机器学习作业 K-均值 python 代码注释版

初学，基本关键句都加了下注释

import matplotlib.pyplot as plt
import seaborn as sns

import numpy as np
import pandas as pd
import scipy.io as sio
mat = sio.loadmat('ex7data2.mat')
data2 = pd.DataFrame(mat.get('X'), columns=['X1', 'X2'])
#给矩阵加一列
def combine_data_C(data, C):
    data_with_c = data.copy()
    data_with_c['C'] = C
    return data_with_c


# k-means fn --------------------------------
# 从data中随机取k个样本
def random_init(data, k):
    """choose k sample from data set as init centroids
    Args:
        data: DataFrame
        k: int
    Returns:
        k samples: ndarray
    """
    return data.sample(k).values

#x：某个样本，centroids：k集（k*n）；寻找某个样本归属的簇；返回的是k的index
def _find_your_cluster(x, centroids):
    """find the right cluster for x with respect to shortest distance
    Args:
        x: ndarray (n, ) -> n features
        centroids: ndarray (k, n)
    Returns:
        k: int
    """
    #这个函数是按照axis方向返回（function（arr））结果是k*1维
    distances = np.apply_along_axis(func1d=np.linalg.norm,  # this give you l2 norm范数函数
                                    axis=1,
                                    arr=centroids - x)  # use ndarray's broadcast
    return np.argmin(distances)

#data：样本集；centroid：簇集；寻找每个样本对应的簇；
def assign_cluster(data, centroids):
    """assign cluster for each node in data
    return C ndarray
    """
    return np.apply_along_axis(lambda x: _find_your_cluster(x, centroids),
                               axis=1,
                               arr=data.values)


def new_centroids(data, C):
    data_with_c = combine_data_C(data, C)
    # 按照：c分组，求均值，再删掉C行
    return data_with_c.groupby('C', as_index=False).mean().sort_values(by='C').drop('C', axis=1).values


def cost(data, centroids, C):
    m = data.shape[0]
    # centroids:3*2,C:m*1,expand_C_with_centroids:m*2 m是样本个数 蛮神奇
    expand_C_with_centroids = centroids[C]
    # 计算样本在该聚类中心下的距离范数 然后求均值
    distances = np.apply_along_axis(func1d=np.linalg.norm,
                                    axis=1,
                                    arr=data.values-expand_C_with_centroids)
    return distances.sum() / m

# 一次性k均值，如果初始化的k点不好 结果可能会很次
def _k_means_iter( data, k, epoch=100, tol=0.0001):
    """one shot k-means
    with early break
    """
    # 先随机初始化k个聚类中心
    centroids = random_init(data, k)
    cost_progress = []
    # 迭代100次
    for i in range(epoch):
        # 标记循环次数
        print('running epoch {}'.format(i))
        # 把样本分配给聚类中心
        C = assign_cluster(data, centroids)
        # 获取聚类中心的均值，生成新的聚类中心组centroids：k*1
        centroids = new_centroids(data, C)
        # 计算本次迭代后的代价函数，c：m*1
        cost_progress.append(cost(data, centroids, C))

        if len(cost_progress) > 1:  # early break
            #距离上次迭代优化率低于0.0001就结束迭代（may局部优化）
            if (np.abs(cost_progress[-1] - cost_progress[-2])) / cost_progress[-1] < tol:
                break

    return C, centroids, cost_progress[-1]


def k_means(data, k, epoch=100, n_init=10):
    """do multiple random init and pick the best one to return
    Args:
        data (pd.DataFrame)
    Returns:
        (C, centroids, least_cost)
    """
    # 进行10次初始化 获得10*3的迭代结果
    tries = np.array([_k_means_iter(data, k, epoch) for _ in range(n_init)])
    # 找到cost函数最小的index
    least_cost_idx = np.argmin(tries[:, -1])

    return tries[least_cost_idx]
# 从样本集中随机选取三个点
init_centroids = random_init(data2, 3)
print(init_centroids)
x = np.array([1, 1])
print(x.shape)
# fig, ax = plt.subplots(figsize=(6, 4))
# ax.scatter(x=init_centroids[:, 0], y=init_centroids[:, 1])
# #enumerate:枚举；annotate：对数据集内的点进行标注
# for i, node in enumerate(init_centroids):
#     ax.annotate('{}: ({},{})'.format(i, node[0], node[1]), node)
# # 标注一个1,1
# ax.scatter(x[0], x[1], marker='x', s=200)
# plt.show()


# 获取样本集中每一个样本对应的簇的index序列
C = assign_cluster(data2, init_centroids)
# 合并样本集和对应的簇index
# data_with_c =combine_data_C(data2, C)
# data_with_c.head()
# sns.lmplot('X1', 'X2', hue='C', data=data_with_c, fit_reg=False)
# plt.show()
# ↑浅试一下样本集分组（因为是随机聚类中心，每次运行都不一样）

final_C, final_centroid, _= _k_means_iter(data2, 3)
data_with_c = combine_data_C(data2, final_C)
# sns.lmplot('X1', 'X2', hue='C', data=data_with_c, fit_reg=False)
# plt.show()
best_C, best_centroids, least_cost = k_means(data2, 3)
data_with_c = combine_data_C(data2, best_C)
sns.lmplot('X1', 'X2', hue='C', data=data_with_c, fit_reg=False)
plt.show()


# 下面是封装好的k-均值函数
from sklearn.cluster import KMeans
# 设聚类中心数量
sk_kmeans = KMeans(n_clusters=3)
# 适配样本集
sk_kmeans.fit(data2)
# 获取样本对应的聚类中心
sk_C = sk_kmeans.predict(data2)
data_with_c = combine_data_C(data2, sk_C)
sns.lmplot('X1', 'X2', hue='C', data=data_with_c, fit_reg=False)
plt.show()