KNN分類原理

TF的KNN程式碼

def load_mnist_data(filename,isbatch=0,train_nums=1000,test_nums=200):
    from tensorflow.examples.tutorials.mnist import input_data
    mnist = input_data.read_data_sets(filename, one_hot=True)
       #2、批量獲取樣本
    if isbatch==1:
        X_train,Y_train = mnist.train.next_batch(train_nums)
        X_test,Y_test = mnist.test.next_batch(test_nums)
        return
 
 X_train,Y_train,X_test,Y_test
    else:
        #1、獲取全部樣本
        X_train = mnist.train.images[0:20000]   #[1:10]
        X_test = mnist.test.images[0:300] 
        Y_train = mnist.train.labels[0:20000] 
        Y_test = mnist.test.labels[0:300] 
        return X_train,Y_train,X_test,Y_test

def KNN_Classifier(X_train,Y_train,X_test,Y_test,K=5
 
,dims=784,dist_metric='L1'):
    # 計算圖輸入佔位符
    xs = tf.placeholder(tf.float32,[None,dims])
    xst  = tf.placeholder(tf.float32,[dims])  
    # 使用 L1 距離進行最近鄰計算
    # L1：dist = sum(|X1-X2|)  或 L2：dist=sqrt(sum(|X1-X2|^2))
    dist = tf.reduce_sum(tf.abs(tf.add(xs,tf.negative(xst))),
                         reduction_indices=1
 
)
    #或dist = tf.reduce_sum(tf.abs(tf.subtract(xtrain, xtest))), axis=1)

    # 預測: 獲得前K個最小距離的索引，用於與正確標籤比較
    #index = tf.arg_min(dist,0)
    if K is None:
        dim = tf.size(tf.shape(dist))   #矩陣元素的個數
        if dim==1:
            K = tf.shape(dist)[0]
        elif dim==2:
            K = tf.shape(dist)[1]
        else:
            K = tf.shape(dist)[-1]
    # 從小到大排序，取前K個最小的
    value,index = tf.nn.top_k(-dist,k=K)
    value = -value


    # 初始化所有變數
    init = tf.global_variables_initializer()    

    #定義一個正確率計算器
    Accuracy = 0 

    #執行會話
    with tf.Session() as sess:
        sess.run(init) 
        # 只能迴圈地對測試樣本進行預測
        for i in range(len(X_test)):    
            idx = sess.run(index,feed_dict={xs:X_train,xst:X_test[i,:]})

            # 計算預測標籤和正確標籤用於比較
            Klabels = np.argmax(Y_train[idx],axis=1)   #統計K行01標籤中為1的下標
            Predict_label = np.argmax(np.bincount(Klabels)) #統計下標陣列中出現次數最多的值
            #print(Y_train[idx],'\n',Klabels,'\n',Predict_label)

            True_label = np.argmax(Y_test[i])

            print("Test Sample",i,"Prediction label:",Predict_label,
                  "True Class label:",True_label)

            # 計算精確度
            if Predict_label == True_label:
                Accuracy +=1
        print("Accuracy=",Accuracy/len(X_test))    

    return Accuracy    

if __name__ == '__main__':  
    X_train,Y_train,X_test,Y_test = load_mnist_data("MNIST_data",isbatch=0,train_nums=1000,test_nums=200)    
    Accuracy =  KNN_Classifier(X_train,Y_train,X_test,Y_test,K=5,dims=784,dist_metric='L1') 

實驗結果比較：
固定選擇訓練樣本前20000個，測試樣本前300個（不是採用batch隨機批量獲取樣本）的實驗結果如下：