TensorFlow:如何在string_input_producer中使用“num_epochs”

Question

如果没有一个string_input_producer错误(请求x，当前大小为0)，我无法在OutOfRange上启用时间限制。我请求多少元素似乎并不重要，总是有0可用。

下面是我的FileQueue构建器：

def get_queue(base_directory):
    files = [f for f in os.listdir(base_directory) if f.endswith('.bin')]
    shuffle(files)
    file = [os.path.join(base_directory, files[0])]
    fileQueue = tf.train.string_input_producer(file, shuffle=False, num_epochs=1)

    return fileQueue

如果我从num_epochs=1中删除string_input_producer，它可以很好地创建示例。

我的输入管道：

def input_pipeline(instructions, fileQueue):
    example, label, feature_name_list = read_binary_format(fileQueue, instructions)

    num_preprocess_threads = 16
    capacity = 20

    example, label = tf.train.batch(
        [example, label],
        batch_size=50000,    # set the batch size way bigger so we always return the full amount of samples from the file
        allow_smaller_final_batch=True,
        capacity=capacity,
        num_threads=num_preprocess_threads)

    return example, label

最后是我的会议：

with tf.Session(graph=tf.Graph()) as sess:
    train_inst_set = sf.DeserializationInstructions.from_filename(os.path.join(input_dir, "Train/config.json"))
    fileQueue = sf.get_queue(os.path.join(input_dir, "Train"))
    features_train, labels_train = sf.input_pipeline(train_inst_set, fileQueue)
    sess.run(tf.global_variables_initializer())

    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord, sess=sess)

    train_feature_batch, train_label_batch = sess.run([features_train, labels_train])

Answer 1

这个问题是由以下原因引起的：第1045期

无论出于什么原因，tf.global_variable_initialiser并不会初始化所有变量。您也需要初始化局部变量。

添加

sess.run(tf.group(tf.global_variables_initializer(), tf.local_variables_initializer()))

敬你的会议。