KNN实现CIFAR-10数据集识别

cs231n链接：http://cs231n.github.io/linear-classify/，

训练集链接：https://download.csdn.net/download/fanzonghao/10592049

KNN缺点：每个测试样本都要循环一遍训练样本。

该数据集由5个data_batch和一个test_batch构成，测试代码

import pickle
import numpy as np
fo=open('./datasets/cifar-10-batches-py/data_batch_1','rb')
dict=pickle.load(fo,encoding='bytes')
print(dict)
print(dict[b'data'].shape)

print(dict[b'labels'])
print(len(dict[b'labels']))

print(dict[b'filenames'])
print(len(dict[b'filenames']))
fo.close()

可看出，一个data_batch由10000个，32×32×3大小的图片组成，5个就是50000个，test_batch也是10000张，故有50000张训练样本，10000张测试样本。

将5个训练集合成一个代码如下：

import pickle
import numpy as np

"""
解压数据集
"""
def unpickle(file):
    fo=open(file,'rb')
    dict=pickle.load(fo,encoding='bytes')
    fo.close()
    return dict
"""
5个data_batch和1个test_batch合成一个
"""
def load_cifar10(file):
    data_train = []
    label_train=[]
    #融合训练集
    for i in range(1,6):
        dic=unpickle(file+'data_batch_'+str(i))
        for i_data in dic[b'data']:
            data_train.append(i_data)
        for i_label in dic[b'labels']:
            label_train.append(i_label)
    # print(np.array(data_train).shape)
    # print(np.array(label_train).shape)
    # 融合测试集
    data_test=[]
    label_test=[]
    dic = unpickle(file + 'test_batch')
    for i_data in dic[b'data']:
        data_test.append(i_data)
    for i_label in dic[b'labels']:
        label_test.append(i_label)
    # print(np.array(data_test).shape)
    # print(np.array(label_test).shape)
    return (np.array(data_train),np.array(label_train),np.array(data_test),np.array(label_test))
path='./datasets/cifar-10-batches-py/'
# #(50000,3072) (50000,) (10000,3072) (10000,)
(data_train,label_train,data_test,label_test)=load_cifar10(path)
print(data_train.shape)
print(label_train.shape)
print(label_train[:10])
print(data_test.shape)
print(label_test.shape)

KNN代码：

import numpy as np
import pickle
"""
程序功能：k近邻实现cifar10上的样本分类 精度低 测试时间长
"""
#输入训练集和测试集
#解压数据集
def unpickle(file):
    fo=open(file,'rb')
    dict=pickle.load(fo,encoding='bytes')
    print(dict)
    fo.close()
    return dict
#融合训练集和测试集作为输出总样本
def load_cifar10(file):
    data_train = []
    label_train=[]
    #融合训练集
    for i in range(1,6):
        dic=unpickle(file+'data_batch_'+str(i))
        for i_data in dic[b'data']:
            data_train.append(i_data)
        for i_label in dic[b'labels']:
            label_train.append(i_label)
    # print(np.array(data_train).shape)
    # print(np.array(label_train).shape)
    # 融合测试集
    data_test=[]
    label_test=[]
    dic = unpickle(file + 'test_batch')
    for i_data in dic[b'data']:
        data_test.append(i_data)
    for i_label in dic[b'labels']:
        label_test.append(i_label)
    # print(np.array(data_test).shape)
    # print(np.array(label_test).shape)
    return (np.array(data_train),np.array(label_train),np.array(data_test),np.array(label_test))
path='./datasets/cifar-10-batches-py/'
#(50000,3072) (50000,) (10000,3072) (10000,)
(data_train,label_train,data_test,label_test)=load_cifar10(path)
#print(label_train)
print(data_train.shape,label_train.shape,data_test.shape,label_test.shape)
#print(data_test.shape[0])

"""
实现最近邻的预测
"""
class NearestNeighbor:
    def __init__(self):
        pass
    def train(self,X,y):
        self.Xtr=X
        self.ytr=y
    def predict(self,X):
        num_test=X.shape[0]
        self.X=X
        Y_pred=np.zeros(num_test,dtype=self.ytr.dtype)
        for i in range(num_test):
            distances=np.sum(np.abs(self.Xtr-self.X[i,:]),axis=1)
            #distances=np.sqrt(np.sum(np.square(self.Xtr-self.X[i,:]),axis=1))
            min_index=np.argmin(distances)
            Y_pred[i]=self.ytr[min_index]
            if i%100==0:
                print('运行到{}步'.format(i))
        return Y_pred
nn=NearestNeighbor()
nn.train(data_train,label_train)
Y_pred=nn.predict(data_test)
accuarcy=np.mean(label_test==Y_pred)
print('accuarcy={}'.format(accuarcy))

打印结果：精度不高，后面引入神经网络

SVM损失函数：

loss.py

import numpy as np
"""
程序功能：利用SVM代价函数实现损失值的积累
"""
def L(X,y,W):
    #X [3073,50000]
    #y 一维（50000,）
    #W [10,3073]
    delta=1.0
    scores=np.dot(W,X)
    #print(y)
    #对应训练样本的输出y
    #print(scores[y, np.arange(scores.shape[1])])
    #(10,50000)
    #SVM函数
    margins=np.maximum(0,scores-scores[y, np.arange(scores.shape[1])]+delta)
    #print('margins.shape={}'.format(margins.shape))
    margins[y,np.arange(scores.shape[1])]=0
    loss=np.mean(margins)
    return loss

optimizer_grand.py

import numpy as np
import pickle
import loss
"""
函数功能：利用随机搜索和局部随机搜索来获取W和b采用SVM损失函数 获取最佳的W和b
"""
#输入训练集和测试集
#解压数据集
def unpickle(file):
    fo=open(file,'rb')
    dict=pickle.load(fo,encoding='bytes')
    fo.close()
    return dict
#融合训练集和测试集作为输出总样本
def load_cifar10(file):
    data_train = []
    label_train=[]
    #融合训练集
    for i in range(1,6):
        dic=unpickle(file+'data_batch_'+str(i))
        for i_data in dic[b'data']:
            data_train.append(i_data)
        for i_label in dic[b'labels']:
            label_train.append(i_label)
    # print(np.array(data_train).shape)
    # print(np.array(label_train).shape)
    # 融合测试集
    data_test=[]
    label_test=[]
    dic = unpickle(file + 'test_batch')
    for i_data in dic[b'data']:
        data_test.append(i_data)
    for i_label in dic[b'labels']:
        label_test.append(i_label)
    # print(np.array(data_test).shape)
    # print(np.array(label_test).shape)
    return (np.array(data_train),np.array(label_train),np.array(data_test),np.array(label_test))
path='./datasets/cifar-10-batches-py/'
#(50000,3072) (50000,) (10000,3072) (10000,)
(data_train,label_train,data_test,label_test)=load_cifar10(path)
#print(label_train)
print(data_train.shape,label_train.shape,data_test.shape,label_test.shape)
#(3072,50000)
train_data=np.transpose(data_train)
#增加一行 处理偏置值
bias=np.ones((1,train_data.shape[1]))
#(3073,50000)
train_data=np.vstack((train_data,bias))
print(train_data.shape)
#随机选择最佳的权值 输出最佳的W
def random_search():
    bestloss=float('inf')
    for number in range(1000):
        # 随机搜索  权值随机更新 选出比较好的
        W = np.random.randn(10, 3073) * 0.0001
        # 计算损失值
        lost = loss.L(train_data, label_train, W)
        if lost<bestloss:
            bestloss=lost
            bestW=W
        if number%100==0:
            print('number={},the lost={},bestloss={}'.format(number,lost,bestloss))
    return bestW
#调用随机产生的最佳权值产生预测值与标签值算精确度
def random_search_accu():
        bestW=random_search()
        #(10,50000)
        scores=np.dot(bestW,train_data)
        #找出每列分数最大值的索引
        Y_predict=np.argmax(scores,axis=0)
        accurarcy=np.mean(Y_predict==label_train)
        print('accurarcy={}'.format(accurarcy))
def random_local_search():
    W = np.random.randn(10, 3073) * 0.001
    bestloss=float('inf')
    for number in range(1000):
        # 随机搜索  权值随机更新 选出比较好的
        step_size=0.0001
        W_try=W+np.random.randn(10, 3073) * step_size
        # 计算损失值
        lost = loss.L(train_data, label_train, W_try)
        if lost<bestloss:
            bestloss=lost
            bestW=W_try
        if number%100==0:
            print('number={},the lost={},bestloss={}'.format(number,lost,bestloss))
    return bestW
#调用随机产生的最佳权值产生预测值与标签值算精确度
def random_local_search_accu():
        bestW=random_local_search()
        #(10,50000)
        scores=np.dot(bestW,train_data)
        #找出每列分数最大值的索引
        Y_predict=np.argmax(scores,axis=0)
        accurarcy=np.mean(Y_predict==label_train)
        print('accurarcy={}'.format(accurarcy))
if __name__ == '__main__':
    #随机搜索
    # random_search_accu()
    #局部随机搜索
    random_local_search_accu()
    #梯度跟随

随机最佳权重的打印结果：

在迭代过程中，权重还变化的结果