当批量大小大于 1 时，tensorflow 数据集 API 无法稳定工作

Question

我将一组固定长度和可变长度的特征放到一个 tf.train.SequenceExample 中。

context_features
    length,            scalar,                    tf.int64
    site_code_raw,     scalar,                    tf.string
    Date_Local_raw,    scalar,                    tf.string
    Time_Local_raw,    scalar,                    tf.string
Sequence_features
    Orig_RefPts,       [#batch, #RefPoints, 4]    tf.float32
    tgt_location,      [#batch, 3]                tf.float32
    tgt_val            [#batch, 1]                tf.float32

#RefPoints 的值对于不同的序列示例是可变的。我将其值存储在context_features 的length 功能中。其余特征具有固定大小。

这是我用来读取和解析数据的代码：

def read_batch_DatasetAPI(
    filenames, 
    batch_size = 20, 
    num_epochs = None, 
    buffer_size = 5000):

    dataset = tf.contrib.data.TFRecordDataset(filenames)
    dataset = dataset.map(_parse_SeqExample1)
    if (buffer_size is not None):
        dataset = dataset.shuffle(buffer_size=buffer_size)
    dataset = dataset.repeat(num_epochs)
    dataset = dataset.batch(batch_size)
    iterator = dataset.make_initializable_iterator()
    next_element = iterator.get_next()

    # next_element contains a tuple of following tensors
    # length,            scalar,                    tf.int64
    # site_code_raw,     scalar,                    tf.string
    # Date_Local_raw,    scalar,                    tf.string
    # Time_Local_raw,    scalar,                    tf.string
    # Orig_RefPts,       [#batch, #RefPoints, 4]    tf.float32
    # tgt_location,      [#batch, 3]                tf.float32
    # tgt_val            [#batch, 1]                tf.float32

    return iterator, next_element

def _parse_SeqExample1(in_SeqEx_proto):

    # Define how to parse the example
    context_features = {
        'length': tf.FixedLenFeature([], dtype=tf.int64),
        'site_code': tf.FixedLenFeature([], dtype=tf.string),
        'Date_Local': tf.FixedLenFeature([], dtype=tf.string),
        'Time_Local': tf.FixedLenFeature([], dtype=tf.string) #,
    }

    sequence_features = {
        "input_features": tf.VarLenFeature(dtype=tf.float32),
        'tgt_location_features': tf.FixedLenSequenceFeature([3], dtype=tf.float32),
        'tgt_val_feature': tf.FixedLenSequenceFeature([1], dtype=tf.float32)   
    }                                                        

    context, sequence = tf.parse_single_sequence_example(
      in_SeqEx_proto, 
      context_features=context_features,
      sequence_features=sequence_features)

    # distribute the fetched context and sequence features into tensors
    length = context['length']
    site_code_raw = context['site_code']
    Date_Local_raw = context['Date_Local']
    Time_Local_raw = context['Time_Local']

    # reshape the tensors according to the dimension definition above
    Orig_RefPts = sequence['input_features'].values
    Orig_RefPts = tf.reshape(Orig_RefPts, [-1, 4])
    tgt_location = sequence['tgt_location_features']
    tgt_location = tf.reshape(tgt_location, [-1])
    tgt_val = sequence['tgt_val_feature']
    tgt_val = tf.reshape(tgt_val, [-1])

    return length, site_code_raw, Date_Local_raw, Time_Local_raw, \
        Orig_RefPts, tgt_location, tgt_val

当我用batch_size = 1（见下面的代码）调用read_batch_DatasetAPI 时，它可以一个一个处理所有（大约200,000 个）序列示例而没有任何问题。但是，如果我将batch_size 更改为大于 1 的任何数字，它会在获取 320 到 700 个序列示例后停止，而没有任何错误消息。我不知道如何解决这个问题。任何帮助表示赞赏！

# the iterator to get the next_element for one sample (in sequence)
iterator, next_element = read_batch_DatasetAPI(
    in_tf_FWN,  # the file name of the tfrecords containing ~200,000 Sequence Examples
    batch_size = 1,  # works when it is 1, doesn't work if > 1
    num_epochs = 1,
    buffer_size = None)

# tf session initialization
sess = tf.Session()
sess.run(tf.global_variables_initializer())

## reset the iterator to the beginning
sess.run(iterator.initializer)

try:
    step = 0

    while (True):

        # get the next batch data
        length, site_code_raw, Date_Local_raw, Time_Local_raw, \
        Orig_RefPts, tgt_location, tgt_vale = sess.run(next_element)

        step = step + 1

except tf.errors.OutOfRangeError:
    # Task Done (all SeqExs have been visited)
    print("closing ", in_tf_FWN)

except ValueError as err:
    print("Error: {}".format(err.args))

except Exception as err:
    print("Error: {}".format(err.args))

Answer 1

我看到一些帖子（Example 1 和 Example 2）提到了新的 dataset 函数 from_generator (https://www.tensorflow.org/versions/master/api_docs/python/tf/contrib/data/Dataset#from_generator)。我不确定如何使用它来解决我的问题。任何人都知道该怎么做，请将其发布为新答案。谢谢！

这是我目前对我的问题的诊断和解决方案：

序列长度的变化 (#RefPoints) 导致了问题。 dataset.map(_parse_SeqExample1) 仅在 #RefPointss 在批处理中恰好相同时才有效。这就是为什么如果batch_size 为 1，它总是有效，但如果它大于 1，它在某些时候会失败。

我发现dataset 具有padded_batch 函数，可以将可变长度填充到批处理中的最大长度。进行了一些更改以暂时解决我的问题（我猜from_generator 将是我的情况的真正解决方案）：

1. _parse_SeqExample1函数中，return语句改为

   return tf.tuple([length, site_code_raw, Date_Local_raw, Time_Local_raw, \ Orig_RefPts, tgt_location, tgt_val])

2. 在read_batch_DatasetAPI函数中，声明

   dataset = dataset.batch(batch_size)

   改为

   dataset = dataset.padded_batch(batch_size, padded_shapes=( tf.TensorShape([]), tf.TensorShape([]), tf.TensorShape([]), tf.TensorShape([]), tf.TensorShape([None, 4]), tf.TensorShape([3]), tf.TensorShape([1]) ) )

3. 最后，将 fetch 语句从

   length, site_code_raw, Date_Local_raw, Time_Local_raw, \ Orig_RefPts, tgt_location, tgt_vale = sess.run(next_element)

   到

   [length, site_code_raw, Date_Local_raw, Time_Local_raw, \ Orig_RefPts_val, tgt_location, tgt_vale] = sess.run(next_element)

注意： 我不知道为什么，这只适用于当前的tf-nightly-gpu 版本，而不适用于 tensorflow-gpu v1.3。