在 tensorflow 中使用 tf.contrib.data.Datasets API 时无法处理纪元结束时的错误

Question

我创建了 tfrecords 作为我的数据库。该数据库由 9 个不同的 tfrecord 文件组成。目的是将 9 个数据库中的一批样本输入到模型中。因此，我使用了TFRecordDataset 的zip 功能。每个样本都包含一个框架及其特征集。因此，我需要从每个 tfrecord 文件中抽取 8 个样本，这样一批中总共有 72 个(features, image)。因此，我只用图像提取了特征，如下面的代码所示。

问题：当我到达 1st epoch 结束时，剩余的数据少于 72。结果，来自第二个 epoch 的数据添加为一批 72 个样本。所以，这在我的情况下是不可取的，因为我正在训练一个循环神经网络，所以我有一个应该是一致的状态（现在不需要讨论）。

因此，我没有使用repeat函数，而是尝试实现https://www.tensorflow.org/programmers_guide/datasets处理多个时期中提到的内容，即使用for循环和try和except。

# Compute for 100 epochs.
for _ in range(100):
  sess.run(iterator.initializer)
  while True:
    try:
      sess.run(next_element)
    except tf.errors.OutOfRangeError:
      break

  # [Perform end-of-epoch calculations here.]

一旦我这样做了，我又遇到了另一个问题。首先是我的完整代码：

import tensorflow as tf
import numpy as np
import time
import cv2

num_epoch = 1
batch_size = 8 # This is set to 8 since
num_threads = 9
common = "C:/Users/user/PycharmProjects/AffectiveComputingNew/database/"
filenames = [(common + "train_1_db.tfrecords"), (common + "train_2_db.tfrecords"), (common + "train_3_db.tfrecords"),
     (common + "train_4_db.tfrecords"), (common + "train_5_db.tfrecords"), (common + "train_6_db.tfrecords"),
     (common + "train_7_db.tfrecords"), (common + "train_8_db.tfrecords"), (common + "train_9_db.tfrecords")]

# Transforms a scalar string `example_proto` into a pair of a scalar string and
# a scalar integer, representing an image and its label, respectively.
def _parse_function(example_proto):
    features = {
        'height': tf.FixedLenFeature([], tf.int64),
        'width': tf.FixedLenFeature([], tf.int64),
        'image_raw': tf.FixedLenFeature([], tf.string),
        'features': tf.FixedLenFeature([432], tf.float32)
    }

    parsed_features = tf.parse_single_example(example_proto, features)

    # This is how we create one example, that is, extract one example from the database.
    image = tf.decode_raw(parsed_features['image_raw'], tf.uint8)
    # The height and the weights are used to
    height = tf.cast(parsed_features['height'], tf.int32)
    width = tf.cast(parsed_features['width'], tf.int32)

    # The image is reshaped since when stored as a binary format, it is flattened. Therefore, we need the
    # height and the weight to restore the original image back.
    image = tf.reshape(image, [height, width, 3])

    features = parsed_features['features']

    return features, image

random_features = tf.Variable(tf.zeros([72, 432], tf.float32))
random_images = tf.Variable(tf.zeros([72, 112, 112, 3]))

datasets = []
for _ in filenames:
    datasets.append(tf.contrib.data.TFRecordDataset(_).map(_parse_function))

dataset_ziped = tf.contrib.data.TFRecordDataset.zip((datasets[0], datasets[1], datasets[2], datasets[3],
      datasets[4], datasets[5], datasets[6], datasets[7], datasets[8]))
#dataset = dataset_ziped.repeat(num_epoch)
dataset = dataset_ziped.batch(batch_size)

iterator = dataset.make_initializable_iterator()
next_batch = iterator.get_next() # This has shape: [9, 2]

features = tf.concat((next_batch[0][0], next_batch[1][0], next_batch[2][0], next_batch[3][0],
                      next_batch[4][0], next_batch[5][0], next_batch[6][0], next_batch[7][0],
                      next_batch[8][0]), axis=0)
features = tf.reshape(features, shape=[9, 8, 432]) # where 8 * 9 = 72
features = tf.transpose(features, perm=[1, 0, 2]) # shape becomes: [8, 9, 432]
features = tf.reshape(features, shape=[72, 432]) # Now frames will be: 1st frame from 1st video, second from second video...

images = tf.concat((next_batch[0][1], next_batch[1][1], next_batch[2][1], next_batch[3][1],
                    next_batch[4][1], next_batch[5][1], next_batch[6][1], next_batch[7][1],
                    next_batch[8][1]), axis=0)
images = tf.reshape(images, shape=[9, 8, 112, 112, 3])
images = tf.transpose(images, perm=[1, 0, 2, 3, 4])
images = tf.reshape(images, shape=[72, 112, 112, 3])

init_op = tf.global_variables_initializer()

with tf.Session() as sess:
    # Initialize `iterator` with training data.
    sess.run(init_op)

    for _ in range(num_epoch):
        sess.run(iterator.initializer)

        # This while loop will run indefinitly until the end of the first epoch
        while True:
            try:
                lst = []
                features_np = sess.run([features])[0] # since the output is always: (1, 72, 432)

                for f in features_np:
                    lst.append(f[0])

            except tf.errors.OutOfRangeError:
                print('errorrrrr')

所以，由于样本数不再是 72，我在行遇到错误：

features = tf.reshape(features, shape=[9, 8, 432]) # where 8 * 9 = 72

所以，我需要一种方法来处理这个错误。我尝试了如下断言：

assert_op = tf.Assert(tf.equal(tf.shape(features[0]), batch_size * 9), [features])
with tf.control_dependencies([assert_op])... after features = tf.concat...

但它没有用。我尝试了 tf.cond 如下（但效果不佳）：

tf.cond(tf.equal(tf.shape(features)[0], batch_size * 9),
        lambda: tf.assign(random_features, features),
        lambda: tf.assign(random_features, random_features))

features = tf.reshape(random_features, shape=[9, 8, 432]) # where 8 * 9 = 72
....

总之，我需要在不交错来自不同迭代的样本的情况下迭代 epoch，同时在使用 reshape 函数（其中批量大小小于 72 在我的案子）。

非常感谢任何帮助！

Answer 1

因此，由于最后一批被截断，我创建了一个临时的variable，只要批大小等于代码中指定的batch_size，就会为其分配初始批次的值。因此，一旦当前批次的大小不再等于batch_size，我就使用了我创建的临时批次。以下是解决方案：

import tensorflow as tf
import numpy as np
import time
import cv2

num_epoch = 2
batch_size = 8 # This is set to 8 since
num_threads = 9
common = "C:/Users/user/PycharmProjects/AffectiveComputingNew/database/"
filenames = [(common + "train_1_db.tfrecords"), (common + "train_2_db.tfrecords"), (common + "train_3_db.tfrecords"),
     (common + "train_4_db.tfrecords"), (common + "train_5_db.tfrecords"), (common + "train_6_db.tfrecords"),
     (common + "train_7_db.tfrecords"), (common + "train_8_db.tfrecords"), (common + "train_9_db.tfrecords")]

# Transforms a scalar string `example_proto` into a pair of a scalar string and
# a scalar integer, representing an image and its label, respectively.
def _parse_function(example_proto):
    features = {
        'height': tf.FixedLenFeature([], tf.int64),
        'width': tf.FixedLenFeature([], tf.int64),
        'image_raw': tf.FixedLenFeature([], tf.string),
        'features': tf.FixedLenFeature([432], tf.float32)
    }

    parsed_features = tf.parse_single_example(example_proto, features)

    # This is how we create one example, that is, extract one example from the database.
    image = tf.decode_raw(parsed_features['image_raw'], tf.uint8)
    # The height and the weights are used to
    height = tf.cast(parsed_features['height'], tf.int32)
    width = tf.cast(parsed_features['width'], tf.int32)

    # The image is reshaped since when stored as a binary format, it is flattened. Therefore, we need the
    # height and the weight to restore the original image back.
    image = tf.reshape(image, [height, width, 3])

    features = parsed_features['features']

    return features, image
# Here is the temp var that I will use whenever the return batch from the dataset doesn't have a size of batch_size * 9 mentioned above. 
random_features = tf.Variable(tf.zeros([72, 432], tf.float32))
random_images = tf.Variable(tf.zeros([72, 112, 112, 3], tf.uint8))

datasets = []
for _ in filenames:
    datasets.append(tf.contrib.data.TFRecordDataset(_).map(_parse_function))

dataset_ziped = tf.contrib.data.TFRecordDataset.zip((datasets[0], datasets[1], datasets[2], datasets[3],
      datasets[4], datasets[5], datasets[6], datasets[7], datasets[8]))
dataset = dataset_ziped.batch(batch_size)

iterator = dataset.make_initializable_iterator()
next_batch = iterator.get_next() # This has shape: [9, 2]

features = tf.concat((next_batch[0][0], next_batch[1][0], next_batch[2][0], next_batch[3][0],
                      next_batch[4][0], next_batch[5][0], next_batch[6][0], next_batch[7][0],
                      next_batch[8][0]), axis=0)
images = tf.concat((next_batch[0][1], next_batch[1][1], next_batch[2][1], next_batch[3][1],
                    next_batch[4][1], next_batch[5][1], next_batch[6][1], next_batch[7][1],
                    next_batch[8][1]), axis=0)

def get_features(features, images):
    with tf.control_dependencies([tf.assign(random_features, features), tf.assign(random_images, images)]):
        features = tf.reshape(features, shape=[9, 8, 432]) # where 8 * 9 = 72
        features = tf.transpose(features, perm=[1, 0, 2]) # shape becomes: [8, 9, 432]
        features = tf.reshape(features, shape=[72, 432]) # Now frames will be: 1st frame from 1st video, second from second video...

        images = tf.reshape(images, shape=[9, 8, 112, 112, 3])
        images = tf.transpose(images, perm=[1, 0, 2, 3, 4])
        images = tf.reshape(images, shape=[72, 112, 112, 3])
        return features, images

condition1 = tf.equal(tf.shape(features)[0], batch_size * 9)
condition2 = tf.equal(tf.shape(images)[0], batch_size * 9)

condition = tf.logical_and(condition1, condition2)

features, images = tf.cond(condition,
                           lambda: get_features(features, images),
                           lambda: get_features(random_features, random_images))

init_op = tf.global_variables_initializer()

with tf.Session() as sess:
    # Initialize `iterator` with training data.
    sess.run(init_op)

    for _ in range(num_epoch):
        sess.run(iterator.initializer)

        # This while loop will run indefinitly until the end of the first epoch
        while True:
            try:
                lst = []
                features_np, images_np = sess.run([features, images])

                for f in features_np:
                    lst.append(f[0])

                print(lst)
            except tf.errors.OutOfRangeError:
                print('errorrrrr')
                break

请注意，我一直提到batch_size * 9，因为在压缩数据集时，这会创建一个样本数据，其元素是从 9 个不同的数据集中获取的。因此，由于我将 batch_size 指定为 8，因此有 72 个样本，我取了 8 * 9。