Cookbook for Practice 6.1

Cong Li 李聪

实践6.1攻略

Suppress Output 减少输出

We will run a large number of training & testing passes. Therefore we need to suppress the output in ‘tester.py’.

我们会运行很多轮训练和评估。因此我们需要减少“tester.py”中的输出。

We will also return the accuracy after each round of training & testing for upper level processing

每一轮训练和评估结束之后，我们需要将准确率返回至上层进行处理

A New Start 重新开始

Start a new file ‘test.py’ & import the relevant components

开始一个新的“test.py”文件，并导入相关组件

Then define some constants: the maximum dimension allowed, relevant attribute number, total rounds, the test size in each round 然后定义常数：最大允许维度、相关属性数、总轮数、每轮测试数据量

Generate Hyperplane �生成超平面 (1)

Write a new function ‘generate_parameters’, starting from randomly determining the attribute importance from a normal distribution 写一个新的“generate_parameters”函数，从一个正态分布中确定属性的重要性

Generate Hyperplane �生成超平面 (2)

Sort the attributes w.r.t. the importance, & then randomly determine the sign 根据重要性对属性排序，并随机确定其符号

Generate a random bias 随机生成偏向

Return the weights & the bias 返回权重和偏向

Generate Data 生成数据 (1)

Write a new function ‘generate_data’ to take inputs of the data dimension & the hyperplane, initializing the attributes array & the summation to get the ground truth class label

写一个“generate_data”的新函数，接受数据维度和超平面的输入，并初始化属性数组和总和变量以获取真实类别标记

Generate Data 生成数据 (2)

Randomly generate the values till the maximum dimension, but only use them as the attributes based on the dimension specified, & also do the calculation based on the hyperplane

依据最大维度生成随机数，但仅依据指定维度用作属性，同时根据超平面进行计算

Generate Data 生成数据 (2)

Randomly generate the values till the maximum dimension, but only use them as the attributes based on the dimension specified, & also do the calculation based on the hyperplane

依据最大维度生成随机数，但仅依据指定维度用作属性，同时根据超平面进行计算

Generate Data 生成数据 (3)

Determine the ground truth class label 确定真实类别标记

Finally return the attributes & the ground truth class label 最后，返回属性和真实类别标记

Main Section 主程序 (1)

Now write the main section, starting w/ parsing the argument for training data size

现在写主程序，从命令行参数解析、获取训练数据量开始

Loop the dimension over 3 to 16, & use a variable to track the average accuracy �从3到16作维度循环，用一个变量来记录平均准确率

Main Section 主程序 (2)

Try numerous rounds. In each round, start w/ a different random seed & generate the hyperplane

尝试很多轮，每一轮都用一个不同的随机种子，并生成超平面

Generate the test data first 先生成测试数据

Main Section 主程序 (3)

Next, generate training data 然后生成训练数据

Now use the perceptron for learning & classification

现在，用感知器进行学习和分类

Main Section 主程序 (4)

Finally, after the rounds for a dimension number, print the average accuracy

最后，当一个维度的所有轮次都结束之后，输出平均准确率

Run it w/ the training size of 20, 50, & 100

尝试不同的训练数据量（20、50和100）运行该程序

The End