linux tenserflow add traning picture dataset to TFRecord

import os
import tensorflow as tf
from PIL import Image  #注意Image,后面会用到
import matplotlib.pyplot as plt
import numpy as np
import cv2
 

 #download pictiures and save picture as

#/tmp/dog/husky/5.jpeg
#/tmp/dog/husky/2.jpeg
#/tmp/dog/husky/1.jpeg
#/tmp/dog/husky/4.jpeg
#/tmp/dog/husky/3.jpeg
#/tmp/dog/chihuahua/5.jpeg
#/tmp/dog/chihuahua/2.jpeg
#/tmp/dog/chihuahua/1.jpeg
#/tmp/dog/chihuahua/4.jpeg
#/tmp/dog/chihuahua/3.jpeg

cwd='/tmp/dog/'
classes={'husky','chihuahua'} #人为 设定 2 类

##opencv check file name

##img = cv2.imread('/tmp/dog/husky/1.jpeg',0)
##cv2.imshow('image',img)
##cv2.waitKey(0)
##cv2.destroyAllWindows()

##cwd='D:\Python\data\dog\\'  windows dir

writer= tf.python_io.TFRecordWriter("dog_train.tfrecords") #要生成的文件
 
for index,name in enumerate(classes):
    class_path=cwd+name+'/'
    for img_name in os.listdir(class_path):
        img_path=class_path+img_name #每一个图片的地址
 
        img=Image.open(img_path)
        img= img.resize((128,128))
##        print(img_path)
##        imgcv = cv2.imread(img_path,0)
##        cv2.imshow(img_path,imgcv)
##        cv2.waitKey(0)
##        cv2.destroyAllWindows()
        img_raw=img.tobytes()#将图片转化为二进制格式
        print(index)
        example = tf.train.Example(features=tf.train.Features(feature={
            "label": tf.train.Feature(int64_list=tf.train.Int64List(value=[index])),
            'img_raw': tf.train.Feature(bytes_list=tf.train.BytesList(value=[img_raw]))
        })) #example对象对label和image数据进行封装
        writer.write(example.SerializeToString())  #序列化为字符串
 
writer.close()

print('-------1211----------')
def read_and_decode(filename): # 读入dog_train.tfrecords
    filename_queue = tf.train.string_input_producer([filename])#生成一个queue队列
 
    reader = tf.TFRecordReader()
    _, serialized_example = reader.read(filename_queue)#返回文件名和文件
    features = tf.parse_single_example(serialized_example,
                                       features={
                                           'label': tf.FixedLenFeature([], tf.int64),
                                           'img_raw' : tf.FixedLenFeature([], tf.string),
                                       })#将image数据和label取出来
 
    img = tf.decode_raw(features['img_raw'], tf.uint8)
    img = tf.reshape(img, [128, 128, 3])  #reshape为128*128的3通道图片
    img = tf.cast(img, tf.float32) * (1. / 255) - 0.5 #在流中抛出img张量
    print('img',img)
    label = tf.cast(features['label'], tf.int32) #在流中抛出label张量
    print('label',label)
    return img, label

read_and_decode('fildog_train.tfrecordsename') # 读入dog_train.tfrecords


filename_queue = tf.train.string_input_producer(["dog_train.tfrecords"]) #读入流中
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)   #返回文件名和文件
features = tf.parse_single_example(serialized_example,
                                   features={
                                       'label': tf.FixedLenFeature([], tf.int64),
                                       'img_raw' : tf.FixedLenFeature([], tf.string),
                                   })  #取出包含image和label的feature对象
image = tf.decode_raw(features['img_raw'], tf.uint8)
image = tf.reshape(image, [128, 128, 3])
label = tf.cast(features['label'], tf.int32)
with tf.Session() as sess: #开始一个会话
    init_op = tf.initialize_all_variables()
    sess.run(init_op)
    coord=tf.train.Coordinator()
    threads= tf.train.start_queue_runners(coord=coord)
    for i in range(10):
        example, l = sess.run([image,label])#在会话中取出image和label
        img=Image.fromarray(example, 'RGB')#这里Image是之前提到的
        img.save(cwd+str(i)+'_''Label_'+str(l)+'.jpg')#存下图
        img.save('/home/duan/'+str(i)+'_''Label_'+str(l)+'.jpg')#存下图片
        
        print(example, l)
    coord.request_stop()
    coord.join(threads)