Cookbook for Practice 9.1

Cong Li 李聪

实践9.1攻略

Different Context Sizes�不同的上下文大小 (1)

Based on Practice 8.2, rename the file ‘interestData.py’ to ‘interestDataEnsemble.py’, change the class name accordingly

基于实践8.2，将文件“interestData.py”重命名为“interestDataEnsemble.py”，对类型名也作相应改动

Different Context Sizes�不同的上下文大小 (2)

In the member function ‘get_items_and_label’, rewrite the for loop to find & return the position of ‘interest’ as well

在成员函数“get_items_and_label”中，重写for循环，找到“interest”的位置并追加返回该位置

Different Context Sizes�不同的上下文大小 (2)

In the member function ‘get_items_and_label’, rewrite the for loop to find & return the position of ‘interest’ as well

在成员函数“get_items_and_label”中，重写for循环，找到“interest”的位置并追加返回该位置

Different Context Sizes�不同的上下文大小 (2)

In the member function ‘get_items_and_label’, rewrite the for loop to find & return the position of ‘interest’ as well

在成员函数“get_items_and_label”中，重写for循环，找到“interest”的位置并追加返回该位置

Different Context Sizes�不同的上下文大小 (2)

In the member function ‘get_items_and_label’, rewrite the for loop to find & return the position of ‘interest’ as well

在成员函数“get_items_and_label”中，重写for循环，找到“interest”的位置并追加返回该位置

Different Context Sizes�不同的上下文大小 (2)

In the member function ‘get_items_and_label’, rewrite the for loop to find & return the position of ‘interest’ as well

在成员函数“get_items_and_label”中，重写for循环，找到“interest”的位置并追加返回该位置

Modify member function ‘build_dict’ to accept the new return value 修改“build_dict”成员函数以接受新返回的值

Different Context Sizes�不同的上下文大小 (3)

Rewrite the member function ‘convert_features_and_label’. Add a new input parameter for context size. Calculate the left & the right side of the context given the position of ‘interest’.

重写成员函数“convert_features_and_label”。为上下文大小加一个新的输入参数。根据“interest”的位置计算上下文的左右边界

Different Context Sizes�不同的上下文大小 (4)

In converting word items to the attributes (features), rewrite the for loop to use based on the left and right side of the context

当把词项转换成属性时，基于上下文左右边界重写for循环

Different Context Sizes�不同的上下文大小 (5)

Now modify the initialization function. Recall that the original initialization function read the input file & convert each valid line into the attributes & a class label. Here we want to convert each valid line to a set of contexts w/ different sizes & a class label

现在修改初始化函数。回忆一下，原先的初始化函数读取输入文件，把每一有效行转换成属性和类别标记的。这里我们要把每一有效行转换成一组不同大小的上下文和类别标记

At the beginning, use an array for different context size: 1, 3, 5, 7, & full context 在起始处，用一个数组记录不同的上下文大小：1，3，5，7和所有上下文

Different Context Sizes�不同的上下文大小 (6)

Then rewrite the processing of each line. Here we use a 2D array for contexts w/ different sizes. Each context is an 1D array of attributes

然后我们重写对每一行的处理。这里我们用一个二维数组来表示不同大小的上下文。每一个上下文是一个一维数组

Modify Classifiers �修改分类器 (1)

In ‘perceptron.py’, in the class ‘Perceptron’ add a new member function ‘score_all’ to return the scores of all the class labels in a dictionary

在“perceptron.py”中的“Perceptron”类型中，加入一个新的成员函数“score_all”，返回所有类别标记的分数的字典

Modify Classifiers �修改分类器 (2)

Similarly in the class ‘LogisticRegression’ in ‘logisticRegression.py’, add a new member function ‘score_all’ to return the scores

类似地，在“logisticRegression.py”中的“LogisticRegression”类型中，加入一个新的成员函数“score_all”，返回所有的分数

Learning Algorithm Instances�学习算法实例 (1)

Create a new file ‘ensemble.py’ & import the required components

创建一个新文件“ensemble.py”，导入需要的组件

Create a class ‘ClassifierFactory’ & accept an algorithm name w/ a set of parameters as the input 创建一个新类型“Classifier-Factory”，在初始化函数中接受一个算法的名字和一组相应的参数

Learning Algorithm Instances�学习算法实例 (2)

Add a member function ‘get_classifier’ to return a new learning algorithm instance based on the classifier name & parameters

加一个成员函数“get_classifier”用作基于学习算法的名字和参数返回一个新的学习算法实例

Ensemble Learning 合奏学习 (1)

Add another class ‘Ensemble’ & in its initialization function, taking a parameter for the factory which will return a new learning algorithm instance on demand

创建另一个新的类型“Ensemble”，在初始化函数中，接受一个生成器参数，能够按需返回一个新的学习算法的实例

Ensemble Learning 合奏学习 (2)

In training, first get the training data to see how many contexts w/ different sizes there are to determine how many classifiers need to be trained. Then in a loop, get the learning algorithm instance from the factory

训练时，先从训练数据中察看有多少不同大小的上下文，以决定要训练多少个分类器。然后在循环中通过生成器获取学习算法的实例

Ensemble Learning 合奏学习 (3)

Now construct a new set of training data for the context with a certain size from the aggregated training data, train the classifier & put the classifier in the ensemble list

从聚合的训练数据中构造一个特定大小上下文的训练数据集，用来训练分类器，并把训练好的分类器放进合奏列表中

Ensemble Learning 合奏学习 (4)

In classification, for each class label, add the scores (weighted voting) from all the classifiers together using a dictionary

分类时，对每一个类别标记，把所有的分类器给出的分数加起来（带权投票），记在一个字典里

Ensemble Learning 合奏学习 (5)

Finally get the result by choosing the highest voted label

最后，获取最高票数的类别标记作为分类结果

Main Program 主程序 (1)

Create a new file ‘test.py’ & import the required components

创建一个新文件“test.py”，导入所需组件

Write the main program, starting from reading the data from the file & constructing the contexts w/ different sizes

写主程序，起始处从文件中读入数据，并构造出不同大小的上下文

Main Program 主程序 (2)

Randomly split the data into the training part & the evaluation part

把数据随机划分成训练部分和评估部分

Main Program 主程序 (3)

Create a factory generating perceptron instances, use it to create an ensemble classifier, & evaluate its performance

创建一个生成感知器实例的生成器，用它创建一个合奏分类器，并评价其准确率

Add a pause before going forward

继续运行前加入一个停顿

Main Program 主程序 (4)

Create another factory generating logistic regression instances, use it to create an ensemble classifier, & evaluate its performance

创建一个生成对数几率回归实例的生成器，用它创建一个合奏分类器，并评价其准确率

Run it 运行之

If you have enough time (e.g., 1 hour), hit enter to continue

如果你有足够的时间（例如1小时），按回车继续

The End