widely used in later deeper models.
In the following sections, we will introduce the following network architectures - VGG, GoogleNet and ResNets.
### VGG
### VGG
The Oxford Visual Geometry Group (VGG) proposed the VGG network in ILSVRC 2014 [11]. This model is deeper and wider than previous neural architectures. It consists of five main groups of convolution operations. Adjacent convolution groups are connected via max-pooling layers. Each group contains a series of 3x3 convolutional layers (i.e. kernels). The number of convolution kernels stays the same within the group and increases from 64 in the first group to 512 in the last one. The total number of learnable layers could be 11, 13, 16, or 19 depending on the number of convolutional layers in each group. Figure 6 illustrates a 16-layer VGG. The neural architecture of VGG is relatively simple and has been adopted by many papers such as the first one that surpassed human-level performance on ImageNet [19].
GoogleNet [12] won the ILSVRC championship in 2014. GoogleNet borrowed some ideas from the Network in Network(NIN) model [13] and is built on the Inception blocks. Let us first familiarize ourselves with these first.
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。
The two main characteristics of the NIN model are:
1) A single-layer convolutional network is replaced with a Multi-Layer Perceptron Convolution (MLPconv). MLPconv is a tiny multi-layer convolutional network. It enhances non-linearity by adding several 1x1 convolutional layers after linear ones.
2) In traditional CNNs, the last fewer layers are usually fully-connected with a large number of parameters. In contrast, NIN replaces all fully-connected layers with convolutional layers with feature maps of the same size as the category dimension and a global average pooling. This replacement of fully-connected layers significantly reduces the number of parameters.
Figure 7 depicts two Inception blocks. Figure 7(a) is the simplest design. The output is a concatenation of features from three convolutional layers and one pooling layer. The disadvantage of this design is that the pooling layer does not change the number of filters and leads to an increase in the number of outputs. After several of such blocks, the number of outputs and parameters become larger and larger and lead to higher computation complexity. To overcome this drawback, the Inception block in Figure 7(b) employs three 1x1 convolutional layers. These reduce dimensions or the number of channels but improve the non-linearity of the network.
<palign="center">
<palign="center">
<imgsrc="image/inception_en.png"width="800"><br/>
<imgsrc="image/inception.png"width="800"><br/>
Figure 7. Inception block
图7. Inception模块
</p>
</p>
GoogleNet consists of multiple stacked Inception blocks followed by an avg-pooling layer as in NIN instead of traditional fully connected layers. The difference between GoogleNet and NIN is that GoogleNet adds a fully connected layer after avg-pooling layer to output a vector of category size. Besides these two characteristics, the features from middle layers of a GoogleNet are also very discriminative. Therefore, GoogeleNet inserts two auxiliary classifiers in the model for enhancing gradient and regularization when doing backpropagation. The loss function of the whole network is the weighted sum of these three classifiers.
Figure 8 illustrates the neural architecture of a GoogleNet which consists of 22 layers: it starts with three regular convolutional layers followed by three groups of sub-networks -- the first group contains two Inception blocks, the second one five, and the third one two. It ends up with an average pooling and a fully-connected layer.
The above model is the first version of GoogleNet or GoogelNet-v1. GoogleNet-v2 [14] introduced BN layer; GoogleNet-v3 [16] further split some convolutional layers, which increases non-linearity and network depth; GoogelNet-v4 [17] leads to the design idea of ResNet which will be introduced in the next section. The evolution from v1 to v4 improved the accuracy rate consistently. We will not go into details of the neural architectures of v2 to v4.
Residual Network(ResNet)[15] won the 2015 championship on three ImageNet competitions -- image classification, object localization, and object detection. The main challenge in training deeper networks is that accuracy degrades with network depth. The authors of ResNet proposed a residual learning approach to ease the difficulty of training deeper networks. Based on the design ideas of BN, small convolutional kernels, full convolutional network, ResNets reformulate the layers as residual blocks, with each block containing two branches, one directly connecting input to the output, the other performing two to three convolutions and calculating the residual function with reference to the layer inputs. The outputs of these two branches are then added up.
Figure 9 illustrates the ResNet architecture. To the left is the basic building block, it consists of two 3x3 convolutional layers of the same channels. To the right is a Bottleneck block. The bottleneck is a 1x1 convolutional layer used to reduce dimension from 256 to 64. The other 1x1 convolutional layer is used to increase dimension from 64 to 256. Thus, the number of input and output channels of the middle 3x3 convolutional layer is 64, which is relatively small.
Figure 10 illustrates ResNets with 50, 101, 152 layers, respectively. All three networks use bottleneck blocks of different numbers of repetitions. ResNet converges very fast and can be trained with hundreds or thousands of layers.
Commonly used public datasets for image classification are [CIFAR](https://www.cs.toronto.edu/~kriz/cifar.html), [ImageNet](http://image-net.org/), [COCO](http://mscoco.org/), etc. Those used for fine-grained image classification are [CUB-200-2011](http://www.vision.caltech.edu/visipedia/CUB-200-2011.html), [Stanford Dog](http://vision.stanford.edu/aditya86/ImageNetDogs/), [Oxford-flowers](http://www.robots.ox.ac.uk/~vgg/data/flowers/), etc. Among these, the ImageNet dataset is the largest. Most research results are reported on ImageNet as mentioned in the Model Overview section. Since 2010, the ImageNet dataset has gone through some changes. The commonly used ImageNet-2012 dataset contains 1000 categories. There are 1,281,167 training images, ranging from 732 to 1200 images per category, and 50,000 validation images with 50 images per category in average.
Since ImageNet is too large to be downloaded and trained efficiently, we use [CIFAR-10](https://www.cs.toronto.edu/~kriz/cifar.html) in this tutorial. The CIFAR-10 dataset consists of 60000 32x32 color images in 10 classes, with 6000 images per class. There are 50000 training images and 10000 test images. Figure 11 shows all the classes in CIFAR-10 as well as 10 images randomly sampled from each category.
`paddle.datasets` package encapsulates multiple public datasets, including `cifar`, `imdb`, `mnist`, `moivelens` and `wmt14`, etc. There's no need to manually download and preprocess CIFAR-10.
We must import and initialize PaddlePaddle (enable/disable GPU, set the number of trainers, etc).
通过 `paddle.init`，初始化Paddle是否使用GPU，trainer的数目等等。
```python
```python
import sys
import sys
...
@@ -220,29 +207,30 @@ from resnet import resnet_cifar10
...
@@ -220,29 +207,30 @@ from resnet import resnet_cifar10
paddle.init(use_gpu=False, trainer_count=1)
paddle.init(use_gpu=False, trainer_count=1)
```
```
As mentioned in section [Model Overview](#model-overview), here we provide the implementations of the VGG and ResNet models.
本教程中我们提供了VGG和ResNet两个模型的配置。
### VGG
#### VGG
First, we use a VGG network. Since the image size and amount of CIFAR10 are relatively small comparing to ImageNet, we use a small version of VGG network for CIFAR10. Convolution groups incorporate BN and dropout operations.
The input to the network is defined as `paddle.layer.data`, or image pixels in the context of image classification. The images in CIFAR10 are 32x32 color images of three channels. Therefore, the size of the input data is 3072 (3x32x32), and the number of categories is 10.
The input to VGG main module is from the data layer. `vgg_bn_drop` defines a 16-layer VGG network, with each convolutional layer followed by BN and dropout layers. Here is the definition in detail:
@@ -275,15 +263,15 @@ First, we use a VGG network. Since the image size and amount of CIFAR10 are rela
...
@@ -275,15 +263,15 @@ First, we use a VGG network. Since the image size and amount of CIFAR10 are rela
return fc2
return fc2
```
```
2.1. First, define a convolution block or conv_block. The default convolution kernel is 3x3, and the default pooling size is 2x2 with stride 2. Dropout specifies the probability in dropout operation. Function `img_conv_group` is defined in `paddle.networks` consisting of a series of `Conv->BN->ReLu->Dropout` and a `Pooling`.
2.2. Five groups of convolutions. The first two groups perform two convolutions, while the last three groups perform three convolutions. The dropout rate of the last convolution in each group is set to 0, which means there is no dropout for this layer.
2.3. The last two layers are fully-connected layers of dimension 512.
2.3. 最后接两层512维的全连接。
3. Define Classifier
3. 定义分类器
The above VGG network extracts high-level features and maps them to a vector of the same size as the categories. Softmax function or classifier is then used for calculating the probability of the image belonging to each category.
@@ -291,9 +279,9 @@ First, we use a VGG network. Since the image size and amount of CIFAR10 are rela
...
@@ -291,9 +279,9 @@ First, we use a VGG network. Since the image size and amount of CIFAR10 are rela
act=paddle.activation.Softmax())
act=paddle.activation.Softmax())
```
```
4. Define Loss Function and Outputs
4. 定义损失函数和网络输出
In the context of supervised learning, labels of training images are defined in `paddle.layer.data` as well. During training, the cross-entropy loss function is used and the loss is the output of the network. During testing, the outputs are the probabilities calculated in the classifier.
Here are some basic functions used in `resnet_cifar10`:
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` : convolutional layer followed by BN.
- `conv_bn_layer` : 带BN的卷积层。
- `shortcut` : the shortcut branch in a residual block. There are two kinds of shortcuts: 1x1 convolution used when the number of channels between input and output is different; direct connection used otherwise.
- `bottleneck` : a bottleneck module as shown in the right of Figure 9, it consists of two 1x1 convolutions with one 3x3 convolution in between branch and a "shortcut" branch.
- `layer_warp` : a group of residual modules consisting of several stacking blocks. In each group, the sliding window size of the first residual block could be different from the rest of blocks, in order to reduce the size of feature maps along horizontal and vertical directions.
Note: besides the first convolutional layer and the last fully-connected layer, the total number of layers in three `layer_warp` should be dividable by 6, that is the depth of `resnet_cifar10` should satisfy $(depth - 2) % 6 == 0$.
The learning rate adjustment policy can be defined with variables `learning_rate_decay_a`($a$), `learning_rate_decay_b`($b$) and `learning_rate_schedule`. In this example, discrete exponential method is used for adjusting learning rate. The formula is as follows,
After training is done, users can use the trained model to classify images. The following code shows how to infer through `paddle.infer` interface. You can remove the comments to change the model name.
```python
```python
from PIL import Image
from PIL import Image
...
@@ -534,22 +526,20 @@ def load_image(file):
...
@@ -534,22 +526,20 @@ def load_image(file):
im = Image.open(file)
im = Image.open(file)
im = im.resize((32, 32), Image.ANTIALIAS)
im = im.resize((32, 32), Image.ANTIALIAS)
im = np.array(im).astype(np.float32)
im = np.array(im).astype(np.float32)
# The storage order of the loaded image is W(widht),
# PIL打开图片存储顺序为H(高度)，W(宽度)，C(通道)。
# H(height), C(channel). PaddlePaddle requires
# PaddlePaddle要求数据顺序为CHW，所以需要转换顺序。
# the CHW order, so transpose them.
im = im.transpose((2, 0, 1)) # CHW
im = im.transpose((2, 0, 1)) # CHW
# In the training phase, the channel order of CIFAR
# CIFAR训练图片通道顺序为B(蓝),G(绿),R(红),
# image is B(Blue), G(green), R(Red). But PIL open
# 而PIL打开图片默认通道顺序为RGB,因为需要交换通道。
# image in RGB mode. It must swap the channel order.
@@ -560,12 +550,12 @@ print "Label of image/dog.png is: %d" % lab[0][0]
...
@@ -560,12 +550,12 @@ print "Label of image/dog.png is: %d" % lab[0][0]
```
```
## Conclusion
## 总结
Traditional image classification methods have complicated frameworks that involve multiple stages of processing. In contrast, CNN models can be trained end-to-end with a significant increase in classification accuracy. In this chapter, we introduced three models -- VGG, GoogleNet, ResNet and provided PaddlePaddle config files for training VGG and ResNet on CIFAR10. We also explained how to perform prediction and feature extraction using the PaddlePaddle API. For other datasets such as ImageNet, the procedure for config and training are the same and you are welcome to give it a try.
[1] D. G. Lowe, [Distinctive image features from scale-invariant keypoints](http://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf). IJCV, 60(2):91-110, 2004.
[1] D. G. Lowe, [Distinctive image features from scale-invariant keypoints](http://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf). IJCV, 60(2):91-110, 2004.
...
@@ -612,7 +602,7 @@ Traditional image classification methods have complicated frameworks that involv
...
@@ -612,7 +602,7 @@ Traditional image classification methods have complicated frameworks that involv
[22] http://cs231n.github.io/classification/
[22] http://cs231n.github.io/classification/
<br/>
<br/>
This tutorial is contributed by <axmlns:cc="http://creativecommons.org/ns#"href="http://book.paddlepaddle.org"property="cc:attributionName"rel="cc:attributionURL">PaddlePaddle</a>, and licensed under a <arel="license"href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<divid="markdown"style='display:none'>
<divid="markdown"style='display:none'>
# Word2Vec
This is intended as a reference tutorial. The source code of this tutorial lives on [book/word2vec](https://github.com/PaddlePaddle/book/tree/develop/04.word2vec).
# 词向量
For instructions on getting started with PaddlePaddle, see [PaddlePaddle installation guide](https://github.com/PaddlePaddle/book/blob/develop/README.en.md#running-the-book).
This section introduces the concept of **word embedding**, which is a vector representation of words. It is a popular technique used in natural language processing. Word embeddings support many Internet services, including search engines, advertising systems, and recommendation systems.
Building these services requires us to quantify the similarity between two words or paragraphs. This calls for a new representation of all the words to make them more suitable for computation. An obvious way to achieve this is through the vector space model, where every word is represented as an **one-hot vector**.
For each word, its vector representation has the corresponding entry in the vector as 1, and all other entries as 0. The lengths of one-hot vectors match the size of the dictionary. Each entry of a vector corresponds to the presence (or absence) of a word in the dictionary.
One-hot vectors are intuitive, yet they have limited usefulness. Take the example of an Internet advertising system: Suppose a customer enters the query "Mother's Day", while an ad bids for the keyword carnations". Because the one-hot vectors of these two words are perpendicular, the metric distance (either Euclidean or cosine similarity) between them would indicate little relevance. However, *we* know that these two queries are connected semantically, since people often gift their mothers bundles of carnation flowers on Mother's Day. This discrepancy is due to the low information capacity in each vector. That is, comparing the vector representations of two words does not assess their relevance sufficiently. To calculate their similarity accurately, we need more information, which could be learned from large amounts of data through machine learning methods.
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：
Like many machine learning models, word embeddings can represent knowledge in various ways. Another model may project an one-hot vector to an embedding vector of lower dimension e.g. $embedding(mother's day) = [0.3, 4.2, -1.5, ...], embedding(carnations) = [0.2, 5.6, -2.3, ...]$. Mapping one-hot vectors onto an embedded vector space has the potential to bring the embedding vectors of similar words (either semantically or usage-wise) closer to each other, so that the cosine similarity between the corresponding vectors for words like "Mother's Day" and "carnations" are no longer zero.
A word embedding model could be a probabilistic model, a co-occurrence matrix model, or a neural network. Before people started using neural networks to generate word embedding, the traditional method was to calculate a co-occurrence matrix $X$ of words. Here, $X$ is a $|V| \times |V|$ matrix, where $X_{ij}$ represents the co-occurrence times of the $i$th and $j$th words in the vocabulary `V` within all corpus, and $|V|$ is the size of the vocabulary. By performing matrix decomposition on $X$ e.g. Singular Value Decomposition \[[5](#References)\]
$$X = USV^T$$
$$X = USV^T$$
the resulting $U$ can be seen as the word embedding of all the words.
1) Since many pairs of words don't co-occur, the co-occurrence matrix is sparse. To achieve good performance of matrix factorization, further treatment on word frequency is needed;
2) The matrix is large, frequently on the order of $10^6*10^6$;
3) We need to manually filter out stop words (like "although", "a", ...), otherwise these frequent words will affect the performance of matrix factorization.
The neural network based model does not require storing huge hash tables of statistics on all of the corpus. It obtains the word embedding by learning from semantic information, hence could avoid the aforementioned problems in the traditional method. In this chapter, we will introduce the details of neural network word embedding model and how to train such model in PaddlePaddle.
In this section, after training the word embedding model, we could use the data visualization algorithm $t-$SNE\[[4](#reference)\] to draw the word embedding vectors after projecting them onto a two-dimensional space (see figure below). From the figure we could see that the semantically relevant words -- *a*, *the*, and *these* or *big* and *huge* -- are close to each other in the projected space, while irrelevant words -- *say* and *business* or *decision* and *japan* -- are far from each other.
On the other hand, we know that the cosine similarity between two vectors falls between $[-1,1]$. Specifically, the cosine similarity is 1 when the vectors are identical, 0 when the vectors are perpendicular, -1 when the are of opposite directions. That is, the cosine similarity between two vectors scales with their relevance. So we can calculate the cosine similarity of two word embedding vectors to represent their relevance:
```
```
please input two words: big huge
similarity: 0.899180685161
similarity: 0.899180685161
please input two words: big huge
please input two words: from company
please input two words: from company
similarity: -0.0997506977351
similarity: -0.0997506977351
```
```
The above results could be obtained by running `calculate_dis.py`, which loads the words in the dictionary and their corresponding trained word embeddings. For detailed instruction, see section [Model Application](#Model Application).
In this section, we will introduce three word embedding models: N-gram model, CBOW, and Skip-gram, which all output the frequency of each word given its immediate context.
## 模型概览
For N-gram model, we will first introduce the concept of language model, and implement it using PaddlePaddle in section [Model Training](#Model Training).
在这里我们介绍三个训练词向量的模型：N-gram模型，CBOW模型和Skip-gram模型，它们的中心思想都是通过上下文得到一个词出现的概率。对于N-gram模型，我们会先介绍语言模型的概念，并在之后的[训练模型](#训练模型)中，带大家用PaddlePaddle实现它。而后两个模型，是近年来最有名的神经元词向量模型，由 Tomas Mikolov 在Google 研发\[[3](#参考文献)\]，虽然它们很浅很简单，但训练效果很好。
The latter two models, which became popular recently, are neural word embedding model developed by Tomas Mikolov at Google \[[3](#reference)\]. Despite their apparent simplicity, these models train very well.
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。
Before diving into word embedding models, we will first introduce the concept of **language model**. Language models build the joint probability function $P(w_1, ..., w_T)$ of a sentence, where $w_i$ is the i-th word in the sentence. The goal is to give higher probabilities to meaningful sentences, and lower probabilities to meaningless constructions.
In general, models that generate the probability of a sequence can be applied to many fields, like machine translation, speech recognition, information retrieval, part-of-speech tagging, and handwriting recognition. Take information retrieval, for example. If you were to search for "how long is a football bame" (where bame is a medical noun), the search engine would have asked if you had meant "how long is a football game" instead. This is because the probability of "how long is a football bame" is very low according to the language model; in addition, among all of the words easily confused with "bame", "game" would build the most probable sentence.
#### Target Probability
For language model's target probability $P(w_1, ..., w_T)$, if the words in the sentence were to be independent, the joint probability of the whole sentence would be the product of each word's probability:
$$P(w_1, ..., w_T) = \prod_{t=1}^TP(w_t)$$
$$P(w_1, ..., w_T) = \prod_{t=1}^TP(w_t)$$
However, the frequency of words in a sentence typically relates to the words before them, so canonical language models are constructed using conditional probability in its target probability:
In computational linguistics, n-gram is an important method to represent text. An n-gram represents a contiguous sequence of n consecutive items given a text. Based on the desired application scenario, each item could be a letter, a syllable or a word. The N-gram model is also an important method in statistical language modeling. When training language models with n-grams, the first (n-1) words of an n-gram are used to predict the *n*th word.
Yoshua Bengio and other scientists describe how to train a word embedding model using neural network in the famous paper of Neural Probabilistic Language Models \[[1](#reference)\] published in 2003. The Neural Network Language Model (NNLM) described in the paper learns the language model and word embedding simultaneously through a linear transformation and a non-linear hidden connection. That is, after training on large amounts of corpus, the model learns the word embedding; then, it computes the probability of the whole sentence, using the embedding. This type of language model can overcome the **curse of dimensionality** i.e. model inaccuracy caused by the difference in dimensionality between training and testing data. Note that the term *neural network language model* is ill-defined, so we will not use the name NNLM but only refer to it as *N-gram neural model* in this section.
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。
We have previously described language model using conditional probability, where the probability of the *t*-th word in a sentence depends on all $t-1$ words before it. Furthermore, since words further prior have less impact on a word, and every word within an n-gram is only effected by its previous n-1 words, we have:
where $f(w_t, w_{t-1}, ..., w_{t-n+1})$ represents the conditional probability of the current word $w_t$ given its previous $n-1$ words, and $R(\theta)$ represents parameter regularization term.
- For each sample, the model gets input $w_{t-n+1},...w_{t-1}$, and outputs the probability that the t-th word is one of `|V|` in the dictionary.
Every input word $w_{t-n+1},...w_{t-1}$ first gets transformed into word embedding $C(w_{t-n+1}),...C(w_{t-1})$ through a transformation matrix.
- 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：
- All the word embeddings concatenate into a single vector, which is mapped (nonlinearly) into the $t$-th word hidden representation:
$$g=Utanh(\theta^Tx + b_1) + Wx + b_2$$
$$g=Utanh(\theta^Tx + b_1) + Wx + b_2$$
where $x$ is the large vector concatenated from all the word embeddings representing the context; $\theta$, $U$, $b_1$, $b_2$ and $W$ are parameters connecting word embedding layers to the hidden layers. $g$ represents the unnormalized probability of the output word, $g_i$ represents the unnormalized probability of the output word being the i-th word in the dictionary.
where $y_k^i$ represents the true label for the $k$-th class in the $i$-th sample ($0$ or $1$), $softmax(g_k^i)$ represents the softmax probability for the $k$-th class in the $i$-th sample.
CBOW model predicts the current word based on the N words both before and after it. When $N=2$, the model is as the figure below:
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：
<palign="center">
<palign="center">
<imgsrc="image/cbow_en.png"width=250><br/>
<imgsrc="image/cbow.png"width=250><br/>
Figure 3. CBOW model
图3. CBOW模型
</p>
</p>
Specifically, by ignoring the order of words in the sequence, CBOW uses the average value of the word embedding of the context to predict the current word:
where $x_t$ is the word embedding of the t-th word, classification score vector is $z=U*\text{context}$, the final classification $y$ uses softmax and the loss function uses multi-class cross-entropy.
The advantages of CBOW is that it smooths over the word embeddings of the context and reduces noise, so it is very effective on small dataset. Skip-gram uses a word to predict its context and get multiple context for the given word, so it can be used in larger datasets.
As illustrated in the figure above, skip-gram model maps the word embedding of the given word onto $2n$ word embeddings (including $n$ words before and $n$ words after the given word), and then combine the classification loss of all those $2n$ words by softmax.
We will use Peen Treebank (PTB) (Tomas Mikolov's pre-processed version) dataset. PTB is a small dataset, used in Recurrent Neural Network Language Modeling Toolkit\[[2](#reference)\]. Its statistics are as follows:
Beginning and end of a sentence have a special meaning, so we will add begin token `<s>` in the front of the sentence. And end token `<e>` in the end of the sentence. By moving the five word window in the sentence, data instances are generated.
- Hidden feature vector will go through another fully conected layer, turn into a $|V|$ dimensional vector. At the same time softmax will be applied to get the probability of each word being generated.
Next, we will begin the training process. `paddle.dataset.imikolov.train()` and `paddle.dataset.imikolov.test()` is our training set and test set. Both of the function will return a **reader**: In PaddlePaddle, reader is a python function which returns a Python iterator which output a single data instance at a time.
`paddle.batch` takes reader as input, outputs a **batched reader**: In PaddlePaddle, a reader outputs a single data instance at a time but batched reader outputs a minibatch of data instances.
Cosine similarity is one way of quantifying the similarity between two vectors. The range of result is $[-1, 1]$. The bigger the value, the similar two vectors are:
This chapter introduces word embedding, the relationship between language model and word embedding, and how to train neural networks to learn word embedding.
In information retrieval, the relevance between the query and document keyword can be computed through the cosine similarity of their word embeddings. In grammar analysis and semantic analysis, a previously trained word embedding can initialize models for better performance. In document classification, clustering the word embedding can group synonyms in the documents. We hope that readers can use word embedding models in their work after reading this chapter.
## Referenes
## 参考文献
1. Bengio Y, Ducharme R, Vincent P, et al. [A neural probabilistic language model](http://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf)[J]. journal of machine learning research, 2003, 3(Feb): 1137-1155.
1. Bengio Y, Ducharme R, Vincent P, et al. [A neural probabilistic language model](http://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf)[J]. journal of machine learning research, 2003, 3(Feb): 1137-1155.
2. Mikolov T, Kombrink S, Deoras A, et al. [Rnnlm-recurrent neural network language modeling toolkit](http://www.fit.vutbr.cz/~imikolov/rnnlm/rnnlm-demo.pdf)[C]//Proc. of the 2011 ASRU Workshop. 2011: 196-201.
2. Mikolov T, Kombrink S, Deoras A, et al. [Rnnlm-recurrent neural network language modeling toolkit](http://www.fit.vutbr.cz/~imikolov/rnnlm/rnnlm-demo.pdf)[C]//Proc. of the 2011 ASRU Workshop. 2011: 196-201.
3. Mikolov T, Chen K, Corrado G, et al. [Efficient estimation of word representations in vector space](https://arxiv.org/pdf/1301.3781.pdf)[J]. arXiv preprint arXiv:1301.3781, 2013.
3. Mikolov T, Chen K, Corrado G, et al. [Efficient estimation of word representations in vector space](https://arxiv.org/pdf/1301.3781.pdf)[J]. arXiv preprint arXiv:1301.3781, 2013.
...
@@ -458,7 +448,7 @@ In information retrieval, the relevance between the query and document keyword c
...
@@ -458,7 +448,7 @@ In information retrieval, the relevance between the query and document keyword c
This tutorial is contributed by <axmlns:cc="http://creativecommons.org/ns#"href="http://book.paddlepaddle.org"property="cc:attributionName"rel="cc:attributionURL">PaddlePaddle</a>, and licensed under a <arel="license"href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.
1. [Peter Brusilovsky](https://en.wikipedia.org/wiki/Peter_Brusilovsky) (2007). *The Adaptive Web*. p. 325.
2. Robin Burke , [Hybrid Web Recommender Systems](http://www.dcs.warwick.ac.uk/~acristea/courses/CS411/2010/Book%20-%20The%20Adaptive%20Web/HybridWebRecommenderSystems.pdf), pp. 377-408, The Adaptive Web, Peter Brusilovsky, Alfred Kobsa, Wolfgang Nejdl (Ed.), Lecture Notes in Computer Science, Springer-Verlag, Berlin, Germany, Lecture Notes in Computer Science, Vol. 4321, May 2007, 978-3-540-72078-2.
3. P. Resnick, N. Iacovou, etc. “[GroupLens: An Open Architecture for Collaborative Filtering of Netnews](http://ccs.mit.edu/papers/CCSWP165.html)”, Proceedings of ACM Conference on Computer Supported Cooperative Work, CSCW 1994. pp.175-186.
4. Sarwar, Badrul, et al. "[Item-based collaborative filtering recommendation algorithms.](http://files.grouplens.org/papers/www10_sarwar.pdf)" *Proceedings of the 10th international conference on World Wide Web*. ACM, 2001.
5. Kautz, Henry, Bart Selman, and Mehul Shah. "[Referral Web: combining social networks and collaborative filtering.](http://www.cs.cornell.edu/selman/papers/pdf/97.cacm.refweb.pdf)" Communications of the ACM 40.3 (1997): 63-65. APA
6. Yuan, Jianbo, et al. ["Solving Cold-Start Problem in Large-scale Recommendation Engines: A Deep Learning Approach."](https://arxiv.org/pdf/1611.05480v1.pdf) *arXiv preprint arXiv:1611.05480* (2016).
7. Covington P, Adams J, Sargin E. [Deep neural networks for youtube recommendations](https://static.googleusercontent.com/media/research.google.com/zh-CN//pubs/archive/45530.pdf)[C]//Proceedings of the 10th ACM Conference on Recommender Systems. ACM, 2016: 191-198.
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<divid="markdown"style='display:none'>
# Personalized Recommendation
The source code of this tutorial is in [book/recommender_system](https://github.com/PaddlePaddle/book/tree/develop/05.recommender_system).
For instructions on getting started with PaddlePaddle, see [PaddlePaddle installation guide](https://github.com/PaddlePaddle/book/blob/develop/README.en.md#running-the-book).
## Background
With the fast growth of e-commerce, online videos, and online reading business, users have to rely on recommender systems to avoid manually browsing tremendous volume of choices. Recommender systems understand users' interest by mining user behavior and other properties of users and products.
Some well know approaches include:
- User behavior-based approach. A well-known method is collaborative filtering. The underlying assumption is that if a person A has the same opinion as a person B on an issue, A is more likely to have B's opinion on a different issue than that of a randomly chosen person.
- Content-based recommendation[[1](#reference)]. This approach infers feature vectors that represent products from their descriptions. It also infers feature vectors that represent users' interests. Then it measures the relevance of users and products by some distances between these feature vectors.
- Hybrid approach[[2](#reference)]: This approach uses the content-based information to help address the cold start problem[[6](#reference)] in behavior-based approach.
Among these options, collaborative filtering might be the most studied one. Some of its variants include user-based[[3](#reference)], item-based [[4](#reference)], social network based[[5](#reference)], and model-based.
This tutorial explains a deep learning based approach and how to implement it using PaddlePaddle. We will train a model using a dataset that includes user information, movie information, and ratings. Once we train the model, we will be able to get a predicted rating given a pair of user and movie IDs.
## Model Overview
To know more about deep learning based recommendation, let us start from going over the Youtube recommender system[[7](#reference)] before introducing our hybrid model.
### YouTube's Deep Learning Recommendation Model
YouTube is a video-sharing Web site with one of the largest user base in the world. Its recommender system serves more than a billion users. This system is composed of two major parts: candidate generation and ranking. The former selects few hundreds of candidates from millions of videos, and the latter ranks and outputs the top 10.
Youtube models candidate generation as a multiclass classification problem with a huge number of classes equal to the number of videos. The architecture of the model is as follows:
The first stage of this model maps watching history and search queries into fixed-length representative features. Then, an MLP (multi-layer perceptron, as described in the [Recognize Digits](https://github.com/PaddlePaddle/book/blob/develop/recognize_digits/README.md) tutorial) takes the concatenation of all representative vectors. The output of the MLP represents the user' *intrinsic interests*. At training time, it is used together with a softmax output layer for minimizing the classification error. At serving time, it is used to compute the relevance of the user with all movies.
For a user $U$, the predicted watching probability of video $i$ is
where $u$ is the representative vector of user $U$, $V$ is the corpus of all videos, $v_i$ is the representative vector of the $i$-th video. $u$ and $v_i$ are vectors of the same length, so we can compute their dot product using a fully connected layer.
This model could have a performance issue as the softmax output covers millions of classification labels. To optimize performance, at the training time, the authors down-sample negative samples, so the actual number of classes is reduced to thousands. At serving time, the authors ignore the normalization of the softmax outputs, because the results are just for ranking.
#### Ranking Network
The architecture of the ranking network is similar to that of the candidate generation network. Similar to ranking models widely used in online advertising, it uses rich features like video ID, last watching time, etc. The output layer of the ranking network is a weighted logistic regression, which rates all candidate videos.
### Hybrid Model
In the section, let us introduce our movie recommendation system. Especially, we feed moives titles into a text convolution network to get a fixed-length representative feature vector. Accordingly we will introduce the convolutional neural network for texts and the hybrid recommendation model respectively.
#### Convolutional Neural Networks for Texts (CNN)
**Convolutional Neural Networks** are frequently applied to data with grid-like topology such as two-dimensional images and one-dimensional texts. A CNN can extract multiple local features, combine them, and produce high-level abstractions, which correspond to semantic understanding. Empirically, CNN is shown to be efficient for image and text modeling.
CNN mainly contains convolution and pooling operation, with versatile combinations in various applications. Here, we briefly describe a CNN as shown in Figure 3.
Let $n$ be the length of the sentence to process, and the $i$-th word has embedding as $x_i\in\mathbb{R}^k$，where $k$ is the embedding dimensionality.
First, we concatenate the words by piecing together every $h$ words, each as a window of length $h$. This window is denoted as $x_{i:i+h-1}$, consisting of $x_{i},x_{i+1},\ldots,x_{i+h-1}$, where $x_i$ is the first word in the window and $i$ takes value ranging from $1$ to $n-h+1$: $x_{i:i+h-1}\in\mathbb{R}^{hk}$.
Next, we apply the convolution operation: we apply the kernel $w\in\mathbb{R}^{hk}$ in each window, extracting features $c_i=f(w\cdot x_{i:i+h-1}+b)$, where $b\in\mathbb{R}$ is the bias and $f$ is a non-linear activation function such as $sigmoid$. Convolving by the kernel at every window ${x_{1:h},x_{2:h+1},\ldots,x_{n-h+1:n}}$ produces a feature map in the following form:
$$c=[c_1,c_2,\ldots,c_{n-h+1}], c \in \mathbb{R}^{n-h+1}$$
Next, we apply *max pooling* over time to represent the whole sentence $\hat c$, which is the maximum element across the feature map:
$$\hat c=max(c)$$
#### Model Structure Of The Hybrid Model
In our network, the input includes features of users and movies. The user feature includes four properties: user ID, gender, occupation, and age. Movie features include their IDs, genres, and titles.
We use fully-connected layers to map user features into representative feature vectors and concatenate them. The process of movie features is similar, except that for movie titles -- we feed titles into a text convolution network as described in the above section to get a fixed-length representative feature vector.
Given the feature vectors of users and movies, we compute the relevance using cosine similarity. We minimize the squared error at training time.
We use the [MovieLens ml-1m](http://files.grouplens.org/datasets/movielens/ml-1m.zip) to train our model. This dataset includes 10,000 ratings of 4,000 movies from 6,000 users to 4,000 movies. Each rate is in the range of 1~5. Thanks to GroupLens Research for collecting, processing and publishing the dataset.
`paddle.v2.datasets` package encapsulates multiple public datasets, including `cifar`, `imdb`, `mnist`, `moivelens` and `wmt14`, etc. There's no need for us to manually download and preprocess `MovieLens` dataset.
The raw `MoiveLens` contains movie ratings, relevant features from both movies and users.
print "User %s rates Movie %s with Score %s"%(user_info[uid], movie_info[mov_id], train_sample[-1])
```
```text
User <UserInfoid(1),gender(F),age(1),job(10)> rates Movie <MovieInfoid(1193),title(OneFlewOvertheCuckoo'sNest),categories(['Drama'])> with Score [5.0]
```
The output shows that user 1 gave movie `1193` a rating of 5.
After issuing a command `python train.py`, training will start immediately. The details will be unpacked by the following sessions to see how it works.
## Model Architecture
### Initialize PaddlePaddle
First, we must import and initialize PaddlePaddle (enable/disable GPU, set the number of trainers, etc).
As shown in the above code, the input is four dimension integers for each user, that is, `user_id`,`gender_id`, `age_id` and `job_id`. In order to deal with these features conveniently, we use the language model in NLP to transform these discrete values into embedding vaules `usr_emb`, `usr_gender_emb`, `usr_age_emb` and `usr_job_emb`.
Movie title, a sequence of words represented by an integer word index sequence, will be feed into a `sequence_conv_pool` layer, which will apply convolution and pooling on time dimension. Because pooling is done on time dimension, the output will be a fixed-length vector regardless the length of the input sequence.
Finally, we can use cosine similarity to calculate the similarity between user characteristics and movie features.
First, we define the model parameters according to the previous model configuration `cost`.
```python
# Create parameters
parameters = paddle.parameters.create(cost)
```
### Create Trainer
Before jumping into creating a training module, algorithm setting is also necessary. Here we specified Adam optimization algorithm via `paddle.optimizer`.
[INFO 2017-03-06 17:12:13,378 networks.py:1472] The input order is [user_id, gender_id, age_id, job_id, movie_id, category_id, movie_title, score]
[INFO 2017-03-06 17:12:13,379 networks.py:1478] The output order is [__mse_cost_0__]
```
### Training
`paddle.dataset.movielens.train` will yield records during each pass, after shuffling, a batch input is generated for training.
```python
reader=paddle.batch(
paddle.reader.shuffle(
paddle.dataset.movielens.train(), buf_size=8192),
batch_size=256)
```
`feeding` is devoted to specifying the correspondence between each yield record and `paddle.layer.data`. For instance, the first column of data generated by `movielens.train` corresponds to `user_id` feature.
```python
feeding = {
'user_id': 0,
'gender_id': 1,
'age_id': 2,
'job_id': 3,
'movie_id': 4,
'category_id': 5,
'movie_title': 6,
'score': 7
}
```
Callback function `event_handler` and `event_handler_plot` will be called during training when a pre-defined event happens.
```python
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0:
print "Pass %d Batch %d Cost %.2f" % (
event.pass_id, event.batch_id, event.cost)
```
```python
from paddle.v2.plot import Ploter
train_title = "Train cost"
test_title = "Test cost"
cost_ploter = Ploter(train_title, test_title)
step = 0
def event_handler_plot(event):
global step
if isinstance(event, paddle.event.EndIteration):
if step % 10 == 0: # every 10 batches, record a train cost
cost_ploter.append(train_title, step, event.cost)
if step % 1000 == 0: # every 1000 batches, record a test cost
result = trainer.test(
reader=paddle.batch(
paddle.dataset.movielens.test(), batch_size=256),
feeding=feeding)
cost_ploter.append(test_title, step, result.cost)
if step % 100 == 0: # every 100 batches, update cost plot
cost_ploter.plot()
step += 1
```
Finally, we can invoke `trainer.train` to start training:
```python
trainer.train(
reader=reader,
event_handler=event_handler_plot,
feeding=feeding,
num_passes=2)
```
## Conclusion
This tutorial goes over traditional approaches in recommender system and a deep learning based approach. We also show that how to train and use the model with PaddlePaddle. Deep learning has been well used in computer vision and NLP, we look forward to its new successes in recommender systems.
## Reference
1. [Peter Brusilovsky](https://en.wikipedia.org/wiki/Peter_Brusilovsky) (2007). *The Adaptive Web*. p. 325.
2. Robin Burke , [Hybrid Web Recommender Systems](http://www.dcs.warwick.ac.uk/~acristea/courses/CS411/2010/Book%20-%20The%20Adaptive%20Web/HybridWebRecommenderSystems.pdf), pp. 377-408, The Adaptive Web, Peter Brusilovsky, Alfred Kobsa, Wolfgang Nejdl (Ed.), Lecture Notes in Computer Science, Springer-Verlag, Berlin, Germany, Lecture Notes in Computer Science, Vol. 4321, May 2007, 978-3-540-72078-2.
3. P. Resnick, N. Iacovou, etc. “[GroupLens: An Open Architecture for Collaborative Filtering of Netnews](http://ccs.mit.edu/papers/CCSWP165.html)”, Proceedings of ACM Conference on Computer Supported Cooperative Work, CSCW 1994. pp.175-186.
4. Sarwar, Badrul, et al. "[Item-based collaborative filtering recommendation algorithms.](http://files.grouplens.org/papers/www10_sarwar.pdf)" *Proceedings of the 10th International Conference on World Wide Web*. ACM, 2001.
5. Kautz, Henry, Bart Selman, and Mehul Shah. "[Referral Web: Combining Social networks and collaborative filtering.](http://www.cs.cornell.edu/selman/papers/pdf/97.cacm.refweb.pdf)" Communications of the ACM 40.3 (1997): 63-65. APA
6. Yuan, Jianbo, et al. ["Solving Cold-Start Problem in Large-scale Recommendation Engines: A Deep Learning Approach."](https://arxiv.org/pdf/1611.05480v1.pdf) *arXiv preprint arXiv:1611.05480* (2016).
7. Covington P, Adams J, Sargin E. [Deep neural networks for youtube recommendations](https://static.googleusercontent.com/media/research.google.com/zh-CN//pubs/archive/45530.pdf)[C]//Proceedings of the 10th ACM Conference on Recommender Systems. ACM, 2016: 191-198.
<br/>
This tutorial is contributed by <axmlns:cc="http://creativecommons.org/ns#"href="http://book.paddlepaddle.org"property="cc:attributionName"rel="cc:attributionURL">PaddlePaddle</a>, and licensed under a <arel="license"href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.
对卷积神经网络来说，首先使用卷积处理输入的词向量序列，产生一个特征图（feature map），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，最后，将所有卷积核得到的特征拼接起来即为文本的定长向量表示，对于文本分类问题，将其连接至softmax即构建出完整的模型。在实际应用中，我们会使用多个卷积核来处理句子，窗口大小相同的卷积核堆叠起来形成一个矩阵，这样可以更高效的完成运算。另外，我们也可使用窗口大小不同的卷积核来处理句子，[推荐系统](https://github.com/PaddlePaddle/book/tree/develop/05.recommender_system)一节的图3作为示意画了四个卷积核，不同颜色表示不同大小的卷积核操作。
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。
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\[[5](#参考文献)\])。
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：**
1. Kim Y. [Convolutional neural networks for sentence classification](http://arxiv.org/pdf/1408.5882)[J]. arXiv preprint arXiv:1408.5882, 2014.
2. Kalchbrenner N, Grefenstette E, Blunsom P. [A convolutional neural network for modelling sentences](http://arxiv.org/pdf/1404.2188.pdf?utm_medium=App.net&utm_source=PourOver)[J]. arXiv preprint arXiv:1404.2188, 2014.
3. Yann N. Dauphin, et al. [Language Modeling with Gated Convolutional Networks](https://arxiv.org/pdf/1612.08083v1.pdf)[J] arXiv preprint arXiv:1612.08083, 2016.
4. Siegelmann H T, Sontag E D. [On the computational power of neural nets](http://research.cs.queensu.ca/home/akl/cisc879/papers/SELECTED_PAPERS_FROM_VARIOUS_SOURCES/05070215382317071.pdf)[C]//Proceedings of the fifth annual workshop on Computational learning theory. ACM, 1992: 440-449.
6. Bengio Y, Simard P, Frasconi P. [Learning long-term dependencies with gradient descent is difficult](http://www-dsi.ing.unifi.it/~paolo/ps/tnn-94-gradient.pdf)[J]. IEEE transactions on neural networks, 1994, 5(2): 157-166.
7. Graves A. [Generating sequences with recurrent neural networks](http://arxiv.org/pdf/1308.0850)[J]. arXiv preprint arXiv:1308.0850, 2013.
8. Cho K, Van Merriënboer B, Gulcehre C, et al. [Learning phrase representations using RNN encoder-decoder for statistical machine translation](http://arxiv.org/pdf/1406.1078)[J]. arXiv preprint arXiv:1406.1078, 2014.
9. Zhou J, Xu W. [End-to-end learning of semantic role labeling using recurrent neural networks](http://www.aclweb.org/anthology/P/P15/P15-1109.pdf)[C]//Proceedings of the Annual Meeting of the Association for Computational Linguistics. 2015.
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<divid="markdown"style='display:none'>
# Sentiment Analysis
The source codes of this section can be located at [book/understand_sentiment](https://github.com/PaddlePaddle/book/tree/develop/06.understand_sentiment). First-time users may refer to PaddlePaddle for [Installation guide](https://github.com/PaddlePaddle/book/blob/develop/README.en.md#running-the-book).
## Background
In natural language processing, sentiment analysis refers to determining the emotion expressed in a piece of text. The text can be a sentence, a paragraph, or a document. Emotion categorization can be binary -- positive/negative or happy/sad -- or in three classes -- positive/neutral/negative. Sentiment analysis is applicable in a wide range of services, such as e-commerce sites like Amazon and Taobao, hospitality services like Airbnb and hotels.com, and movie rating sites like Rotten Tomatoes and IMDB. It can be used to gauge from the reviews how the customers feel about the product. Table 1 illustrates an example of sentiment analysis in movie reviews:
| Movie Review | Category |
| -------- | ----- |
| Best movie of Xiaogang Feng in recent years!| Positive |
| Pretty bad. Feels like a tv-series from a local TV-channel | Negative |
| Politically correct version of Taken ... and boring as Heck| Negative|
|delightful, mesmerizing, and completely unexpected. The plot is nicely designed.|Positive|
<palign="center">Table 1 Sentiment Analysis in Movie Reviews</p>
In natural language processing, sentiment analysis can be categorized as a **Text Classification problem**, i.e., to categorize a piece of text to a specific class. It involves two related tasks: text representation and classification. Before the emergence of deep learning techniques, the mainstream methods for text representation include BOW (*bag of words*) and topic modeling, while the latter contain SVM (*support vector machine*) and LR (*logistic regression*).
The BOW model does not capture all the information in a piece of text, as it ignores syntax and grammar and just treats the text as a set of words. For example, “this movie is extremely bad“ and “boring, dull, and empty work” describe very similar semantic meaning, yet their BOW representations have with little similarity. Furthermore, “the movie is bad“ and “the movie is not bad“ have high similarity with BOW features, but they express completely opposite semantics.
This chapter introduces a deep learning model that handles these issues in BOW. Our model embeds texts into a low-dimensional space and takes word order into consideration. It is an end-to-end framework and it has large performance improvement over traditional methods \[[1](#Reference)\].
## Model Overview
The model we used in this chapter uses **Convolutional Neural Networks** (**CNNs**) and **Recurrent Neural Networks** (**RNNs**) with some specific extensions.
### Revisit to the Convolutional Neural Networks for Texts (CNN)
The convolutional neural network for texts is introduced in chapter [recommender_system](https://github.com/PaddlePaddle/book/tree/develop/05.recommender_system), here we make a brief overview.
CNN mainly contains convolution and pooling operation, with versatile combinations in various applications. We first apply the convolution operation: we apply the kernel in each window, extracting features. Convolving by the kernel at every window produces a feature map. Next, we apply *max pooling* over time to represent the whole sentence, which is the maximum element across the feature map. In real applications, we will apply multiple CNN kernels on the sentences. It can be implemented efficiently by concatenating the kernels together as a matrix. Also, we can use CNN kernels with different kernel size. Finally, concatenating the resulting features produces a fixed-length representation, which can be combined with a softmax to form the model for the sentiment analysis problem.
For short texts, the aforementioned CNN model can achieve very high accuracy \[[1](#Reference)\]. If we want to extract more abstract representations, we may apply a deeper CNN model \[[2](#Reference),[3](#Reference)\].
### Recurrent Neural Network (RNN)
RNN is an effective model for sequential data. In terms of computability, the RNN is Turing-complete \[[4](#Reference)\]. Since NLP is a classical problem on sequential data, the RNN, especially its variant LSTM\[[5](#Reference)\]), achieves state-of-the-art performance on various NLP tasks, such as language modeling, syntax parsing, POS-tagging, image captioning, dialog, machine translation, and so forth.
Figure 1. An illustration of an unfolded RNN in time.
</p>
As shown in Figure 1, we unfold an RNN: at the $t$-th time step, the network takes two inputs: the $t$-th input vector $\vec{x_t}$ and the latent state from the last time-step $\vec{h_{t-1}}$. From those, it computes the latent state of the current step $\vec{h_t}$. This process is repeated until all inputs are consumed. Denoting the RNN as function $f$, it can be formulated as follows:
where $W_{xh}$ is the weight matrix to feed into the latent layer; $W_{hh}$ is the latent-to-latent matrix; $b_h$ is the latent bias and $\sigma$ refers to the $sigmoid$ function.
In NLP, words are often represented as a one-hot vectors and then mapped to an embedding. The embedded feature goes through an RNN as input $x_t$ at every time step. Moreover, we can add other layers on top of RNN, such as a deep or stacked RNN. Finally, the last latent state may be used as a feature for sentence classification.
### Long-Short Term Memory (LSTM)
Training an RNN on long sequential data sometimes leads to the gradient vanishing or exploding\[[6](#)\]. To solve this problem Hochreiter S, Schmidhuber J. (1997) proposed **Long Short Term Memory** (LSTM)\[[5](#Reference)\]).
Compared to the structure of a simple RNN, an LSTM includes memory cell $c$, input gate $i$, forget gate $f$ and output gate $o$. These gates and memory cells dramatically improve the ability for the network to handle long sequences. We can formulate the **LSTM-RNN**, denoted as a function $F$, as follows：
$$ h_t=F(x_t,h_{t-1})$$
$F$ contains following formulations\[[7](#Reference)\]：
In the equation，$i_t, f_t, c_t, o_t$ stand for input gate, forget gate, memory cell and output gate, respectively. $W$ and $b$ are model parameters, $\tanh$ is a hyperbolic tangent, and $\odot$ denotes an element-wise product operation. The input gate controls the magnitude of the new input into the memory cell $c$; the forget gate controls the memory propagated from the last time step; the output gate controls the magnitutde of the output. The three gates are computed similarly with different parameters, and they influence memory cell $c$ separately, as shown in Figure 2:
LSTM enhances the ability of considering long-term reliance, with the help of memory cell and gate. Similar structures are also proposed in Gated Recurrent Unit (GRU)\[[8](Reference)\] with simpler design. **The structures are still similar to RNN, though with some modifications (As shown in Figure 2), i.e., latent status depends on input as well as the latent status of last time-step, and the process goes on recurrently until all input are consumed:**
$$ h_t=Recrurent(x_t,h_{t-1})$$
where $Recrurent$ is a simple RNN, GRU or LSTM.
### Stacked Bidirectional LSTM
For vanilla LSTM, $h_t$ contains input information from previous time-step $1..t-1$ context. We can also apply an RNN with reverse-direction to take successive context $t+1…n$ into consideration. Combining constructing deep RNN (deeper RNN can contain more abstract and higher level semantic), we can design structures with deep stacked bidirectional LSTM to model sequential data\[[9](#Reference)\].
As shown in Figure 3 (3-layer RNN), odd/even layers are forward/reverse LSTM. Higher layers of LSTM take lower-layers LSTM as input, and the top-layer LSTM produces a fixed length vector by max-pooling (this representation considers contexts from previous and successive words for higher-level abstractions). Finally, we concatenate the output to a softmax layer for classification.
Figure 3. Stacked Bidirectional LSTM for NLP modeling.
</p>
## Dataset
We use [IMDB](http://ai.stanford.edu/%7Eamaas/data/sentiment/) dataset for sentiment analysis in this tutorial, which consists of 50,000 movie reviews split evenly into 25k train and 25k test sets. In the labeled train/test sets, a negative review has a score <=4outof10,andapositivereviewhasascore>= 7 out of 10.
`paddle.datasets` package encapsulates multiple public datasets, including `cifar`, `imdb`, `mnist`, `moivelens`, and `wmt14`, etc. There's no need for us to manually download and preprocess IMDB.
After issuing a command `python train.py`, training will start immediately. The details will be unpacked by the following sessions to see how it works.
## Model Structure
### Initialize PaddlePaddle
We must import and initialize PaddlePaddle (enable/disable GPU, set the number of trainers, etc).
```python
import sys
import paddle.v2 as paddle
# PaddlePaddle init
paddle.init(use_gpu=False, trainer_count=1)
```
As alluded to in section [Model Overview](#model-overview), here we provide the implementations of both Text CNN and Stacked-bidirectional LSTM models.
### Text Convolution Neural Network (Text CNN)
We create a neural network `convolution_net` as the following snippet code.
Note: `paddle.networks.sequence_conv_pool` includes both convolution and pooling layer operations.
Parameter `input_dim` denotes the dictionary size, and `class_dim` is the number of categories. In `convolution_net`, the input to the network is defined in `paddle.layer.data`.
1. Define Classifier
The above Text CNN network extracts high-level features and maps them to a vector of the same size as the categories. `paddle.activation.Softmax` function or classifier is then used for calculating the probability of the sentence belonging to each category.
1. Define Loss Function
In the context of supervised learning, labels of the training set are defined in `paddle.layer.data`, too. During training, cross-entropy is used as loss function in `paddle.layer.classification_cost` and as the output of the network; During testing, the outputs are the probabilities calculated in the classifier.
#### Stacked bidirectional LSTM
We create a neural network `stacked_lstm_net` as below.
```python
def stacked_lstm_net(input_dim,
class_dim=2,
emb_dim=128,
hid_dim=512,
stacked_num=3):
"""
A Wrapper for sentiment classification task.
This network uses bi-directional recurrent network,
consisting three LSTM layers. This configure is referred to
Parameter `input_dim` denotes the dictionary size, and `class_dim` is the number of categories. In `stacked_lstm_net`, the input to the network is defined in `paddle.layer.data`.
1. Define Classifier
The above stacked bidirectional LSTM network extracts high-level features and maps them to a vector of the same size as the categories. `paddle.activation.Softmax` function or classifier is then used for calculating the probability of the sentence belonging to each category.
1. Define Loss Function
In the context of supervised learning, labels of the training set are defined in `paddle.layer.data`, too. During training, cross-entropy is used as loss function in `paddle.layer.classification_cost` and as the output of the network; During testing, the outputs are the probabilities calculated in the classifier.
To reiterate, we can either invoke `convolution_net` or `stacked_lstm_net`.
`feeding` is devoted to specifying the correspondence between each yield record and `paddle.layer.data`. For instance, the first column of data generated by `paddle.dataset.imdb.train()` corresponds to `word` feature.
```python
feeding = {'word': 0, 'label': 1}
```
Callback function `event_handler` will be invoked to track training progress when a pre-defined event happens.
result = trainer.test(reader=test_reader, feeding=feeding)
print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics)
```
Finally, we can invoke `trainer.train` to start training:
```python
trainer.train(
reader=train_reader,
event_handler=event_handler,
feeding=feeding,
num_passes=10)
```
## Conclusion
In this chapter, we use sentiment analysis as an example to introduce applying deep learning models on end-to-end short text classification, as well as how to use PaddlePaddle to implement the model. Meanwhile, we briefly introduce two models for text processing: CNN and RNN. In following chapters, we will see how these models can be applied in other tasks.
## Reference
1. Kim Y. [Convolutional neural networks for sentence classification](http://arxiv.org/pdf/1408.5882)[J]. arXiv preprint arXiv:1408.5882, 2014.
2. Kalchbrenner N, Grefenstette E, Blunsom P. [A convolutional neural network for modelling sentences](http://arxiv.org/pdf/1404.2188.pdf?utm_medium=App.net&utm_source=PourOver)[J]. arXiv preprint arXiv:1404.2188, 2014.
3. Yann N. Dauphin, et al. [Language Modeling with Gated Convolutional Networks](https://arxiv.org/pdf/1612.08083v1.pdf)[J] arXiv preprint arXiv:1612.08083, 2016.
4. Siegelmann H T, Sontag E D. [On the computational power of neural nets](http://research.cs.queensu.ca/home/akl/cisc879/papers/SELECTED_PAPERS_FROM_VARIOUS_SOURCES/05070215382317071.pdf)[C]//Proceedings of the fifth annual workshop on Computational learning theory. ACM, 1992: 440-449.
6. Bengio Y, Simard P, Frasconi P. [Learning long-term dependencies with gradient descent is difficult](http://www-dsi.ing.unifi.it/~paolo/ps/tnn-94-gradient.pdf)[J]. IEEE transactions on neural networks, 1994, 5(2): 157-166.
7. Graves A. [Generating sequences with recurrent neural networks](http://arxiv.org/pdf/1308.0850)[J]. arXiv preprint arXiv:1308.0850, 2013.
8. Cho K, Van Merriënboer B, Gulcehre C, et al. [Learning phrase representations using RNN encoder-decoder for statistical machine translation](http://arxiv.org/pdf/1406.1078)[J]. arXiv preprint arXiv:1406.1078, 2014.
9. Zhou J, Xu W. [End-to-end learning of semantic role labeling using recurrent neural networks](http://www.aclweb.org/anthology/P/P15/P15-1109.pdf)[C]//Proceedings of the Annual Meeting of the Association for Computational Linguistics. 2015.
<br/>
This tutorial is contributed by <axmlns:cc="http://creativecommons.org/ns#"href="http://book.paddlepaddle.org"property="cc:attributionName"rel="cc:attributionURL">PaddlePaddle</a>, and licensed under a <arel="license"href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.
For instructions on getting started with PaddlePaddle, see [PaddlePaddle installation guide](https://github.com/PaddlePaddle/book/blob/develop/README.en.md#running-the-book).
Natural language analysis techniques consist of lexical, syntactic, and semantic analysis. **Semantic Role Labeling (SRL)** is an instance of **Shallow Semantic Analysis**.
In a sentence, a **predicate** states a property or a characterization of a *subject*, such as what it does and what it is like. The predicate represents the core of an event, whereas the words accompanying the predicate are **arguments**. A **semantic role** refers to the abstract role an argument of a predicate take on in the event, including *agent*, *patient*, *theme*, *experiencer*, *beneficiary*, *instrument*, *location*, *goal*, and *source*.
In the following example of a Chinese sentence, "to encounter" is the predicate (*pred*); "Ming" is the *agent*; "Hong" is the *patient*; "yesterday" and "evening" are the *time*; finally, "the park" is the *location*.
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。
$$\mbox{[小明 Ming]}_{\mbox{Agent}}\mbox{[昨天 yesterday]}_{\mbox{Time}}\mbox{[晚上 evening]}_\mbox{Time}\mbox{在[公园 a park]}_{\mbox{Location}}\mbox{[遇到 to encounter]}_{\mbox{Predicate}}\mbox{了[小红 Hong]}_{\mbox{Patient}}\mbox{。}$$
传统的SRL系统大多建立在句法分析基础之上，通常包括5个流程：
Instead of analyzing the semantic information, **Semantic Role Labeling** (**SRL**) identifies the relation between the predicate and the other constituents surrounding it. The predicate-argument structures are labeled as specific semantic roles. A wide range of natural language understanding tasks, including *information extraction*, *discourse analysis*, and *deepQA*. Research usually assumes a predicate of a sentence to be specified; the only task is to identify its arguments and their semantic roles.
Conventional SRL systems mostly build on top of syntactic analysis, usually consisting of five steps:
1. Construct a syntax tree, as shown in Fig. 1
2. Identity the candidate arguments of the given predicate on the tree.
3. Prune the most unlikely candidate arguments.
4. Identify the real arguments, often by a binary classifier.
5. Multi-classify on results from step 4 to label the semantic roles. Steps 2 and 3 usually introduce hand-designed features based on syntactic analysis (step 1).
However, a complete syntactic analysis requires identifying the relation among all constituents. Thus, the accuracy of SRL is sensitive to the preciseness of the syntactic analysis, making SRL challenging. To reduce its complexity and obtain some information on the syntactic structures, we often use *shallow syntactic analysis* a.k.a. partial parsing or chunking. Unlike complete syntactic analysis, which requires the construction of the complete parsing tree, *Shallow Syntactic Analysis* only requires identifying some independent constituents with relatively simple structures, such as verb phrases (chunk). To avoid difficulties in constructing a syntax tree with high accuracy, some work\[[1](#Reference)\] proposed semantic chunking-based SRL methods, which reduces SRL into a sequence tagging problem. Sequence tagging tasks classify syntactic chunks using **BIO representation**. For syntactic chunks forming role A, its first chunk receives the B-A tag (Begin) and the remaining ones receive the tag I-A (Inside); in the end, the chunks left out receive the tag O.
我们继续以上面的这句话为例，图1展示了BIO表示方法。
The BIO representation of above example is shown in Fig.1.
1. It only relies on shallow syntactic analysis, reduces the precision requirement of syntactic analysis;
2. Pruning the candidate arguments is no longer necessary;
3. Arguments are identified and tagged at the same time. Simplifying the workflow reduces the risk of accumulating errors; oftentimes, methods that unify multiple steps boost performance.
In this tutorial, our SRL system is built as an end-to-end system via a neural network. The system takes only text sequences as input, without using any syntactic parsing results or complex hand-designed features. The public dataset [CoNLL-2004 and CoNLL-2005 Shared Tasks](http://www.cs.upc.edu/~srlconll/) is used for the following task: given a sentence with predicates marked, identify the corresponding arguments and their semantic roles through sequence tagging.
## Model
与基于语块的SRL方法类似，在本教程中我们也将SRL看作一个序列标注问题，不同的是，我们只依赖输入文本序列，不依赖任何额外的语法解析结果或是复杂的人造特征，利用深度神经网络构建一个端到端学习的SRL系统。我们以[CoNLL-2004 and CoNLL-2005 Shared Tasks](http://www.cs.upc.edu/~srlconll/)任务中SRL任务的公开数据集为例，实践下面的任务：给定一句话和这句话里的一个谓词，通过序列标注的方式，从句子中找到谓词对应的论元，同时标注它们的语义角色。
**Recurrent Neural Networks** (*RNN*) are important tools for sequence modeling and have been successfully used in some natural language processing tasks. Unlike feed-forward neural networks, RNNs can model the dependencies between elements of sequences. As a variant of RNNs', LSTMs aim model long-term dependency in long sequences. We have introduced this in [understand_sentiment](https://github.com/PaddlePaddle/book/tree/develop/05.understand_sentiment). In this chapter, we continue to use LSTMs to solve SRL problems.
*Deep Neural Networks* can extract hierarchical representations. The higher layers can form relatively abstract/complex representations, based on primitive features discovered through the lower layers. Unfolding LSTMs through time results in a deep feed-forward neural network. This is because any computational path between the input at time $k <t$totheoutputattime$t$crossesseveralnonlinearlayers.Ontheotherhand,duetoparametersharingovertime,LSTMsarealso*shallow*;thatis,thecomputationcarriedoutateachtime-stepisjustalineartransformation.DeepLSTMnetworksaretypicallyconstructedbystackingmultipleLSTMlayersontopofeachotherandtakingtheoutputfromlowerLSTMlayerattime$t$astheinputofupperLSTMlayerattime$t$.Deep,hierarchicalneuralnetworkscanbeefficientatrepresentingsomefunctionsandmodelingvarying-lengthdependencies\[[2](#Reference)\].
While LSTMs can summarize the history -- all the previous input seen up until now -- they can not see the future. Because most NLP (natural language processing) tasks provide the entirety of sentences, sequential learning can benefit from having the future encoded as well as the history.
To address, we can design a bidirectional recurrent neural network by making a minor modification. A higher LSTM layer can process the sequence in reversed direction with regards to its immediate lower LSTM layer, i.e., deep LSTM layers take turns to train on input sequences from left-to-right and right-to-left. Therefore, LSTM layers at time-step $t$ can see both histories and the future, starting from the second layer. Fig. 4 illustrates the bidirectional recurrent neural networks.
Note that, this bidirectional RNNs is different with the one proposed by Bengio et al. in machine translation tasks \[[3](#Reference), [4](#Reference)\]. We will introduce another bidirectional RNNs in the following tasks [machine translation](https://github.com/PaddlePaddle/book/blob/develop/machine_translation/README.en.md)
Typically, a neural network's lower layers learn representations while its very top layer learns the final task. These principles can guide our problem-solving approaches. In SRL tasks, a **Conditional Random Field** (*CRF*) is built on top of the network in order to perform the final prediction to tag sequences. It takes as input the representations provided by the last LSTM layer.
### 条件随机场 (Conditional Random Field)
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。
The CRF is an undirected probabilistic graph with nodes denoting random variables and edges denoting dependencies between these variables. In essence, CRFs learn the conditional probability $P(Y|X)$, where $X = (x_1, x_2, ... , x_n)$ are sequences of input and $Y = (y_1, y_2, ... , y_n)$ are label sequences; to decode, simply search through $Y$ for a sequence that maximizes the conditional probability $P(Y|X)$, i.e., $Y^* = \mbox{arg max}_{Y} P(Y | X)$。
Sequence tagging tasks do not assume a lot of conditional independence, because they are only concerned with the input and the output being linear sequences. Thus, the graph model of sequence tagging tasks is usually a simple chain or line, which results in a **Linear-Chain Conditional Random Field**, shown in Fig.5.
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）。
where, $Z(X)$ is normalization constant, ${t_j}$ represents the feature functions defined on edges called the *transition feature*, which denotes the transition probabilities from $y_{i-1}$ to $y_i$ given input sequence $X$. ${s_k}$ represents the feature function defined on nodes, called the state feature, denoting the probability of $y_i$ given input sequence $X$. In addition, $\lambda_j$ and $\mu_k$ are weights corresponding to $t_j$ and $s_k$. Alternatively, $t$ and $s$ can be written in the same form that depends on $y_{i - 1}$, $y_i$, $X$, and $i$. Taking its summation over all nodes $i$, we have: $f_{k}(Y, X) = \sum_{i=1}^{n}f_k({y_{i - 1}, y_i, X, i})$, which defines the *feature function* $f$. Thus, $P(Y|X)$ can be written as:
where $\omega$ are the weights to the feature function that the CRF learns. While training, given input sequences and label sequences $D = \left[(X_1, Y_1), (X_2 , Y_2) , ... , (X_N, Y_N)\right]$, by maximum likelihood estimation (**MLE**), we construct the following objective function:
This objective function can be solved via back-propagation in an end-to-end manner. While decoding, given input sequences $X$, search for sequence $\bar{Y}$ to maximize the conditional probability $\bar{P}(Y|X)$ via decoding methods (such as *Viterbi*, or [Beam Search Algorithm](https://github.com/PaddlePaddle/book/blob/develop/07.machine_translation/README.en.md#Beam%20Search%20Algorithm)).
### 深度双向LSTM（DB-LSTM）SRL模型
### Deep Bidirectional LSTM (DB-LSTM) SRL model
Given predicates and a sentence, SRL tasks aim to identify arguments of the given predicate and their semantic roles. If a sequence has $n$ predicates, we will process this sequence $n$ times. Here is the breakdown of a straight-forward model:
- expand input 1 into a sequence of the same length with input 2's sentence, using one-hot representation;
2. Convert the one-hot sequences from step 1 to vector sequences via a word embedding's lookup table;
3. Learn the representation of input sequences by taking vector sequences from step 2 as inputs;
4. Take the representation from step 3 as input, label sequence as supervisory signal, and realize sequence tagging tasks.
Here, we propose some improvements by introducing two simple but effective features:
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；
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；
- predicate context (**ctx-p**): A single predicate word may not describe all the predicate information, especially when the same words appear multiple times in a sentence. With the expanded context, the ambiguity can be largely eliminated. Thus, we extract $n$ words before and after predicate to construct a window chunk.
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：
- region mark ($m_r$): The binary marker on a word, $m_r$, takes the value of $1$ when the word is in the predicate context region, and $0$ if not.
After these modifications, the model is as follows, as illustrated in Figure 6:
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. Construct inputs
- Input 1: word sequence. Input 2: predicate. Input 3: predicate context, extract $n$ words before and after predicate. Input 4: region mark sequence, where an entry is 1 if word is located in the predicate context region, 0 otherwise.
- expand input 2~3 into sequences with the same length with input 1
2. Convert input 1~4 to vector sequences via word embedding lookup tables; While input 1 and 3 shares the same lookup table, input 2 and 4 have separate lookup tables.
3. Take the four vector sequences from step 2 as inputs to bidirectional LSTMs; Train the LSTMs to update representations.
4. Take the representation from step 3 as input to CRF, label sequence as supervisory signal, and complete sequence tagging tasks.
In the tutorial, we use [CoNLL 2005](http://www.cs.upc.edu/~srlconll/) SRL task open dataset as an example. Note that the training set and development set of the CoNLL 2005 SRL task are not free to download after the competition. Currently, only the test set can be obtained, including 23 sections of the Wall Street Journal and three sections of the Brown corpus. In this tutorial, we use the WSJ corpus as the training dataset to explain the model. However, since the training set is small, for a usable neural network SRL system, please consider paying for the full corpus.
## 数据介绍
在此教程中，我们选用[CoNLL 2005](http://www.cs.upc.edu/~srlconll/)SRL任务开放出的数据集作为示例。需要特别说明的是，CoNLL 2005 SRL任务的训练数集和开发集在比赛之后并非免费进行公开，目前，能够获取到的只有测试集，包括Wall Street Journal的23节和Brown语料集中的3节。在本教程中，我们以测试集中的WSJ数据为训练集来讲解模型。但是，由于测试集中样本的数量远远不够，如果希望训练一个可用的神经网络SRL系统，请考虑付费获取全量数据。
The original data includes a variety of information such as POS tagging, naming entity recognition, syntax tree, etc. In this tutorial, we only use the data under `test.wsj/words/` (text sequence) and `test.wsj/props/` (label results). The data directory used in this tutorial is as follows:
The annotation information is derived from the results of Penn TreeBank\[[7](#references)\] and PropBank \[[8](# references)\]. The labeling of the PropBank is different from the labeling methods mentioned before, but shares with it the same underlying principle. For descriptions of the labeling, please refer to the paper \[[9](#references)\].
1. Merge the text sequence and the tag sequence into the same record;
2. If a sentence contains $n$ predicates, the sentence will be processed $n$ times into $n$ separate training samples, each sample with a different predicate;
3. Extract the predicate context and construct the predicate context region marker;
4. Construct the markings in BIO format;
5. Obtain the integer index corresponding to the word according to the dictionary.
```python
```python
# import paddle.v2.dataset.conll05 as conll05
# import paddle.v2.dataset.conll05 as conll05
# conll05.corpus_reader does step 1 and 2 as mentioned above.
# conll05.corpus_reader函数完成上面第1步和第2步.
# conll05.reader_creator does step 3 to 5.
# conll05.reader_creator函数完成上面第3步到第5步.
# conll05.test gets preprocessed training instances.
# conll05.test函数可以获取处理之后的每条样本来供PaddlePaddle训练.
```
```
After preprocessing, a training sample contains nine features, namely: word sequence, predicate, predicate context (5 columns), region mark sequence, label sequence. The following table is an example of a training sample.
| word sequence | predicate | predicate context（5 columns） | region mark sequence | label sequence|
| 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 |
|---|---|---|---|---|
|---|---|---|---|---|
| A | set | n't been set . × | 0 | B-A1 |
| A | set | n't been set . × | 0 | B-A1 |
| record | set | n't been set . × | 0 | I-A1 |
| record | set | n't been set . × | 0 | I-A1 |
...
@@ -237,18 +214,19 @@ After preprocessing, a training sample contains nine features, namely: word sequ
...
@@ -237,18 +214,19 @@ After preprocessing, a training sample contains nine features, namely: word sequ
| set | set | n't been set . × | 1 | B-V |
| set | set | n't been set . × | 1 | B-V |
| . | set | n't been set . × | 1 | O |
| . | set | n't been set . × | 1 | O |
In addition to the data, we provide following resources:
| filename | explanation |
除数据之外，我们同时提供了以下资源：
| 文件名称 | 说明 |
|---|---|
|---|---|
| word_dict | dictionary of input sentences, total 44068 words |
| word_dict | 输入句子的词典，共计44068个词 |
| label_dict | dictionary of labels, total 106 labels |
| label_dict | 标记的词典，共计106个标记 |
| predicate_dict | predicate dictionary, total 3162 predicates |
| predicate_dict | 谓词的词典，共计3162个词 |
| emb | a pre-trained word vector lookup table, 32-dimentional |
| emb | 一个训练好的词表，32维 |
We trained a language model on the English Wikipedia to get a word vector lookup table used to initialize the SRL model. While training the SRL model, the word vector lookup table is no longer updated. To learn more about the language model and the word vector lookup table, please refer to the tutorial [word vector](https://github.com/PaddlePaddle/book/blob/develop/04.word2vec/README.md). There are 995,000,000 tokens in the training corpus, and the dictionary size is 4900,000 words. In the CoNLL 2005 training corpus, 5% of the words are not in the 4900,000 words, and we see them all as unknown words, represented by `<unk>`.
Note that `hidden_dim = 512` means a LSTM hidden vector of 128 dimension (512/4). Please refer to PaddlePaddle's official documentation for detail: [lstmemory](http://www.paddlepaddle.org/doc/ui/api/trainer_config_helpers/layers.html#lstmemory)。
- 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 paddle.layer.mixed (paddle.layer.fc can be uesed as well), and the CRF layer in PaddlePaddle: paddle.layer.crf only learns the transition features, which is a cost layer and is the last layer of the network. paddle.layer.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.
- 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, `paddle.layer.crf_decoding` calculates labelling error for each input token and `evaluator.sum` sum the error over the entire sequence. Otherwise, `paddle.layer.crf_decoding` generates the labelling tags.
We will create trainer given model topology, parameters, and optimization method. We will use the most basic **SGD** method, which is a momentum optimizer with 0 momentum. Meanwhile, we will set learning rate and regularization.
As mentioned in data preparation section, we will use CoNLL 2005 test corpus as the training data set. `conll05.test()` outputs one training instance at a time. It is shuffled and batched into mini batches, and used as input.
`feeding` is used to specify the correspondence between data instance and data layer. For example, according to following `feeding`, the 0th column of data instance produced by`conll05.test()` is matched to the data layer named `word_data`.
`event_handler` can be used as callback for training events, it will be used as an argument for the `train` method. Following `event_handler` prints cost during training.
print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics)
print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics)
```
```
`trainer.train` will train the model.
通过`trainer.train`函数训练：
```python
```python
trainer.train(
trainer.train(
reader=reader,
reader=reader,
event_handler=event_handler,
event_handler=event_handler,
num_passes=10000,
num_passes=1,
feeding=feeding)
feeding=feeding)
```
```
### Application
### 应用模型
Aftern training is done, we need to select an optimal model based one performance index to do inference. In this task, one can simply select the model with the least number of marks on the test set. The `paddle.layer.crf_decoding` layer is used in the inference, but its inputs does not include the ground truth label.
@@ -561,11 +542,11 @@ pre_lab = [labels_reverse[i] for i in labs]
...
@@ -561,11 +542,11 @@ pre_lab = [labels_reverse[i] for i in labs]
print pre_lab
print pre_lab
```
```
## Conclusion
## 总结
Semantic Role Labeling is an important intermediate step in a wide range of natural language processing tasks. In this tutorial, we use SRL as an example to illustrate using PaddlePaddle to do sequence tagging tasks. The models proposed are from our published paper\[[10](#Reference)\]. We only use test data for illustration since the training data on the CoNLL 2005 dataset is not completely public. This aims to propose an end-to-end neural network model with fewer dependencies on natural language processing tools but is comparable, or even better than traditional models in terms of performance. Please check out our paper for more information and discussions.
1. Sun W, Sui Z, Wang M, et al. [Chinese semantic role labeling with shallow parsing](http://www.aclweb.org/anthology/D09-1#page=1513)[C]//Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 3-Volume 3. Association for Computational Linguistics, 2009: 1475-1483.
1. Sun W, Sui Z, Wang M, et al. [Chinese semantic role labeling with shallow parsing](http://www.aclweb.org/anthology/D09-1#page=1513)[C]//Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 3-Volume 3. Association for Computational Linguistics, 2009: 1475-1483.
2. Pascanu R, Gulcehre C, Cho K, et al. [How to construct deep recurrent neural networks](https://arxiv.org/abs/1312.6026)[J]. arXiv preprint arXiv:1312.6026, 2013.
2. Pascanu R, Gulcehre C, Cho K, et al. [How to construct deep recurrent neural networks](https://arxiv.org/abs/1312.6026)[J]. arXiv preprint arXiv:1312.6026, 2013.
3. Cho K, Van Merriënboer B, Gulcehre C, et al. [Learning phrase representations using RNN encoder-decoder for statistical machine translation](https://arxiv.org/abs/1406.1078)[J]. arXiv preprint arXiv:1406.1078, 2014.
3. Cho K, Van Merriënboer B, Gulcehre C, et al. [Learning phrase representations using RNN encoder-decoder for statistical machine translation](https://arxiv.org/abs/1406.1078)[J]. arXiv preprint arXiv:1406.1078, 2014.
...
@@ -578,7 +559,7 @@ Semantic Role Labeling is an important intermediate step in a wide range of natu
...
@@ -578,7 +559,7 @@ Semantic Role Labeling is an important intermediate step in a wide range of natu
10. Zhou J, Xu W. [End-to-end learning of semantic role labeling using recurrent neural networks](http://www.aclweb.org/anthology/P/P15/P15-1109.pdf)[C]//Proceedings of the Annual Meeting of the Association for Computational Linguistics. 2015.
10. Zhou J, Xu W. [End-to-end learning of semantic role labeling using recurrent neural networks](http://www.aclweb.org/anthology/P/P15/P15-1109.pdf)[C]//Proceedings of the Annual Meeting of the Association for Computational Linguistics. 2015.
<br/>
<br/>
This tutorial is contributed by <axmlns:cc="http://creativecommons.org/ns#"href="http://book.paddlepaddle.org"property="cc:attributionName"rel="cc:attributionURL">PaddlePaddle</a>, and licensed under a <arel="license"href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<!-- This block will be replaced by each markdown file content. Please do not change lines below.-->
<divid="markdown"style='display:none'>
<divid="markdown"style='display:none'>
# Machine Translation
# 机器翻译
The source codes is located at [book/machine_translation](https://github.com/PaddlePaddle/book/tree/develop/08.machine_translation). Please refer to the PaddlePaddle [installation tutorial](https://github.com/PaddlePaddle/book/blob/develop/README.en.md#running-the-book) if you are a first time user.
Machine translation (MT) leverages computers to translate from one language to another. The language to be translated is referred to as the source language, while the language to be translated into is referred to as the target language. Thus, Machine translation is the process of translating from the source language to the target language. It is one of the most important research topics in the field of natural language processing.
Early machine translation systems are mainly rule-based i.e. they rely on a language expert to specify the translation rules between the two languages. It is quite difficult to cover all the rules used in one languge. So it is quite a challenge for language experts to specify all possible rules in two or more different languages. Hence, a major challenge in conventional machine translation has been the difficulty in obtaining a complete rule set \[[1](#References)\]。
To address the aforementioned problems, statistical machine translation techniques have been developed. These techniques learn the translation rules from a large corpus, instead of being designed by a language expert. While these techniques overcome the bottleneck of knowledge acquisition, there are still quite a lot of challenges, for example:
1. human designed features cannot cover all possible linguistic variations;
2. it is difficult to use global features;
3. the techniques heavily rely on pre-processing techniques like word alignment, word segmentation and tokenization, rule-extraction and syntactic parsing etc. The error introduced in any of these steps could accumulate and impact translation quality.
The recent development of deep learning provides new solutions to these challenges. The two main categories for deep learning based machine translation techniques are:
1. techniques based on the statistical machine translation system but with some key components improved with neural networks, e.g., language model, reordering model (please refer to the left part of Figure 1);
2. techniques mapping from source language to target language directly using a neural network, or end-to-end neural machine translation (NMT).
<palign="center">
<palign="center">
<imgsrc="image/nmt_en.png"width=400><br/>
<imgsrc="image/nmt.png"width=400><br/>
Figure 1. Neural Network based Machine Translation
图1. 基于神经网络的机器翻译系统
</p>
</p>
本教程主要介绍NMT模型，以及如何用PaddlePaddle来训练一个NMT模型。
This tutorial will mainly introduce an NMT model and how to use PaddlePaddle to train it.
## 效果展示
## Illustrative Results
以中英翻译（中文翻译到英文）的模型为例，当模型训练完毕时，如果输入如下已分词的中文句子：
Let's consider an example of Chinese-to-English translation. The model is given the following segmented sentence in Chinese
```text
```text
这些 是 希望 的 曙光 和 解脱 的 迹象 .
这些 是 希望 的 曙光 和 解脱 的 迹象 .
```
```
After training and with a beam-search size of 3, the generated translations are as follows:
如果设定显示翻译结果的条数（即[柱搜索算法](#柱搜索算法)的宽度）为3，生成的英语句子如下：
```text
```text
0 -5.36816 These are signs of hope and relief . <e>
0 -5.36816 These are signs of hope and relief . <e>
1 -6.23177 These are the light of hope and relief . <e>
1 -6.23177 These are the light of hope and relief . <e>
2 -7.7914 These are the light of hope and the relief of hope . <e>
2 -7.7914 These are the light of hope and the relief of hope . <e>
```
```
- The first column corresponds to the id of the generated sentence; the second column corresponds to the score of the generated sentence (in descending order), where a larger value indicates better quality; the last column corresponds to the generated sentence.
This section will introduce Gated Recurrent Unit (GRU), Bi-directional Recurrent Neural Network, the Encoder-Decoder framework used in NMT, attention mechanism, as well as the beam search algorithm.
We already introduced RNN and LSTM in the [Sentiment Analysis](https://github.com/PaddlePaddle/book/blob/develop/understand_sentiment/README.md) chapter.
### GRU
Compared to a simple RNN, the LSTM added memory cell, input gate, forget gate and output gate. These gates combined with the memory cell greatly improve the ability to handle long-term dependencies.
GRU\[[2](#References)\] proposed by Cho et al is a simplified LSTM and an extension of a simple RNN. It is shown in the figure below.
- reset gate: when this gate is closed, the history information is discarded, i.e., the irrelevant historical information has no effect on the future output.
- update gate: it combines the input gate and the forget gate and is used to control the impact of historical information on the hidden output. The historical information is passed over when the update gate is close to 1.
Generally speaking, sequences with short distance dependencies will have an active reset gate while sequences with long distance dependency will have an active update date.
In addition, Chung et al.\[[3](#References)\] have empirically shown that although GRU has less parameters, it has similar performance to LSTM on several different tasks.
### Bi-directional Recurrent Neural Network
### 双向循环神经网络
We already introduced an instance of bi-directional RNN in the [Semantic Role Labeling](https://github.com/PaddlePaddle/book/blob/develop/label_semantic_roles/README.md) chapter. Here we present another bi-directional RNN model with a different architecture proposed by Bengio et al. in \[[2](#References),[4](#References)\]. This model takes a sequence as input and outputs a fixed dimensional feature vector at each step, encoding the context information at the corresponding time step.
Specifically, this bi-directional RNN processes the input sequence in the original and reverse order respectively, and then concatenates the output feature vectors at each time step as the final output. Thus the output node at each time step contains information from the past and future as context. The figure below shows an unrolled bi-directional RNN. This network contains a forward RNN and backward RNN with six weight matrices: weight matrices from input to forward hidden layer and backward hidden ($W_1, W_3$), weight matrices from hidden to itself ($W_2, W_5$), matrices from forward hidden and backward hidden to output layer ($W_4, W_6$). Note that there are no connections between forward hidden and backward hidden layers.
The Encoder-Decoder\[[2](#References)\] framework aims to solve the mapping of a sequence to another sequence, for sequences with arbitrary lengths. The source sequence is encoded into a vector via an encoder, which is then decoded to a target sequence via a decoder by maximizing the predictive probability. Both the encoder and the decoder are typically implemented via RNN.
1. One-hot vector representation of a word: Each word $x_i$ in the source sentence $x=\left \{ x_1,x_2,...,x_T \right \}$ is represented as a vector $w_i\epsilon \left \{ 0,1 \right \}^{\left | V \right |},i=1,2,...,T$ where $w_i$ has the same dimensionality as the size of the dictionary, i.e., $\left | V \right |$, and has an element of one at the location corresponding to the location of the word in the dictionary and zero elsewhere.
2. Word embedding as a representation in the low-dimensional semantic space: There are two problems with one-hot vector representation
编码阶段分为三步：
* the dimensionality of the vector is typically large, leading to the curse of dimensionality;
* it is hard to capture the relationships between words, i.e., semantic similarities. Therefore, it is useful to project the one-hot vector into a low-dimensional semantic space as a dense vector with fixed dimensions, i.e., $s_i=Cw_i$ for the $i$-th word, with $C\epsilon R^{K\times \left | V \right |}$ as the projection matrix and $K$ is the dimensionality of the word embedding vector.
2. 映射到低维语义空间的词向量：one-hot vector表示存在两个问题，1）生成的向量维度往往很大，容易造成维数灾难；2）难以刻画词与词之间的关系（如语义相似性，也就是无法很好地表达语义）。因此，需再one-hot vector映射到低维的语义空间，由一个固定维度的稠密向量（称为词向量）表示。记映射矩阵为$C\epsilon R^{K\times \left | V \right |}$，用$s_i=Cw_i$表示第$i$个词的词向量，$K$为向量维度。
3. Encoding of the source sequence via RNN: This can be described mathematically as:
$\varnothing _\theta$ is a non-linear activation function, and
$\mathbf{h}=\left \{ h_1,..., h_T \right \}$
is the sequential encoding of the first $T$ words from the source sequence. The vector representation of the whole sentence can be represented as the encoding vector at the last time step $T$ from $\mathbf{h}$, or by temporal pooling over $\mathbf{h}$.
Bi-directional RNN can also be used in step (3) for more a complicated sentence encoding. This can be implemented using a bi-directional GRU. Forward GRU encodes the source sequence in its original order $(x_1,x_2,...,x_T)$, and generates a sequence of hidden states $(\overrightarrow{h_1},\overrightarrow{h_2},...,\overrightarrow{h_T})$. The backward GRU encodes the source sequence in reverse order, i.e., $(x_T,x_T-1,...,x_1)$ and generates $(\overleftarrow{h_1},\overleftarrow{h_2},...,\overleftarrow{h_T})$. Then for each word $x_i$, its complete hidden state is the concatenation of the corresponding hidden states from the two GRUs, i.e., $h_i=\left [ \overrightarrow{h_i^T},\overleftarrow{h_i^T} \right ]^{T}$.
The goal of the decoder is to maximize the probability of the next correct word in the target language. The main idea is as follows:
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. At each time step $i$, given the encoding vector (or context vector) $c$ of the source sentence, the $i$-th word $u_i$ from the ground-truth target language and the RNN hidden state $z_i$, the next hidden state $z_{i+1}$ is computed as:
where $\phi _{\theta '}$ is a non-linear activation function and $c=q\mathbf{h}$ is the context vector of the source sentence. Without using [attention](#Attention Mechanism), if the output of the [encoder](#Encoder) is the encoding vector at the last time step of the source sentence, then $c$ can be defined as $c=h_T$. $u_i$ denotes the $i$-th word from the target language sentence and $u_0$ denotes the beginning of the target language sentence (i.e., `<s>`), indicating the beginning of decoding. $z_i$ is the RNN hidden state at time step $i$ and $z_0$ is an all zero vector.
2. Calculate the probability $p_{i+1}$ for the $i+1$-th word in the target language sequence by normalizing $z_{i+1}$ using `softmax` as follows
3. Compute the cost accoding to $p_{i+1}$ and $u_{i+1}$.
3. 根据$p_{i+1}$和$u_{i+1}$计算代价。
4. Repeat Steps 1-3, until all the words in the target language sentence have been processed.
4. 重复步骤1~3，直到目标语言序列中的所有词处理完毕。
The generation process of machine translation is to translate the source sentence into a sentence in the target language according to a pre-trained model. There are some differences between the decoding step in generation and training. Please refer to [Beam Search Algorithm](#Beam Search Algorithm) for details.
* It is very challenging to encode both the semantic and syntactic information a sentence with a fixed dimensional vector regardless of the length of the sentence.
* Intuitively, when translating a sentence, we typically pay more attention to the parts in the source sentence more relevant to the current translation. Moreover, the focus changes along the process of the translation. With a fixed dimensional vector, all the information from the source sentence is treated equally in terms of attention. This is not reasonable. Therefore, Bahdanau et al. \[[4](#References)\] introduced attention mechanism, which can decode based on different fragments of the context sequence in order to address the difficulty of feature learning for long sentences. Decoder with attention will be explained in the following.
Different from the simple decoder, $z_i$ is computed as:
It is observed that for each word $u_i$ in the target language sentence, there is a corresponding context vector $c_i$ as the encoding of the source sentence, which is computed as:
It is noted that the attention mechanism is achieved by a weighted average over the RNN hidden states $h_j$. The weight $a_{ij}$ denotes the strength of attention of the $i$-th word in the target language sentence to the $j$-th word in the source sentence and is calculated as
where $align$ is an alignment model that measures the fitness between the $i$-th word in the target language sentence and the $j$-th word in the source sentence. More concretely, the fitness is computed with the $i$-th hidden state $z_i$ of the decoder RNN and the $j$-th context vector $h_j$ of the source sentence. Hard alignment is used in the conventional alignment model, which means each word in the target language explicitly corresponds to one or more words from the target language sentence. In an attention model, soft alignment is used, where any word in source sentence is related to any word in the target language sentence, where the strength of the relation is a real number computed via the model, thus can be incorporated into the NMT framework and can be trained via back-propagation.
[Beam Search](http://en.wikipedia.org/wiki/Beam_search) is a heuristic search algorithm that explores a graph by expanding the most promising node in a limited set. It is typically used when the solution space is huge (e.g., for machine translation, speech recognition), and there is not enough memory for all the possible solutions. For example, if we want to translate “`<s>你好<e>`” into English, even if there are only three words in the dictionary (`<s>`, `<e>`, `hello`), it is still possible to generate an infinite number of sentences, where the word `hello` can appear different number of times. Beam search could be used to find a good translation among them.
Beam search builds a search tree using breadth first search and sorts the nodes according to a heuristic cost (sum of the log probability of the generated words) at each level of the tree. Only a fixed number of nodes according to the pre-specified beam size (or beam width) are considered. Thus, only nodes with highest scores are expanded in the next level. This reduces the space and time requirements significantly. However, a globally optimal solution is not guaranteed.
The goal is to maximize the probability of the generated sequence when using beam search in decoding, The procedure is as follows:
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. At each time step $i$, compute the hidden state $z_{i+1}$ of the next time step according to the context vector $c$ of the source sentence, the $i$-th word $u_i$ generated for the target language sentence and the RNN hidden state $z_i$.
3. Sample the word $u_{i+1}$ according to $p_{i+1}$.
3. 根据$p_{i+1}$采样出单词$u_{i+1}$。
4. Repeat Steps 1-3, until end-of-sentence token `<e>` is generated or the maximum length of the sentence is reached.
4. 重复步骤1~3，直到获得句子结束标记`<e>`或超过句子的最大生成长度为止。
Note: $z_{i+1}$ and $p_{i+1}$ are computed the same way as in [Decoder](#Decoder). In generation mode, each step is greedy in so there is no guarantee of a global optimum.
This tutorial uses a dataset from [WMT-14](http://www-lium.univ-lemans.fr/~schwenk/cslm_joint_paper/), where [bitexts (after selection)](http://www-lium.univ-lemans.fr/~schwenk/cslm_joint_paper/data/bitexts.tgz) is used as the training set, and [dev+test data](http://www-lium.univ-lemans.fr/~schwenk/cslm_joint_paper/data/dev+test.tgz) is used as test and generation set.
- Merge the source and target parallel corpus files into one file
- Merge `XXX.src` and `XXX.trg` file pair as `XXX`
- The $i$-th row in `XXX` is the concatenation of the $i$-th row from `XXX.src` with the $i$-th row from `XXX.trg`, separated with '\t'.
- Create source dictionary and target dictionary, each containing **DICTSIZE** number of words, including the most frequent (DICTSIZE - 3) fo word from the corpus and 3 special token `<s>` (begin of sequence), `<e>` (end of sequence) and `<unk>` (unknown words that are not in the vocabulary).
Because the full dataset is very big, to reduce the time for downloading the full dataset. PadddlePaddle package `paddle.dataset.wmt14` provides a preprocessed `subset of dataset`(http://paddlepaddle.bj.bcebos.com/demo/wmt_shrinked_data/wmt14.tgz).
## 流程说明
This subset has 193319 instances of training data and 6003 instances of test data. Dictionary size is 30000. Because of the limitation of size of the subset, the effectiveness of trained model from this subset is not guaranteed.
### paddle初始化
## Training Instructions
### Initialize PaddlePaddle
```python
```python
# 加载 paddle的python包
import sys
import sys
import paddle.v2 as paddle
import paddle.v2 as paddle
# train with a single CPU
# 配置只使用cpu，并且使用一个cpu进行训练
paddle.init(use_gpu=False, trainer_count=1)
paddle.init(use_gpu=False, trainer_count=1)
# False: training, True: generating
# 训练模式False，生成模式True
is_generating = False
is_generating = False
```
```
### Model Configuration
### 模型结构
1. 首先，定义了一些全局变量。
1. Define some global variables
```python
```python
dict_size = 30000 # dict dim
dict_size = 30000 # 字典维度
source_dict_dim = dict_size # source language dictionary size
source_dict_dim = dict_size # 源语言字典维度
target_dict_dim = dict_size # destination language dictionary size
target_dict_dim = dict_size # 目标语言字典维度
word_vector_dim = 512 # word embedding dimension
word_vector_dim = 512 # 词向量维度
encoder_size = 512 # hidden layer size of GRU in encoder
encoder_size = 512 # 编码器中的GRU隐层大小
decoder_size = 512 # hidden layer size of GRU in decoder
decoder_size = 512 # 解码器中的GRU隐层大小
beam_size = 3 # expand width in beam search
beam_size = 3 # 柱宽度
max_length = 250 # a stop condition of sequence generation
max_length = 250 # 生成句子的最大长度
```
```
2. Implement Encoder as follows:
2. 其次，实现编码器框架。分为三步：
- Input is a sequence of words represented by an integer word index sequence. So we define data layer of data type `integer_value_sequence`. The value range of each element in the sequence is `[0, source_dict_dim)`
with paddle.layer.mixed(size=decoder_size) as encoded_proj:
with paddle.layer.mixed(size=decoder_size) as encoded_proj:
...
@@ -312,7 +275,7 @@ is_generating = False
...
@@ -312,7 +275,7 @@ is_generating = False
input=encoded_vector)
input=encoded_vector)
```
```
- Use a non-linear transformation of the last hidden state of the backward GRU on the source language sentence as the initial state of the decoder RNN $c_0=h_T$
- Define the computation in each time step for the decoder RNN, i.e., according to the current context vector $c_i$, hidden state for the decoder $z_i$ and the $i$-th word $u_i$ in the target language to predict the probability $p_{i+1}$ for the $i+1$-th word.
- context is computed via `simple_attention` as $c_i=\sum {j=1}^{T}a_{ij}h_j$, where enc_vec is the projection of $h_j$ and enc_proj is the projection of $h_j$ (c.f. 3.1). $a_{ij}$ is calculated within `simple_attention`.
- Softmax normalization is used in the end to computed the probability of words, i.e., $p\left ( u_i|u_{<i},\mathbf{x} \right )=softmax(W_sz_i+b_z)$. The output is returned.
4. Define the name for the decoder and the first two input for `gru_decoder_with_attention`. Note that `StaticInput` is used for the two inputs. Please refer to [StaticInput Document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/deep_model/rnn/recurrent_group_cn.md#输入) for more details.
- the decoder predicts a next target word based on the the last generated target word. Embedding of the last generated word is automatically gotten by GeneratedInputs.
Note: Our configuration is based on Bahdanau et al. \[[4](#Reference)\] but with a few simplifications. Please refer to [issue #1133](https://github.com/PaddlePaddle/Paddle/issues/1133) for more details.
Create every parameter that `cost` layer needs. And we can get parameter names. If the parameter name is not specified during model configuration, it will be generated.
As the training of an NMT model is very time consuming, we provide a pre-trained model. The model is trained with a cluster of 50 physical nodes (each node has two 6-core CPU) over 5 days. The provided model has the [BLEU Score](#BLEU Score) of 26.92, and the size of 205M.
src: <s> Les <unk> se <unk> au sujet de la largeur des sièges alors que de grosses commandes sont en jeu <e>
src: <s> Les <unk> se <unk> au sujet de la largeur des sièges alors que de grosses commandes sont en jeu <e>
...
@@ -577,11 +541,11 @@ Note: Our configuration is based on Bahdanau et al. \[[4](#Reference)\] but with
...
@@ -577,11 +541,11 @@ Note: Our configuration is based on Bahdanau et al. \[[4](#Reference)\] but with
prob = -19.512890: The <unk> will be rotated about the width of the seats , while large commands are at play . <e>
prob = -19.512890: The <unk> will be rotated about the width of the seats , while large commands are at play . <e>
```
```
## Summary
## 总结
End-to-end neural machine translation is a recently developed way to perform machine translations. In this chapter, we introduced the typical "Encoder-Decoder" framework and "attention" mechanism. Since NMT is a typical Sequence-to-Sequence (Seq2Seq) learning problem, tasks such as query rewriting, abstraction generation, and single-turn dialogues can all be solved with the model presented in this chapter.
端到端的神经网络机器翻译是近几年兴起的一种全新的机器翻译方法。本章中，我们介绍了NMT中典型的“编码器-解码器”框架和“注意力”机制。由于NMT是一个典型的Seq2Seq（Sequence to Sequence，序列到序列）学习问题，因此，Seq2Seq中的query改写（query rewriting）、摘要、单轮对话等问题都可以用本教程的模型来解决。
## References
## 参考文献
1. Koehn P. [Statistical machine translation](https://books.google.com.hk/books?id=4v_Cx1wIMLkC&printsec=frontcover&hl=zh-CN&source=gbs_ge_summary_r&cad=0#v=onepage&q&f=false)[M]. Cambridge University Press, 2009.
1. Koehn P. [Statistical machine translation](https://books.google.com.hk/books?id=4v_Cx1wIMLkC&printsec=frontcover&hl=zh-CN&source=gbs_ge_summary_r&cad=0#v=onepage&q&f=false)[M]. Cambridge University Press, 2009.
2. Cho K, Van Merriënboer B, Gulcehre C, et al. [Learning phrase representations using RNN encoder-decoder for statistical machine translation](http://www.aclweb.org/anthology/D/D14/D14-1179.pdf)[C]//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014: 1724-1734.
2. Cho K, Van Merriënboer B, Gulcehre C, et al. [Learning phrase representations using RNN encoder-decoder for statistical machine translation](http://www.aclweb.org/anthology/D/D14/D14-1179.pdf)[C]//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014: 1724-1734.
...
@@ -590,7 +554,7 @@ End-to-end neural machine translation is a recently developed way to perform mac
...
@@ -590,7 +554,7 @@ End-to-end neural machine translation is a recently developed way to perform mac
5. Papineni K, Roukos S, Ward T, et al. [BLEU: a method for automatic evaluation of machine translation](http://dl.acm.org/citation.cfm?id=1073135)[C]//Proceedings of the 40th annual meeting on association for computational linguistics. Association for Computational Linguistics, 2002: 311-318.
5. Papineni K, Roukos S, Ward T, et al. [BLEU: a method for automatic evaluation of machine translation](http://dl.acm.org/citation.cfm?id=1073135)[C]//Proceedings of the 40th annual meeting on association for computational linguistics. Association for Computational Linguistics, 2002: 311-318.
<br/>
<br/>
This tutorial is contributed by <axmlns:cc="http://creativecommons.org/ns#"href="http://book.paddlepaddle.org"property="cc:attributionName"rel="cc:attributionURL">PaddlePaddle</a>, and licensed under a <arel="license"href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.
