Python不借助深度学习框架完成MNIST手写数字识别

2024-07-15 19:37| 来源: 网络整理| 查看: 265

这学期选修了一门《深度学习导论》，本来以为是门水课（指最后大作业蹭框架就能完成或者直接交个啥论文的那种），结果没想到最后要求不使用框架完成一个卷积神经网络的搭建与训练，要求训练限时七分钟，训练好后的分类器在MNIST训练测试集的识别率在97%以上。当然大部分代码老师都以MATLAB形式给出了，我自己照葫芦画瓢做了个python版本，仅供参考。 MNIST数据来源：THE MNIST DATABASE

# -*- coding:utf-8 -*- """ @author:UESTC_Sicca @file:V12-手写数字识别.py @time:2020/11/9 """ import numpy as np import struct import matplotlib.pyplot as plt from scipy.signal import correlate2d import time # 训练集文件 train_images_idx3_ubyte_file = '/Users/violet/Documents/计算机视觉/MNIST_Database/train-images.idx3-ubyte' # 训练集标签文件 train_labels_idx1_ubyte_file = '/Users/violet/Documents/计算机视觉/MNIST_Database/train-labels.idx1-ubyte' # 测试集文件 test_images_idx3_ubyte_file = '/Users/violet/Documents/计算机视觉/MNIST_Database/t10k-images.idx3-ubyte' # 测试集标签文件 test_labels_idx1_ubyte_file = '/Users/violet/Documents/计算机视觉/MNIST_Database/t10k-labels.idx1-ubyte' def decode_idx3_ubyte(idx3_ubyte_file): """ 解析idx3文件的通用函数 :param idx3_ubyte_file: idx3文件路径 :return: 数据集 """ # 读取二进制数据 bin_data = open(idx3_ubyte_file, 'rb').read() # 解析文件头信息，依次为魔数、图片数量、每张图片高、每张图片宽 offset = 0 fmt_header = '>iiii' #因为数据结构中前4行的数据类型都是32位整型，所以采用i格式，但我们需要读取前4行数据，所以需要4个i。我们后面会看到标签集中，只使用2个ii。 magic_number, num_images, num_rows, num_cols = struct.unpack_from(fmt_header, bin_data, offset) print('魔数:%d, 图片数量: %d张, 图片大小: %d*%d' % (magic_number, num_images, num_rows, num_cols)) # 解析数据集 image_size = num_rows * num_cols offset += struct.calcsize(fmt_header) #获得数据在缓存中的指针位置，从前面介绍的数据结构可以看出，读取了前4行之后，指针位置（即偏移位置offset）指向0016。 print(offset) fmt_image = '>' + str(image_size) + 'B' #图像数据像素值的类型为unsigned char型，对应的format格式为B。这里还有加上图像大小784，是为了读取784个B格式数据，如果没有则只会读取一个值（即一副图像中的一个像素值） print(fmt_image,offset,struct.calcsize(fmt_image)) images = np.empty((num_images, num_rows, num_cols)) #plt.figure() for i in range(num_images): if (i + 1) % 10000 == 0: print('已解析 %d' % (i + 1) + '张') print(offset) images[i] = np.array(struct.unpack_from(fmt_image, bin_data, offset)).reshape((num_rows, num_cols)) #print(images[i]) offset += struct.calcsize(fmt_image) # plt.imshow(images[i],'gray') # plt.pause(0.00001) # plt.show() #plt.show() return images def decode_idx1_ubyte(idx1_ubyte_file): """ 解析idx1文件的通用函数 :param idx1_ubyte_file: idx1文件路径 :return: 数据集 """ # 读取二进制数据 bin_data = open(idx1_ubyte_file, 'rb').read() # 解析文件头信息，依次为模样数和标签数 offset = 0 fmt_header = '>ii' magic_number, num_images = struct.unpack_from(fmt_header, bin_data, offset) print('魔数:%d, 图片数量: %d张' % (magic_number, num_images)) # 解析数据集 offset += struct.calcsize(fmt_header) fmt_image = '>B' labels = np.empty(num_images) for i in range(num_images): if (i + 1) % 10000 == 0: print ('已解析 %d' % (i + 1) + '张') labels[i] = struct.unpack_from(fmt_image, bin_data, offset)[0] offset += struct.calcsize(fmt_image) return labels def load_train_images(idx_ubyte_file=train_images_idx3_ubyte_file): """ TRAINING SET IMAGE FILE (train-images-idx3-ubyte): [offset] [type] [value] [description] 0000 32 bit integer 0x00000803(2051) magic number 0004 32 bit integer 60000 number of images 0008 32 bit integer 28 number of rows 0012 32 bit integer 28 number of columns 0016 unsigned byte ?? pixel 0017 unsigned byte ?? pixel ........ xxxx unsigned byte ?? pixel Pixels are organized row-wise. Pixel values are 0 to 255. 0 means background (white), 255 means foreground (black). :param idx_ubyte_file: idx文件路径 :return: n*row*col维np.array对象，n为图片数量 """ return decode_idx3_ubyte(idx_ubyte_file) def load_train_labels(idx_ubyte_file=train_labels_idx1_ubyte_file): """ TRAINING SET LABEL FILE (train-labels-idx1-ubyte): [offset] [type] [value] [description] 0000 32 bit integer 0x00000801(2049) magic number (MSB first) 0004 32 bit integer 60000 number of items 0008 unsigned byte ?? label 0009 unsigned byte ?? label ........ xxxx unsigned byte ?? label The labels values are 0 to 9. :param idx_ubyte_file: idx文件路径 :return: n*1维np.array对象，n为图片数量 """ return decode_idx1_ubyte(idx_ubyte_file) def load_test_images(idx_ubyte_file=test_images_idx3_ubyte_file): """ TEST SET IMAGE FILE (t10k-images-idx3-ubyte): [offset] [type] [value] [description] 0000 32 bit integer 0x00000803(2051) magic number 0004 32 bit integer 10000 number of images 0008 32 bit integer 28 number of rows 0012 32 bit integer 28 number of columns 0016 unsigned byte ?? pixel 0017 unsigned byte ?? pixel ........ xxxx unsigned byte ?? pixel Pixels are organized row-wise. Pixel values are 0 to 255. 0 means background (white), 255 means foreground (black). :param idx_ubyte_file: idx文件路径 :return: n*row*col维np.array对象，n为图片数量 """ return decode_idx3_ubyte(idx_ubyte_file) def load_test_labels(idx_ubyte_file=test_labels_idx1_ubyte_file): """ TEST SET LABEL FILE (t10k-labels-idx1-ubyte): [offset] [type] [value] [description] 0000 32 bit integer 0x00000801(2049) magic number (MSB first) 0004 32 bit integer 10000 number of items 0008 unsigned byte ?? label 0009 unsigned byte ?? label ........ xxxx unsigned byte ?? label The labels values are 0 to 9. :param idx_ubyte_file: idx文件路径 :return: n*1维np.array对象，n为图片数量 """ return decode_idx1_ubyte(idx_ubyte_file) def load_data(): ''' :return: 依次为训练样本及其标签，测试样本及其标签 ''' train_images = load_train_images() train_labels = load_train_labels() test_images = load_test_images() test_labels = load_test_labels() train_images = train_images/255 test_images = test_images/255 # # # # 查看前十个数据及其标签以读取是否正确 # plt.figure(figsize=(2,5)) # for i in range(10): # plt.subplot(2, 5, i + 1) # plt.xticks([]) # plt.yticks([]) # plt.imshow(train_images[i]) # plt.xlabel(train_labels[i]) # print(train_labels[i]) # plt.show() return train_images,train_labels,test_images,test_labels def Softmax(X): ex = np.exp(X) return ex/np.sum(ex) def one_hot_label(label): lab = np.zeros((label.size, 10, 1)) for i in range(len(lab)): lab[i][int(label[i])] = 1 return lab train_images,train_labels,test_images,test_labels = load_data() train_labels = one_hot_label(train_labels) test_labels = one_hot_label(test_labels) time_start = time.time() W1 = np.random.randn(9,9,20) W3 = (2*np.random.rand(100,2000)-1)/20 W4 = (2*np.random.rand(10,100)-1)/10 alpha=0.01 # S_dW1,S_dW3,S_dW4 = 0,0,0 for i in range(len(train_images)): V1 = np.zeros((20,20,20)) for k in range(20): V1[:,:,k] = correlate2d(train_images[i],W1[:,:,k],'valid') Y1 = np.maximum(0,V1) Y2 = (Y1[::2,::2] + Y1[1::2,::2] + Y1[::2,1::2] + Y1[1::2,1::2])/4 y2 = np.reshape(Y2,[2000,1]) V3 = np.dot(W3,y2) Y3 = np.maximum(0,V3) V = np.dot(W4,Y3) # if i == 11:print(Y) Y = Softmax(V) d = train_labels[i] e = d-Y delta = e e3 = np.dot(W4.T,delta) delta3 = (V3>0)*e3 e2 = np.dot(W3.T,delta3) dW4 = np.dot(alpha*delta,Y3.T) dW3 = np.dot(alpha*delta3,y2.T) E2 = np.reshape(e2,[10,10,20]) E1 = np.zeros(np.shape(Y1)) E2_4 = E2/4 E1[::2,::2,:] = E2_4 E1[::2,1::2,:] = E2_4 E1[1::2,::2,:] = E2_4 E1[1::2,1::2,:] = E2_4 delta1 = (V1>0)*E1 dW1 = np.zeros(np.shape(W1)) for k in range(20): dW1[:,:,k] = np.dot(alpha,correlate2d(train_images[i],delta1[:,:,k],'valid')) # if (i+1)%100 == 0: W1 += dW1 W3 += dW3 W4 += dW4 # S_dW1,S_dW2,S_dW3 = 0,0,0 # else: # S_dW1 += dW1 # S_dW3 += dW3 # S_dW4 += dW4 print("已训练"+str((i+1)/len(train_images)*100)+"%") time_end = time.time() s = 0 for i in range(len(test_images)): V1 = np.zeros((20, 20, 20)) for k in range(20): V1[:, :, k]=correlate2d(test_images[i], W1[:, :, k], 'valid') Y1 = np.maximum(0, V1) Y2 = (Y1[::2, ::2] + Y1[1::2, ::2] + Y1[::2, 1::2] + Y1[1::2, 1::2]) / 4 y2 = np.reshape(Y2, [2000, 1]) V3 = np.dot(W3, y2) Y3 = np.maximum(0, V3) V = np.dot(W4, Y3) Y = Softmax(V) Y_pred = np.argmax(Y) Y_test = np.argmax(test_labels[i]) if Y_pred != Y_test:s+=1 print("训练用时为",time_end-time_start,"秒") print("正确率为"+str((1-s/len(test_labels))*100)+"%")

其中load_data是参考了网上的开源代码完成的，是SGD单轮训练，其他的训练方法暂时不想搞了，等以后再说吧。结果大概是四分多钟能有97%出头的准确率。

【本文地址】

公司简介

联系我们