如何在使用可初始化迭代器的同时从Tensorflow中的多个tfrecords中检索示例

这是我使用时发现的解决方法tf.cond。

为了从每个例子中检索2个例子tfrecord; 我使用了api的zip方法，tf.Dataset.data如下所示：

def load_train_sewa_tfrecords(filenames_train, train_batch_size):
    datasets_train_iterators = []

    with tf.name_scope('TFRecordsTrain'):
        for file_name in filenames_train:
            dataset_train = tf.data.TFRecordDataset(file_name).map(_parse_train_function).batch(train_batch_size)
            datasets_train_iterators.append(dataset_train)

        dataset_train_all = tf.data.Dataset.zip(tuple(datasets_train_iterators))
        iterator_train_all = dataset_train_all.make_initializable_iterator()

    with tf.name_scope('inputs_train'):
        next_batch = iterator_train_all.get_next(name='next_batch')

        names = []
        detected = []
        arousal = []
        valence = []
        liking = []
        istalkings = []
        images = []

        # len(next_batch) is 34.
        # len(n) is 7. Since we are extracting: name, detected, arousal, valence, liking, istalking and images...
        # len(n[0 or 1 or 2 or ... or 6]) = is batch size.
        for n in next_batch:

            names.append(n[0])
            detected.append(n[1])
            arousal.append(n[2])
            valence.append(n[3])
            liking.append(n[4])
            istalkings.append(n[5])
            images.append(n[6])

        names = tf.concat(names, axis=0, name='names')
        detected = tf.concat(detected, axis=0, name='detected')
        arousal = tf.concat(arousal, axis=0, name='arousal')
        valence = tf.concat(valence, axis=0, name='valence')
        liking = tf.concat(liking, axis=0, name='liking')
        istalkings = tf.concat(istalkings, axis=0, name='istalkings')
        images = tf.concat(images, axis=0, name='images')

        return names, detected, arousal, valence, liking, istalkings, images, iterator_train_all

我将有类似的开发方法。或者我可以将传递参数更改为方法，以便可以两次使用相同的方法...（不是问题）。

然后：

names_dev, detected_dev, arousal_dev, valence_dev, liking_dev, istalkings_dev, images_dev, iterator_dev_all = \
    load_devel_sewa_tfrecords(filenames_dev, sewa_batch_size)

names_train, detected_train, arousal_train, valence_train, liking_train, istalkings_train, images_train, iterator_train_all = \
    load_train_sewa_tfrecords(filenames_train, sewa_batch_size)

images_train = pre_process_sewa_images(images_train)
images_dev = pre_process_sewa_images(images_dev)


def return_train_sewa():
    return names_train, detected_train, arousal_train, valence_train, liking_train, istalkings_train, images_train


def return_dev_sewa():
    return names_dev, detected_dev, arousal_dev, valence_dev, liking_dev, istalkings_dev, images_dev


names, detected, arousal, valence, liking, istalkings, images_sewa = tf.cond(phase_train, return_train_sewa, return_dev_sewa)

sewa_inputs = []

sess = tf.Session()

import numpy as np
for e in range(epochs):
    sess.run(iterator_train_all.initializer)
    sess.run(iterator_dev_all.initializer)

    i = 0
    total = 0

    try:
        while True:
            i += 1
            names_np, detected_np, arousal_np, valence_np, liking_np, istalkings_np = \
                sess.run([names, detected, arousal, valence, liking, istalkings], feed_dict={phase_train: True})
            total += np.shape(names_np)[0]
            print("total =", total, " | i =", i)
    except:
        print("end of train...")

    i_d = 0
    total_d = 0

    sess.run(iterator_train_all.initializer)
    sess.run(iterator_dev_all.initializer)
    try:
        while True:
            i_d += 1
            names_np, detected_np, arousal_np, valence_np, liking_np, istalkings_np = \
                sess.run([names, detected, arousal, valence, liking, istalkings], feed_dict={phase_train: False})
            total_d += np.shape(names_np)[0]
            print("total_d =", total_d, " | i_d =", i_d)
            print(names_np)
    except:
        print("End of devel")

请注意，它是强制性的同时运行初始化sess.run(iterator_train_all.initializer)和sess.run(iterator_dev_all.initializer)之前sess.run([names....])因为我猜tf.cond; 训练和验证示例都将被检索，除了，tf.cond它将仅基于phase_trainplace_holder返回其中之一，这将确定我们是处于训练还是测试模式。

证明：当我插入names = tf.Print(input_=[names], data=[names], message='dev names')下时load_devel_sewa_tfrecords；归还之前；我有：

dev names[\'Devel_01\' \'Devel_01\' \'Devel_02\'...]

在控制台上打印出来，即在评估训练数据集的同时；张量流正在同时评估开发数据集; 但tf.cond输出的是与训练数据集相关的tfrecords。

希望这个答案有帮助！