在这次练习中将实现K-means 聚类算法并应用它压缩图片，第二部分，将使用主成分分析算法去找到一个脸部图片的低维描述。

K-means Clustering

Implementing K-means

K-means算法是一种自动将相似的数据样本聚在一起的方法,K-means背后的直观是一个迭代过程，它从猜测初始的质心开始，然后通过重复地将示例分配到最接近的质心，然后根据分配重新计算质心，来细化这个猜测。
具体步骤：

1. 随机初始K个质心
2. 开始迭代
3. 为每个样本找到最近的质心
4. 用分配给每个质心的点来计算每个质心的平均值，作为新的质心
5. 回到三
6. 收敛于最终的均值集
   在实践中，K-means算法通常使用不同的随机初始化运行几次。从不同的随机初始化中选择不同的解的一种方法是选择成本函数值(失真)最低的解。

Finding closest centroids

Computing centroid means

for every centroid k we set

新的形心重置为所有聚集该形心和的平均值

K-means on example dataset

def find_closest_centroids(X, centroids):idx = np.zeros((len(X), 1))print(idx.shape)K = len(centroids)print(K)t = [0 for i in range(K)]print(t)for i in range(len(X)):for j in range(K):temp = centroids[j, :] - X[i, :]t[j] = temp.dot(temp.T)index = t.index(min(t)) + 1print(index)idx[i] = indexreturn np.array(idx)

mat = loadmat("ex7data2.mat")
X = mat['X']
init_centroids = np.array([[3, 3], [6, 2], [8, 5]])
idx = find_closest_centroids(X, init_centroids)
print(idx[0:3])

计算出聚集样本的平均数，得到新的centroids

def compute_centroids(X, idx):centroids = []# print(idx)K = len(np.unique(idx))m, n = X.shapecentroids = np.zeros((K, n))temp = np.zeros((K, n))count = np.zeros((K, 1))print(temp.shape)print(count.shape)for i in range(m):for j in range(K):if idx[i] == j:temp[j, :] = temp[j, :] + X[i, :]count[j] = count[j] + 1centroids = temp/countreturn centroids

绘制聚类过程

def plotData(X, centroids, idx=None):"""可视化数据，并自动分开着色。idx: 最后一次迭代生成的idx向量，存储每个样本分配的簇中心点的值centroids: 包含每次中心点历史记录"""colors = ['b', 'g', 'gold', 'darkorange', 'salmon', 'olivedrab','maroon', 'navy', 'sienna', 'tomato', 'lightgray', 'gainsboro''coral', 'aliceblue', 'dimgray', 'mintcream','mintcream']assert len(centroids[0]) <= len(colors), 'colors not enough 'subX = []  # 分号类的样本点if idx is not None:for i in range(centroids[0].shape[0]):x_i = X[idx[:, 0] == i]subX.append(x_i)else:subX = [X]  # 将X转化为一个元素的列表，每个元素为每个簇的样本集，方便下方绘图# 分别画出每个簇的点，并着不同的颜色plt.figure(figsize=(8, 5))for i in range(len(subX)):xx = subX[i]plt.scatter(xx[:, 0], xx[:, 1], c=colors[i], label='Cluster %d' % i)plt.legend()plt.grid(True)plt.xlabel('x1', fontsize=14)plt.ylabel('x2', fontsize=14)plt.title('Plot of X Points', fontsize=16)# 画出簇中心点的移动轨迹xx, yy = [], []for centroid in centroids:xx.append(centroid[:, 0])yy.append(centroid[:, 1])plt.plot(xx, yy, 'rx--', markersize=8)plotData(X, [init_centroids])
plt.show()

def run_k_means(X, centroids, max_iters):K = len(centroids)centroids_all = []centroids_all.append(centroids)centroid_i = centroidsfor i in range(max_iters):idx = find_closest_centroids(X, centroid_i)centroid_i = compute_centroids(X, idx)centroids_all.append(centroid_i)return idx, centroids_allidx, centroids_all = run_k_means(X, init_centroids, 20)
plotData(X, centroids_all, idx)
plt.show()

Random initialization

1. 洗牌所有数据
2. 取前k个作为centroids
   随机选取K个样本作为centroids

def initCentroids(X, K):m, n = X.shapeidx = np.random.choice(m, K)centroids = X[idx]return centroids

for i in range(K):centroids = initCentroids(X, K)idx, centroids_all = run_k_means(X, centroids, 10)plotData(X, centroids_all, idx)

Image compression with K-means

读入图片，128*128 RGB编码图片

from skimage import ioA = io.imread('bird_small.png')
print(A.shape)
plt.imshow(A)
A = A/255.

K-means on pixels

X = A.reshape(-1, 3)
K = 16
centroids = initCentroids(X, K)
idx, centroids_all = run_k_means(X, centroids, 10)img = np.zeros(X.shape)
centroids = centroids_all[-1]
for i in range(len(centroids)):img[idx == i] = centroids[i]img = img.reshape((128, 128, 3))fig, axes = plt.subplots(1, 2, figsize=(12, 6))
axes[0].imshow(A)
axes[1].imshow(img)

Principal Component Analysis

Example Dataset

mat = loadmat('ex7data1.mat')
X = mat['X']
print(X.shape)
plt.scatter(X[:,0], X[:,1], facecolors='none', edgecolors='b')
plt.show()

Implementing PCA

两个步骤：

1. 计算矩阵数据的协方差
2. 使用SVD 方法得出特征向量
   使用归一化是每个特征在同一范围内。

def feature_normalize(X):means = X.mean(axis=0)stds = X.std(axis=0, ddof=1)X_norm = (X - means) / stdsreturn X_norm, means, stds

def pca(X):sigma = (X.T @ X) / len(X)U, S, V = np.linalg.svd(sigma)return U, S, V

X_norm, means, stds = feature_normalize(X)
U, S, V = pca(X_norm)

print(U[:,0])
plt.figure(figsize=(7, 5))
plt.scatter(X[:,0], X[:,1], facecolors='none', edgecolors='b')plt.plot([means[0], means[0] + 1.5*S[0]*U[0,0]],[means[1], means[1] + 1.5*S[0]*U[0,1]],c='r', linewidth=3, label='First Principal Component')
plt.plot([means[0], means[0] + 1.5*S[1]*U[1,0]],[means[1], means[1] + 1.5*S[1]*U[1,1]],c='g', linewidth=3, label='Second Principal Component')
plt.grid()
plt.axis("equal")
plt.legend()

Dimensionality Reduction with PCA

Projecting the data onto the principal components

def project_data(X, U, K):Z = X @ U[:, :K]return Z

def recover_data(Z, U, K):X_rec = Z @ U[:, :K].Treturn X_rec

Reconstructing an approximation of the data

X_rec = recover_data(Z, U, 1)
X_rec[0]

Visualizing the projections

plt.figure(figsize=(7,5))
plt.axis("equal")
plot = plt.scatter(X_norm[:,0], X_norm[:,1], s=30, facecolors='none',edgecolors='b',label='Original Data Points')
plot = plt.scatter(X_rec[:,0], X_rec[:,1], s=30, facecolors='none',edgecolors='r',label='PCA Reduced Data Points')plt.title("Example Dataset: Reduced Dimension Points Shown",fontsize=14)
plt.xlabel('x1 [Feature Normalized]',fontsize=14)
plt.ylabel('x2 [Feature Normalized]',fontsize=14)
plt.grid(True)for x in range(X_norm.shape[0]):plt.plot([X_norm[x,0],X_rec[x,0]],[X_norm[x,1],X_rec[x,1]],'k--')# 输入第一项全是X坐标，第二项都是Y坐标
plt.legend()

Face Image Dataset

mat = loadmat('ex7faces.mat')
X = mat['X']
print(X.shape)

def display_data(X, row, col):fig, axs = plt.subplots(row, col, figsize=(8, 8))for r in range(row):for c in range(col):axs[r][c].imshow(X[r * col + c].reshape(32, 32).T, cmap='Greys_r')axs[r][c].set_xticks([])axs[r][c].set_yticks([])display_data(X, 10, 10)

PCA on Faces

X_norm, means, stds = feature_normalize(X)
U, S, V = pca(X_norm)
display_data(U[:,:36].T, 6, 6)

Dimensionality Reduction

z = project_data(X_norm, U, K=36)
X_rec = recover_data(z, U, K=36)
display_data(X_rec, 10, 10)