聚类（kmeans）-吴恩达机器学习基于python

导入包

import scipy.io as sio # load mat
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
import numpy as np
from sklearn.cluster import KMeans

visualize data

## data1
mat1 = sio.loadmat(r'D:\python_try\5. AndrewNg_ML\data\kmeans and pca\ex7data1.mat')
data1 = pd.DataFrame(mat1.get('X'), columns=['X1', 'X2'])

## data2
mat2 = sio.loadmat(r'D:\python_try\5. AndrewNg_ML\data\kmeans and pca\ex7data2.mat')
data2 = pd.DataFrame(mat2.get('X'), columns=['X1', 'X2'])

画出data2的图形

sns.lmplot('X1', 'X2', data=data2, fit_reg=None)
plt.show()

2D Kmeans

原理步骤

1. 选择k个随机的点，成为聚类中心（cluster centroids）
   
2. 对于数据集中的每一个数据，按照距离k个中心点的距离，将其与距离最近的中心点关联起来，与同一个中心点关联的所有点聚成一类
3. 计算每一组的平均值，将该组所关联的中心点移动到平均值的位置
   

重复step2、3直至中心点不发生改变

步骤编程

生成随机的k个中心

思路：sample(k)

def random_init(data, k):
    return data.sample(k).as_matrix() #as_matrix(): 转换成array

分配聚类

思路：

1. 单个分配
2. 集体分配：套用单个

# 单个找寻聚类
def find_cluster(x, centroids):
    distances = np.apply_along_axis(func1d=np.linalg.norm, axis=1, arr=x-centroids)
    return np.argmin(distances)

# 集体data聚类标签
def assign_cluster(data, centroids):
    return np.apply_along_axis(lambda x: find_cluster(x, centroids), axis=1, arr=data.as_matrix())

计算每个聚类的中心，生成新的聚类中心

#  data中增加一列聚类标签C
def combineDataC(data, C):
    dataC = data.copy()
    dataC['C'] = C
    return dataC

# 新中心点，同时去掉C， 再转换成array数组
def newCentroids(data, C):
    dataC = combineDataC(data, C)
    return dataC.groupby('C', as_index=False).mean().sort_values(by='C').drop('C', axis=1).as_matrix()


# 损失函数
def cost(data, centroids, C):
    m = data.shape[0] # 样本量
    dataCentroids = centroids[C] # 各行的中心坐标
    distances = np.apply_along_axis(func1d=np.linalg.norm, axis=1, arr=data.as_matrix()-dataCentroids)
    return distances.sum()/m

执行步骤

# kmeans通道，运行一次
def kMeansIter(data, k, epoch=100, tol=0.0001):
    # 生成最初的中心坐标
    centroids = random_init(data, k)
    
    costProgress = [] # 用来存放递归聚类的每次损失
    # 分配聚类标签
    for i in range(epoch):
        # print('running epoch{}'.format(i))
        C = assign_cluster(data, centroids)
        centroids = newCentroids(data, C )
        costProgress.append(cost(data, centroids, C))
        
        if len(costProgress)>1:
            if np.abs(costProgress[-1]-costProgress[-2])/costProgress[-1] < tol:
                break
                
    return C, centroids, costProgress[-1]

# 每个k运行n_init次，套用kmeans通道
def kMeans(data, k, epoch=100, n_init=10):
    tries = np.array([kMeansIter(data, k) for _ in range(n_init)])
    leasrCostIndex = np.argmin(tries[:,-1])
    
    return tries[leasrCostIndex]    

调用函数进行聚类

CList = [] # 存放每个k的数据的最佳标签
centroidsList = [] # 存放每个k的数据的最佳中心
costList = [] # 存放每个k的最小损失函数
for k in range(2,10):
    kFinalC, kFinalCentroids, kFinalCost = kMeans(data2, k)
    CList.append(kFinalC)
    centroidsList.append(kFinalCentroids)
    costList.append(kFinalCost)

运用肘部法确定k

plt.plot([i for i in range(2,10)], costList, 'bo-', markersize=20)
plt.show()

用肘部法选择$k=3$进行聚类

k =3 # 确认的k值
bestC = CList[k-2]
bestCentroids = centroidsList[k-2]
dataWithC = combineDataC(data2, bestC)

画图

fig, ax = plt.subplots(figsize=(8,6))
ax.scatter(dataWithC['X1'], dataWithC['X2'], s=50, c=dataWithC['C'], cmap='Accent', label='data')
ax.scatter(pd.DataFrame(bestCentroids)[0],pd.DataFrame(bestCentroids)[1], s=100, c=['r'], label='centroids')
ax.legend()
plt.show()

try sklearn kmeans

KMeans网址
调用KMeans()， 查看参数设置

KMeans()

sk_kmeans = KMeans(n_clusters=3)
sk_kmeans.fit(data2)

sk_kmeans.cluster_centers_ # 中心点
sk_kmeans.labels_ # 所属类
sk_kmeans.inertia_ # 各点到其重点的距离和

画图

fig, ax = plt.subplots(figsize=(8,6))
ax.scatter(SKdataWithC['X1'], SKdataWithC['X2'], s=50, c=sk_kmeans.labels_, cmap='Accent', label='data')
ax.scatter(pd.DataFrame(sk_kmeans.cluster_centers_)[0],pd.DataFrame(sk_kmeans.cluster_centers_)[1], s=100, c=['r'], label='centroids')
ax.legend()
plt.show()

Kmeans for image compression

load image

from skimage import io

pic = io.imread(r'D:\python_try\5. AndrewNg_ML\data\kmeans and pca\bird_small.png')/255.
pic.shape

(128, 128, 3)
转换成2D

data3 = pic.reshape(128*128, 3)

do kmeans

model = KMeans(n_clusters=16, n_init=100, n_jobs=-1)
model.fit(data3)

picCentroids = model.cluster_centers_
picC = model.labels_
compressionPic = picCentroids[picC].reshape((128, 128, 3))

fig, ax = plt.subplots(1, 2, figsize=(12,8))
ax[0].imshow(pic)
ax[1].imshow(compressionPic)
plt.show()

thousands of colors, reduce the number of color to 16