k-NN回顾

k-NN基本实现

  1. import numpy as np
  2. import pandas as pd
  1. ## 读取数据
  2. root = '../cifar-10-batches-py/'
  4. # a = pd.read_pickle(path=root+'data_batch_1')
  5. # print(a.keys())
  6. Xtr = np.zeros((50000, 32*32*3))
  7. ytr = np.zeros((50000, 1))
  8. for i in range(1, 6):
  9.     Xtr[(i-1)*10000:i*10000, :] = pd.read_pickle(path=root+'data_batch_'+str(i))['data']
  10.     ytr[(i-1)*10000:i*10000] = np.reshape(pd.read_pickle(path=root+'data_batch_'+str(i))['labels'],(10000,1))
  11. Xte = pd.read_pickle(path=root+'test_batch')['data']
  12. yte = np.reshape(pd.read_pickle(path=root+'test_batch')['labels'],(10000,1))
  1. ## 定义KNN
  2. class kNN(object):
  3.     def __init__(self,X,y):
  4.         self.Xtr = X
  5.         self.ytr = y
  7.     def forward(self, Xte, K):
  8.         """
  9.         Xte:测试集输入
  10.         K：kNN的k
  11.         """
  12.         Ypred = np.zeros((Xte.shape[0], K),dtype = self.ytr.dtype)
  13.         for i in range(Xte.shape[0]):
  14.             distances = np.sum(np.abs(self.Xtr - Xte[i,:]), axis = 1) # 计算训练集里面每一幅图片和测试图片的距离
  15.             min_indexs = np.argsort(distances)[0:K]
  16. #             print(min_indexs)
  17.             ## 从min_indexs中获取重复次数最多的一个类
  18.             min_lables = [self.ytr[i][0] for i in min_indexs] # ytr[i][0] 是因为type(ytr[i]) => numpy.ndarray
  19. #             print(np.array(min_lables))
  20.             Ypred[i] = np.argmax(np.bincount(min_lables))
  21. #             print(i)
  22.         return Ypred
  1. nn = kNN(Xtr,ytr)
  2. N = 10
  3. Yte_predict = nn.forward(Xte[0:N,:],5)  #取100个数据验证
  4. print('ACC:%f'%(np.mean(Yte_predict==yte[0:N])))
  1. ACC:0.500000
  1. !jupyter nbconvert --to markdown k-NN.ipynb
  1. [NbConvertApp] Converting notebook k-NN.ipynb to markdown
  2. [NbConvertApp] Support files will be in k-NN_files\
  3. [NbConvertApp] Making directory k-NN_files
  4. [NbConvertApp] Writing 4017 bytes to k-NN.md