61. knn实现手写数字识别

# 1）样本  2）KNNN 3） CNN
# 旧瓶装新酒：MNist数字识别的的不同 
# ->  网络 每一级 原理后代吗
# !) 样本 yann.lecun.com/exdb/mnist
# 2) knn最近邻域法  
# 本质：k个与当前相似的，其中相似性最大的数字作为结果（10图中有8个描述为1）

# 步骤： 1 load data   
# 2 距离计算 knn test and train distance 5 * 500 = 2500个距离
# 3 knn 利用距离找到k个最近的图片 test5 与 train100, 从train500中找出4张与测试图片最接近的图片
# 4 解析k个最近的图片-> parse content label(得到label)
# 5 label -> 数字
# 6 检测概率统计
# 小结： 

import numpy as np 
import random
import tensorflow as tf

# 解决mnist版本弃用警告问题
old_v = tf.logging.get_verbosity()
tf.logging.set_verbosity(tf.logging.ERROR)
from tensorflow.examples.tutorials.mnist import input_data

# 1) load data  filename: 'MNIST_data'   ont_hot:数组中一个为1其余为0
mnist = input_data.read_data_sets('MNIST_data', one_hot = True)
# 属性设置
trainNum = 55000
testNum = 10000
trainSize = 500
testSize = 5
k = 4 #   k个与样本最接近

# data分解  trainSize个  0~trainNum之间选取，replace: false表示不可重复
trainIndex = np.random.choice(trainNum, trainSize, replace = False)
testIndex = np.random.choice(testNum, testSize, replace = False)

trainData = mnist.train.images[trainIndex] # 训练图片
trainLabel = mnist.train.labels[trainIndex] # 训练标签
testData = mnist.test.images[testIndex]
testLabel = mnist.test.labels[testIndex]
print('trainData.shape = ',  trainData.shape)
print('trainLabel.shape = ', trainLabel.shape)
print('testData.shape = ', testData.shape)
print('testLabel.shape = ', testLabel.shape)
# [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.] 1:testData[0]  
# [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.] 9::testData[1]  一共五个值 ...
print('testLabel = ', testLabel)  

# tf input  建立占位符  784 ->  image
trainDataInput = tf.placeholder(shape = [None, 784], dtype = tf.float32)  # (500, 784)
trainLabelInput = tf.placeholder(shape = [None, 10], dtype = tf.float32)   # (500, 10)
testDataInput = tf.placeholder(shape = [None, 784], dtype = tf.float32)   # (5, 784)
testLabelInput = tf.placeholder(shape = [None, 10], dtype = tf.float32)    # (5, 10)

# 2) knn distance   5 * 784 -> 5 * 1 * 784 
# 扩展维度： 测试5 训练500 像素点784(维)=>  2500 *784
f1 = tf.expand_dims(testDataInput, 1)

# 作差-> 差值放入784维中 -> sum(784)
f2 = tf.subtract(trainDataInput, f1) 

# f3: 5 * 500 每一点表示某一个测试点与某一个训练点二者之间的距离之差
f3 = tf.reduce_sum(tf.abs(f2), reduction_indices = 2) # 在第二维度(即784)进行数据累加， 784个像素点之间的差值，取abs

# 3) 找出4张与测试图片最近的图片
# 取反
f4 = tf.negative(f3)

# 选取f4中最大的4个值(f3中最小的4个值)
f5, f6 = tf.nn.top_k(f4, k = 4)

# 4) 解析内容标签
# f6 index -> trainLabelInput
f7 = tf.gather(trainLabelInput, f6)

# 5) 通过label -> 数字
# num
f8 = tf.reduce_sum(f7, reduction_indices = 1)
    
# 一维，最大概率的下标
f9 = tf.argmax(f8, dimension = 1)

# 6) 通过f10统计检测概率
f10 = tf.argmax(testLabel[0 : 5], axis = 1)


with tf.Session() as sess:
    # f1 <- 五张testData待检测的手写数字图片
    p1 = sess.run(f1, feed_dict = {testDataInput : testData[0 : 5]})
    print('p1.shape = ', p1.shape)  # 维度(5, 1, 784)  [28行 * 28列]
    
    p2 = sess.run(f2, feed_dict = {trainDataInput : trainData, testDataInput : testData[0 : 5]})
    # (5, 500, 784)  p2实现对应像素作差: (1, 100)则是第二张测试图片与第一百零一张训练图片之间的距离
    print('p2.shape = ', p2.shape) 
    
    p3 = sess.run(f3, feed_dict = {trainDataInput : trainData, testDataInput : testData[0 : 5]})
    print('p3.shape = ', p3.shape) # (5, 500)
    print('p3[0, 0]', p3[0, 0])  # 距离差值knn distance 114.54903
    
    p4 = sess.run(f4, feed_dict = {trainDataInput : trainData, testDataInput : testData[0 : 5]})
    print('p4.shape = ', p4.shape) # p4.shape =  (5, 500)
    print('p4[0, 0] = ', p4[0, 0]) # p3[0, 0] 122.364716【每一次不同取决于随机生成的train与test】    p4[0, 0] =  -122.364716
    
    # p6下标 p5内容
    p5, p6 = sess.run((f5, f6), feed_dict = {trainDataInput : trainData, testDataInput : testData[0 : 5]})
    print('p5.shape = ', p5.shape) # (5, 4)
    print('p6.shape = ', p6.shape) # (5, 4)
    print('p5[0, 0] = ', p5[0, 0]) # -43.44706  内容
    print('p6[0, 0] = ', p6[0, 0]) #  390  下标
    print('p5[0] = ', p5[0]) 
    print('p6[0] = ', p6[0]) 

    
    p7 = sess.run(f7, feed_dict = {trainDataInput : trainData, testDataInput : testData[0 : 5], trainLabelInput : trainLabel})
    print('p7.shape = ', p7.shape) # (5, 4, 10)
    print('p7 = ', p7) # 5组 每组整体是一个测试图片，每行代表一个最近的测试图片，每个数字代表一个标签，一行表示一个数


    p8 = sess.run(f8, feed_dict = {trainDataInput : trainData, testDataInput : testData, trainLabelInput : trainLabel})
    print('p8.shape = ', p8.shape)     # (5, 10) 
    #   [[[0. 0. 0. 1. 0. 0. 0. 0. 0. 0.]
    #     [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.]
    #     [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.]
    #     [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.]]
    #   数值方向上的累加 -> [[0. 0. 0. 4. 0. 0. 0. 0. 0. 0.]
    print('p8 = ', p8)    
        
    p9 = sess.run(f9, feed_dict = {trainDataInput : trainData, testDataInput : testData, trainLabelInput : trainLabel})    
    print('p9.shape = ', p9.shape) #(5, ) 即是5维数据
    # p8: [[0. 0. 4. 0. 0. 0. 0. 0. 0. 0.]
    #     [0. 4. 0. 0. 0. 0. 0. 0. 0. 0.]
    #     [0. 0. 0. 0. 3. 0. 0. 0. 1. 0.]
    #     [0. 4. 0. 0. 0. 0. 0. 0. 0. 0.]
    #     [0. 4. 0. 0. 0. 0. 0. 0. 0. 0.]]
    print('p9 = ', p9)  # 最大值的下标 [2 1 4 1 1]
    
    p10 = sess.run(f10, feed_dict = {trainDataInput : trainData, testDataInput : testData, trainLabelInput : trainLabel})
    print('p10 = ', p10) #    p9 =  [3 4 2 6 4]    p10 =  [3 4 8 6 4] 相差一个
    
j = 0
for i in range(0, 5):
    if p10[i] == p9[i]:
        j = j + 1
print('ac = ', j * 100 / 5)  # 训练数据很多 测试数据很少 所以检测概率较大60 80 100%..
     
        

准确度为100%或80%……
打印值与准确度结果如下：

image.png image.png