def customKmeans(vec_array, n_clusters=3, epochs=50):
    '''
    @description: 手动实现kmeans，以向量间的夹角余弦作为相似度
                 根据上述tf-idf得到的7条文本向量tfidf_array，进行聚类算法
    @param {type} vec_array- 向量组, n_clusters-聚类数目, epochs- 训练次数
    @return: cluster_labels- 分类标签
    '''

    # 初始化质心的位置
    cluster_centers = []
    cluster_centers_indx = []
    while (len(cluster_centers_indx) < n_clusters):
        indx = random.randint(0, len(vec_array)-1)
        if indx not in cluster_centers_indx:
            cluster_centers_indx.append(indx)
            cluster_centers.append(vec_array[indx])

    # 初始化向量类别
    cluster_labels = [0] * len(vec_array)
    max_similarity = [-1] * len(vec_array)
    epoch = 0
    while(epoch < epochs):
        # 计算每个向量与质心的相似性，并将其归纳到最近的质心集群中
        for i in range(0, len(vec_array)):
            max_similarity[i] = computeCosine(vec_array[i], cluster_centers[0])
            for j in range(1, n_clusters):
                temp = computeCosine(vec_array[i], cluster_centers[j])
                if (temp > max_similarity[i]):
                    max_similarity[i] = temp
                    cluster_labels[i] = j

        # 更新质心位置
        for i in range(0, n_clusters):
            # 找到集群对应原向量的下标
            indxs = [indx for indx, cluster in enumerate(
                cluster_labels) if cluster == i]
            # 根据集群向量的平均值更新质点位置
            cluster_centers[i] = np.average(
                [vec_array[indx] for indx in indxs], axis=0)

        # 当满足迭代次数或平均最大相似性时退出算法
        epoch += 1
        if (np.average(max_similarity) > 0.9):
            break
    return cluster_labels


def computeCosine(vec1, vec2):
    # 计算向量间的余弦值
    vecCosine = np.dot(vec1, vec2) / \
        (np.linalg.norm(vec1) * np.linalg.norm(vec2))
    return vecCosine


# 应用customkmeans
labels = customKmeans(tfidf_array, n_clusters=3, epochs=1000)
print(labels)

输出结果：
[0 2 0 1 2 2 1]