MNIST手写数字识别-Python+纯numpy实现

1.X版本说明

此版本基于老师所给的MATLAB代码改写，并对隐藏层大小、数据读取方式进行修改，提高收敛速度以及准确率

一、数据处理

import scipy.io as scio
## import torch
import numpy as np
## import torchvision
## 使用GPU资源
## device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

1. 直接读取mat文件

path=''
data = scio.loadmat(path)

## 查看数据类型
data

{'__header__': b'MATLAB 5.0 MAT-file, Platform: MACI64, Created on: Wed Nov 20 21:23:10 2019',
 '__version__': '1.0',
 '__globals__': [],
 'test_ima': array([[0, 0, 0, ..., 0, 0, 0],
        [0, 0, 0, ..., 0, 0, 0],
        [0, 0, 0, ..., 0, 0, 0],
        ...,
        [0, 0, 0, ..., 0, 0, 0],
        [0, 0, 0, ..., 0, 0, 0],
        [0, 0, 0, ..., 0, 0, 0]], dtype=uint8),
 'test_lab': array([[7],
        [2],
        [1],
        ...,
        [4],
        [5],
        [6]], dtype=uint8),
 'train_ima': array([[0, 0, 0, ..., 0, 0, 0],
        [0, 0, 0, ..., 0, 0, 0],
        [0, 0, 0, ..., 0, 0, 0],
        ...,
        [0, 0, 0, ..., 0, 0, 0],
        [0, 0, 0, ..., 0, 0, 0],
        [0, 0, 0, ..., 0, 0, 0]], dtype=uint8),
 'train_lab': array([[5],
        [0],
        [4],
        ...,
        [5],
        [6],
        [8]], dtype=uint8)}

2. 将数据集转换为train、valid、test

## 制作训练集
train_ima=np.array(data['train_ima'])
train_lab=np.array(data['train_lab'])
length=len(train_lab)
train_feature=train_ima[:,:int(length*0.8)]
train_label=train_lab[:int(length*0.8):]
## 制作测试集
valid_feature=train_ima[:,int(length*0.8):]
valid_label=train_lab[int(length*0.8):]
## 制作检验集
test_feature=np.array(data['test_ima'])
test_label=np.array(data['test_lab'])

train_sum=len(train_label)
valid_sum=len(valid_label)
test_sum=len(test_label)
## 查看大小，确保划分的正确性
print(train_sum)

48000

3. 对MATLAB代码进行复现

import numpy as np
 
## 初始化模型参数
label_num=10
num_inputs = 784
num_hidden = 20
num_outputs = 10

W1 = np.random.normal(0, 0.01, (num_inputs, num_hidden))
b1 = np.zeros((num_hidden,1))
W2 = np.random.normal(0, 0.01, (num_hidden, num_outputs))
b2 = np.zeros((num_outputs,1))
## 设置梯度运算
rate1 = 0.05;
rate2 = 0.05;                #置学习率

## 定义需要用到的函数
def sigmord(x):
    output =1./(1+np.exp(-x))
    return output

def forward(x,w,b):
    y=np.dot(x.reshape(1,-1),w).T + b
    return y

## 训练与检测
## 在这一部分中，针对原代码进行了两项修改：

## 用作三层神经网络的暂时存储变量
temp1 = np.zeros((num_hidden,1))
net = temp1
temp2 = np.zeros((num_outputs,1))
z = temp2

batch_size=720
epochs=1

for epoch in range(epochs):
    print('epoch:',epoch)
    for num in range(int(train_sum/batch_size)):
        ## 随机读取数据
        rand=np.random.randint(0,train_sum-1,batch_size)
        ## 训练
        for i in rand:
            label = np.zeros((label_num,1))
            label[int(train_label[i])]=1

            ## 前向计算
            temp1=forward(train_feature[:,i],W1,b1)
            #temp1=np.dot(train_feature[:,i].reshape(1,784),W1).T + b1
            net = sigmord(temp1).T
            temp2=forward(net,W2,b2)
            #temp2 = np.dot(net , W2).T + b2
            z = sigmord(temp2)

            ## 误差计算
            error = label - z
            deltaZ = (error*z*(1-z))
            deltaNet =(net*(1-net)).T*np.dot(W2,deltaZ)

            ## 参数更新
            for j in range(num_outputs):
                W2[:,j] = W2[:,j] + rate2*deltaZ[j]*net
            for j in range(num_hidden):
                W1[:,j] = W1[:,j] + (rate1*deltaNet[j]*train_feature[:,i])
            b2 = b2 + rate2*deltaZ
            b1 = b1 + rate1*deltaNet
            
        ## 检测
        valid_sum=60000*0.2
        count = 0
        if num%10==0:
            for i in range(int(valid_sum)):
                temp1=np.dot(valid_feature[:,i].reshape(1,784),W1).T + b1
                net = sigmord(temp1).T

                temp2 = np.dot(net , W2).T + b2
                z = sigmord(temp2)

                inx =np.argmax(z)
                inx=inx
                if inx == valid_label[i]:
                    count=count+1 
            correctRate=count/valid_sum
            print('step: ',num,'acc：',correctRate)
        

epoch: 0
step:  0 acc： 0.27841666666666665
step:  10 acc： 0.8729166666666667
step:  20 acc： 0.8755
step:  30 acc： 0.8913333333333333
step:  40 acc： 0.8919166666666667
step:  50 acc： 0.8993333333333333
step:  60 acc： 0.89525

## test
count=0 
for i in range(test_sum):
    temp1=np.dot(test_feature[:,i].reshape(1,784),W1).T + b1
    #print(temp1.shape[0])
    net = sigmord(temp1).T
    #print(net.shape[1])

    temp2 = np.dot(net , W2).T + b2
    z = sigmord(temp2)
    
    inx =np.argmax(z)
    inx=inx
    if inx == test_label[i]:
        count=count+1 
correctRate=count/test_sum
print('the test\'s acc：',correctRate)

the test's acc： 0.9013

## 保存参数
import csv
file = open('mnist_4.0.csv','w',encoding='utf-8')
# 2. 基于文件对象构建 csv写入对象
csv_writer = csv.writer(file)
# 3. 构建列表头
csv_writer.writerow(["batch_size","epochs","w1","w2","b1","b2","num_hidden"])
csv_writer.writerow([512,10,W1,W2,b1,b2,num_hidden])
# 5. 关闭文件
file.close()

对个别参数进行了“调乱”，直接跑无法得到相应精度，自行完善
欢迎私信或评论区讨论