tensorflow 读取图片

用tf.record形式处理打的图片数据比较方便，而且后面sess.run的时候不需要feeddict数据

首先，建议直接把读取的数据封装成一个类：

class data_reader():    def __init__(self,dict):        self.train_images=dict['train']        self.test_images = dict['test']        self.validation_images = dict['validation']

2.读取图像，这里以list形式存放了图像名称

def create_img_lists(img_dir,test_percent=0.0,validation_percent=0.0):    if not os.path.exists(img_dir):        print("Image directory '" + img_dir + "' not found.")        return None    train_imgs=[]    format_extends=['jpg', 'jpeg', 'JPG', 'JPEG','png']#image format     #sub_dir=[x[0] for x in os.walk(img_dir)] #dirpath, dirnames, filenames return for walk    file_lists=[]    for extends in format_extends:        file_full_path=os.path.join(img_dir,'*.'+extends) #per image  full name        file_lists.extend(glob(file_full_path)) #append    if not file_lists:        print("No files(image) found")    else:        train_imgs.extend([im for im in file_lists]) #直接就是file_lists就好了，无所谓的    random.shuffle(train_imgs)#permute    num_imgs=len(train_imgs)#if initial data_set has validation set,following need tobe changed    validation_offset=int(validation_percent*num_imgs) #以下是中规中矩的划分图像集    validation_imgs=train_imgs[:validation_offset]    test_offset = int(test_percent * num_imgs)    test_imgs = train_imgs[validation_offset:validation_offset+test_offset]    training_imgs=train_imgs[validation_offset+test_offset:]    result={        'train':training_imgs,        'test':test_imgs,        'validation':validation_imgs,    }    return result 

这里采用了cPickle方式读取，注意这里是直接将整个文件都写入.pkl中的，也就说，你的数据就是整个数据集，要划分batch您得自己动手

def read_data(data_dir):    pickle_file_name='data_reader.pkl'    pickle_file_path=os.path.join(data_dir,pickle_file_name)    if not os.path.exists(pickle_file_path):    #如果没有该文件，会生成.pkl文件        result=create_img_lists(data_dir)        print ("Training set: %d" % len(result['train']))        print ("Test set: %d" % len(result['test']))        print ("Validation set: %d" % len(result['validation']))        print ("Pickling ...")        with open(pickle_file_path,'wb') as f:            pickle.dump(result,f,pickle.HIGHEST_PROTOCOL)    with open(pickle_file_path,'rb') as f:        result=pickle.load(f)    return  result

3.这里可以考虑用tf的读入方式将数据直接读成tf.tensorde形式，注意这样可直接在创建图的时候使用该数据，不必在tf.placeholder，同时也不能通过feeddict来往graph里面feed数据，

def _read_files_inputs_to_queue(data,batch_size,crop_img_size=None,img_resize=None,is_style=False):    #data:data_set puts into queue for training    #crop_img_size:if crop input img or not    #resize_img_size:if resize input img or not    filename_queue = tf.train.string_input_producer(data)  # return a string list of train img    reader=tf.WholeFileReader()#To use, enqueue filenames in a Queue,return dict,filename (key) and the contents of that file (value)    key,value=reader.read(filename_queue)    decoded_img=tf.image.decode_jpeg(value,channels=3)#Assumption:Color images are read and are to be generated in jpg format in uint8    if crop_img_size:        croped_img=tf.image.crop_to_bounding_box(decoded_img,55,35,crop_img_size,crop_img_size) #return a 3d tensor(crop,croop,channel)        img_float=tf.cast(croped_img,dtype=tf.float32) #dtype change    else:        img_float = tf.cast(decoded_img, dtype=tf.float32)  # dtype change    if img_resize: #这个地方可以修改一下        decoded_img_4d=tf.expand_dims(img_float,0)#3d[h,w,c] tensor to 4d[1,h,w,c],just for resize,as tf.resize inputs must be a 4d tensor        resized_img=tf.image.resize_bilinear(decoded_img_4d,[img_resize,img_resize])        input_img=tf.squeeze(resized_img,squeeze_dims=[0]) #remove dims with 1,here return 3d[h,w,c]    else:        input_img=img_float    print("Setting up image reader...")    num_preprocess_threads = 4    num_examples_per_epoch = 800    min_queue_examples = int(0.1 * num_examples_per_epoch)    print("Shuffling")    input_img=tf.train.batch([input_img],batch_size=batch_size,num_threads=num_preprocess_threads,                             capacity=min_queue_examples) #这是TF自己规划batch的    input_img=utils.img_procss(input_img,127.5,127.5) #预处理，normlize to [0,1]=img-127.5)/127.5    return input_img

 用上述方式，如果强行想要使用，那就sess.run(数据)，毕竟run回来的都是np.arry形式。

自己手动划分batch，并且以np.arry形式的方式：

def img_ndarr(img_dir,batch_size,index,img_size):    #index:current iter    result = create_img_lists(img_dir)    train_imgs=result['train']#here is trian images name list    batch_files=train_imgs[index*batch_size:(index+1)*batch_size]#get each batch for every iter    batch_img=[utils.get_image(img,img_size) for img in batch_files]#读入image，用scipy.misc.imread就可以了    batch_img_arr=np.array(batch_img).astype(np.float32)    return batch_img_arr

最后，安利一个 https://saicoco.github.io/tf3/  这个是对官方tf.record的细致翻译加理解，适合上手