吴恩达机器学习作业3---Multi-class Classification and Neural Networks

多元分类和神经网络

代码分析

首先导入需要的类库

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import scipy.io #Used to load the OCTAVE *.mat files
import scipy.misc #Used to show matrix as an image
import matplotlib.cm as cm #Used to display images in a specific colormap
import random #To pick random images to display
from scipy.special import expit #Vectorized sigmoid function
from scipy import optimize
#可选
%matplotlib inline

1.Multi-class Classification

导入数据

#导入.mat数据
datafile = 'data/ex3data1.mat'
mat = scipy.io.loadmat( datafile )
X, y = mat['X'], mat['y']        
#在X矩阵前插入全为1的一列
X = np.insert(X,0,1,axis=1)
#X有5000张图片，每张图片大小20x20，共有400+1=401个像素点
#y有5000个值，值的集合为0-9

测试

print ("'y' shape: %s. Unique elements in y: %s"%(y.shape,np.unique(y)))
print ("'X' shape: %s. X[0] shape: %s"%(X.shape,X[0].shape))

输出

'y' shape: (5000, 1). Unique elements in y: [ 1  2  3  4  5  6  7  8  9 10]
'X' shape: (5000, 401). X[0] shape: (401,)

将400的行向量转为20x20的ndarray的函数

def getDatumImg(row):
    width, height = 20, 20
    #row.shape=401
    #将400的行向量转为20x20的narray
    square = row[1:].reshape(width,height)
    return square.T

可视化数据为黑白图片

def displayData(X,indices_to_display = None):
    #图片格式为20x20
    width, height = 20, 20
    #10x10的图像网格
    nrows, ncols = 10, 10
    #5000张图片中随机抽取100张
    if not indices_to_display:
        indices_to_display = random.sample(range(X.shape[0]), nrows*ncols)
    #200x200的narray
    big_picture = np.zeros((height*nrows,width*ncols))
    #遍历图片集,为空白的大图片赋值
    irow, icol = 0, 0
    for idx in indices_to_display:
        if icol == ncols:
            irow += 1
            icol  = 0
        iimg = getDatumImg(X[idx])
        #将这块区域的图片赋值
        big_picture[irow*height:irow*height+iimg.shape[0],icol*width:icol*width+iimg.shape[1]] = iimg
        icol += 1
    #输出图片
    fig = plt.figure(figsize=(6,6))
    plt.imshow(big_picture,cmap ='gray')

测试

displayData(X)

下面我们将使用logistic regression算法来进行多元分类

如果使用向量化代替for-loop来遍历样本集，模型的训练速度会提升

下面开始将logistic regression算法向量化

首先，这是带有正则化参数的logistic regression的代价函数

J ( θ ) = − 1 m [ ∑ i = 1 m y (