'''
author: white mai
program: knn demo
date: 2021,9,24
'''
from sklearn.datasets import load_iris
from sklearn.preprocessing import MinMaxScaler
import numpy as np
from numpy import tile
import matplotlib.pyplot as plt
from scipy.spatial.distance import pdist

# 数据初始化
train_spt = 0.8


# 数据导入
def load_data():
    iris = load_iris()
    data = iris.data
    target = iris.target
    # 打乱数据
    data = np.concatenate([data, target.reshape(-1, 1)], axis=1)
    np.random.shuffle(data)
    target = data[:, 4]
    data = data[:, 0:4]
    return data, target


# 训练集和测试集分离
def train_test_split(datasets, train_spt):
    train_size = int(len(datasets) * train_spt)
    return datasets[:train_size], datasets[train_size:]


# knn算法部分
# 求距离
def distance(train_data, test_data, k):
    # 欧几里得距离
    if k == 1:
        test_data = tile(test_data, train_data.shape[0]).reshape(-1, train_data.shape[1])
        dist = ((train_data - test_data)**2).sum(axis=1)**0.5
        return dist
    # 曼哈顿距离
    elif k == 2:
        test_data = tile(test_data, train_data.shape[0]).reshape(-1, train_data.shape[1])
        dist = (abs(train_data - test_data)).sum(axis=1)**0.5
        return dist
    # 闵可夫斯基距离(Minkowski Distance)
    elif k == 3:
        test_data = tile(test_data, train_data.shape[0]).reshape(-1, train_data.shape[1])
        dist = ((train_data - test_data)**train_data.shape[1]).sum(axis=1) ** (1/train_data.shape[1])
        return dist
    # 余弦距离函数:
    elif k == 4:
        test_data = tile(test_data, train_data.shape[0]).reshape(-1, train_data.shape[1])
        distes = np.array([])
        for i in range(test_data.shape[0]):
            Vec = np.vstack([train_data[i], test_data[i]])
            dist = 1 - pdist(Vec, 'cosine')
            distes = np.append(distes, dist)
        return distes


# 判断函数
def decide(data):
    data = list(data)
    return max(set(data), key=data.count)


def knn(train_data, test_data, train_label, test_label, k):
    loss = 0
    matrix_dist = np.array([])
    dist = 0
    for i in range(test_data.shape[0]):
        dist = distance(train_data, test_data[i, :], 2)
        dist = np.concatenate([dist.reshape(-1, 1), train_label.reshape(-1, 1)], axis=1)
        dist = dist[np.argsort(dist, 0)[:, 0], :]
        # 取前K个元素
        ans_label = decide(dist[0:k, 1])
        if ans_label != test_label[i]:
            loss += 1
        # matrix_dist = matrix_dist.reshape(-1, test_data.shape[0])
    loss = loss / test_label.shape[0]
    return dist, loss


if __name__ == '__main__':

    # 模型预处理
    data, target = load_data()
    data_train, data_test = train_test_split(data, train_spt)
    label_train, label_test = train_test_split(target, train_spt)

    # min_max归一化
    sc = MinMaxScaler(feature_range=(0, 1))
    data_train_scaled = sc.fit_transform(data_train)
    data_test_scaled = sc.transform(data_test)
    losses = np.array([])
    for k in range(1, 50):
        C, loss = knn(data_train_scaled, data_test_scaled, label_train, label_test, k)
        print(loss)
        losses = np.append(losses, loss)


ks = np.array([i for i in range(1, 50)])
plt.title('the relationship between k and loss')
plt.xlabel('k')
plt.ylabel('loss')
plt.scatter(ks, losses)
plt.show()