cifar数据集的基本情况：
该数据集共有60000张彩色图像，这些图像是32*32，分为10个类，每类6000张图。这里面有50000张用于训练，构成了5个训练批，每一批10000张图；另外10000用于测试，单独构成一批。测试批的数据里，取自10类中的每一类，每一类随机取1000张。抽剩下的就随机排列组成了训练批。注意一个训练批中的各类图像并不一定数量相同，总的来看训练批，每一类都有5000张图。
我们可以看到下载后的数据集文件夹包括：
其中data_batch就是用于训练的数据集，一共有五个；test_batch就是用于测试的数据集，一共有一个；batchs.meta其中包含的信息是类别名称，下面我们就仔细看看各个文件包含的具体内容：
各文件详细内容
- 1. batch.meta
import pickle
def unpickle(file):
with open(file, 'rb') as fo:
dict = pickle.load(fo, encoding='latin-1')
return dict
cc=unpickle("./dataset/cifar-10/cifar-10-batches-py/batches.meta") #路径需要自行修改
print(cc)
运行结果：
可以看到：batch应该是一个batch包含的图片数量，label_names应该是类别名字，num_vis应该是32*32（图片大小）*3（RGB三通道），个人理解为描述一副图像所需要的值。
- 2. data_batch_1
debug结果：代码不加赘述，上面的改下路径就行
可以看到，其中包含batch_label，labels，data，filenames四项，其中batch_label就是第几个batch，labels就是第几类，data就是上面3072对应的具体值啦（不止一个3017，因为不止一幅图，大家自行理解哈），其实就是RGB值，filename就是图片的名字，是不是挺简单的。test_batch也是一样的样的道理。
接下来就是具体自己的数据集转换的代码：
# -*- coding: utf-8 -*-
"""
@author: zhangjiaqing 有借鉴
"""
import numpy as np
import chardet
from PIL import Image
import operator
from os import listdir
import sys
import pickle
import random
def unpickle(file):
with open(file, 'rb') as fo:
dict = pickle.load(fo, encoding='latin-1')
return dict
#cc=unpickle("./dataset/cifar-10/cifar-10-batches-py/data_batch_1")
#print(cc)
data={}
list1=[]
list2=[]
list3=[]
#将图片转化为32*32的三通道图片
def img_tra():
for k in range(0,num):
currentpath=folder+"/"+imglist[k]
im=Image.open(currentpath)
#width=im.size[0]
#height=im.size[1]
x_s=32
y_s=32
out = im.resize((x_s,y_s),Image.ANTIALIAS)
out.save(folder_ad+"/"+str(imglist[k]))
def addWord(theIndex,word,adder):
theIndex.setdefault(word,[]).append(adder)
def seplabel(fname):
filestr=fname.split(".")[0]
label=int(filestr.split("_")[0]) #图片的命名 _前面是类别
return label
def mkcf():
global data
global list1
global list2
global list3
for k in range(0,num):
currentpath=folder_ad+"/"+imglist[k]
im=Image.open(currentpath)
with open(binpath, 'a') as f:
for i in range (0,32):
for j in range (0,32):
cl=im.getpixel((i,j))
list1.append(cl[0]) #R
for i in range (0,32):
for j in range (0,32):
cl=im.getpixel((i,j))
#with open(binpath, 'a') as f:
#mid=str(cl[1])
#f.write(mid)
list1.append(cl[1]) #G
for i in range (0,32):
for j in range (0,32):
cl=im.getpixel((i,j))
list1.append(cl[2]) ##B
list2.append(list1)
list1=[]
f.close()
print("image"+str(k+1)+"saved.")
list3.append(imglist[k]) #name of pictures
arr2=np.array(list2,dtype=np.uint8)
data['batch_label']='training batch 5 of 5' #training batch 1 of 5 testing batch 1 of 1
data.setdefault('labels',label)
data.setdefault('data',arr2)
data.setdefault('filenames',list3)
output = open(binpath, 'wb')
pickle.dump(data, output)
output.close()
folder="./cloud/train_batch_5" #自己图片的路径 train_batch_5 test
folder_ad="./cloud/train_batch_5_ad" #将图片转化为32*32的三通道图片的路径 train_batch_5_ad test_ad
imglist=listdir(folder) #这里原作者好像写错了，我自行修改了，目测现在是对的
num=len(imglist)
img_tra()
label=[]
for i in range (0,num):
label.append(seplabel(imglist[i]))
binpath="./dataset/cloud/cloud-5-batches-py/data_batch_5" #保存的路径 data_batch_5 test_batch
print(binpath)
mkcf()
给大家看一下转的结果
解释一下为什么这里少了batch.meta，感觉这里的信息没什么用，如果是自己的数据集，可以直接写一句代码就行：
classes = ['A', 'B', 'C', 'D', 'E']
注意自己数据集图片的命名：‘_‘前面是类别
代码改进：训练测试集路径不用一直修改
# -*- coding: utf-8 -*-
"""
@author: zhangjiaqing
"""
import numpy as np
import chardet
from PIL import Image
import operator
from os import listdir
import sys
import pickle
import random
from skimage.util.shape import view_as_windows
def unpickle(file):
with open(file, 'rb') as fo:
dict = pickle.load(fo, encoding='latin-1')
return dict
#cc=unpickle("./dataset/cloud/cloud-5-batches-py/test_batch")
#cc=unpickle("./dataset/cifar-10/cifar-10-batches-py/data_batch_1")
#print(cc)
data = {}
list1 = []
list2 = []
list3 = []
label = []
size = 64
def split():
global label
for k in range(0,num):
currentpath=folder+imglist[k]
img=Image.open(currentpath)
img = np.array(img)
img_block_2 = view_as_windows(img, (size, size, 3), step=size)
hang = img_block_2.shape[0]
lie = img_block_2.shape[1]
label=[]
img_block = np.zeros((size,size,3))
for i in range (hang):
for j in range(lie):
label.append(random.randint(0, 4))
img_block = img_block_2[i,j,0,:,:,:]
image = Image.fromarray(img_block.astype('uint8'))
image.save(folder_ad + '%s_%d_%d.jpg'%(str(imglist[k]),i*lie+j,label[i*lie+j]))
#out.save(folder_ad+"/"+str(imglist[k]))
def addWord(theIndex,word,adder):
theIndex.setdefault(word,[]).append(adder)
def mkcf():
global data
global list1
global list2
global list3
global train
for k in range(0,number):
currentpath=folder_ad+imagelist[k]
im=Image.open(currentpath)
with open(binpath, 'a') as f:
for i in range (0,size):
for j in range (0,size):
cl=im.getpixel((i,j))
list1.append(cl[0]) #R
for i in range (0,size):
for j in range (0,size):
cl=im.getpixel((i,j))
list1.append(cl[1]) #G
for i in range (0,size):
for j in range (0,size):
cl=im.getpixel((i,j))
list1.append(cl[2]) ##B
list2.append(list1)
list1=[]
f.close()
print("image"+str(k+1)+"saved.")
list3.append(imagelist[k]) #name of pictures
arr2=np.array(list2,dtype=np.uint8)
if train:
data['batch_label']='training batch 1 of 1' #training batch 1 of 5 testing batch 1 of 1
else:
data['batch_label']='testing batch 1 of 1' #training batch 1 of 5 testing batch 1 of 1
data.setdefault('labels',label)
data.setdefault('data',arr2)
data.setdefault('filenames',list3)
output = open(binpath, 'wb')
pickle.dump(data, output)
output.close()
train = False #true就是训练集路径 #false就是测试集路径
if train:
folder="./cloud/train_batch_1/" # train_batch_5 test
folder_ad="./cloud/train_batch_1_ad/" #将图片转化为32*32的三通道图片 train_batch_5_ad test_ad
binpath="./dataset/cloud/cloud-5-batches-py/data_batch_1" # data_batch_5 test_batch
else:
folder="./cloud/test/" # train_batch_5 test
folder_ad="./cloud/test_ad/" #将图片转化为32*32的三通道图片 train_batch_5_ad test_ad
binpath="./dataset/cloud/cloud-5-batches-py/test_batch" # data_batch_5 test_batch
imglist=listdir(folder)
num=len(imglist)
split()
imagelist=listdir(folder_ad)
number=len(imagelist)
mkcf()
print('the work is finished!')
原创文章，作者：ItWorker，如若转载，请注明出处：https://blog.ytso.com/tech/pnotes/279260.html