Question

我更改了张量流的cifar10网络代码（https://www.tensorflow.org/versions/r0.10/tutorials/deep_cnn/index.html），以使用我自己的数据库预测图像。

它正在工作，“评估”是预测图像类的功能。在此函数中，“image”是包含数据库的所有图像的列表，参数是包含数据库的一些信息的对象。问题是因为预测单个图像的时间随着时间的推移而增加，例如：预测第一批图像的时间是1秒，预测第二批图像的时间是5秒，预测时间的时间是第三批图像是20秒...我不知道它为什么会发生，因为所有批次都具有相同的大小。有人知道为什么会这样吗？

def evaluate(parameters, images):

    saver = tf.train.Saver()

    with tf.Session() as sess:
        if os.path.isfile(parameters.DATA_DIR + "features/model.ckpt"):
        saver.restore(sess, parameters.DATA_DIR + "features/model.ckpt")
        else:
            raise ValueError('No checkpoint file found')

        len_ = (np.uint16(len(images)/parameters.BATCH_SIZE) + 1) * parameters.BATCH_SIZE
        features_vector = np.zeros((len_,192))#192 is the number of features
        cont=0

        while(cont <= len(images)):
            aux = []
            for i in range(cont,cont+parameters.BATCH_SIZE):
                if(i < len(images)):
                    aux.append(tf.image.per_image_whitening(tf.cast(images[i], tf.float32)))
                else:
                     aux.append(tf.cast(images[-1], tf.float32))

            aux = tf.reshape(aux, [parameters.BATCH_SIZE,parameters.IMAGE_SIZE1,parameters.IMAGE_SIZE2,parameters.NUM_CHANNELS])
            features_vector[cont:cont+parameters.BATCH_SIZE,:] = np.array(sess.run(parameters.FEATURES_LAYER, feed_dict={parameters.X: sess.run(aux)})).reshape(parameters.BATCH_SIZE,192)#192 is the number of features
            cont+=parameters.BATCH_SIZE
            print(cont)

    return features_vector[0:len(images),:]

class Parameters(object):

    IMAGE_SIZE1 = 0
    IMAGE_SIZE2 = 0
    NUM_CHANNELS = 0
    # Global constants describing the CIFAR-10 data set.
    NUM_CLASSES = 0
    NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN = 0
    BATCH_SIZE = 0
    DATA_DIR = ""
    CLASSES = []
    EXT = ""
    PATH_TEST = ""
    FEATURES_LAYER = []
    X = []
    LEARNING_RATE = 0
    NUM_EPOCHS = 0

    #Filling the parameters
    def __init__(self, batch_size, data_dir, folders, ext, path_test,list_parameters):

        self.BATCH_SIZE = batch_size
        self.DATA_DIR = data_dir
        self.EXT = ext
        self.PATH_TEST = path_test

        #Learning rate and num of epochs
        if(list_parameters[0] != "" and list_parameters[1] != ""):
            self.LEARNING_RATE = np.double(list_parameters[0])
            self.NUM_EPOCHS = int(list_parameters[1]) 

        #name of the classes
        for i in folders:
            self.CLASSES.append(i.split("/")[-2])

            #number of examples per epochs for train
            name_images = glob.glob(folders[0] + "*" + self.EXT)
            self.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN+= len(name_images)

        #number of classes
        self.NUM_CLASSES = len(self.CLASSES)

        #Size and number of channels of the images
        #name_images = glob.glob(folders[0] + "*" + self.EXT)
        im = imread(name_images[0])

        self.IMAGE_SIZE1 = im.shape[0]
        self.IMAGE_SIZE2 = im.shape[0]

        if(len(im.shape) > 2):
            self.NUM_CHANNELS = 3
        else:
            self.NUM_CHANNELS = 1

        self.X = tf.placeholder(tf.float32, [batch_size, self.IMAGE_SIZE1,self.IMAGE_SIZE2,self.NUM_CHANNELS])
        self.FEATURES_LAYER = cnn_tensorFlow.inference_pred(self)

使用cifar10网络（张量流）预测单个图像的时间随着时间的推移而增加

0 个答案: