Question

最佳

我想阅读一些TF记录数据这是有效的，但仅适用于固定长度数据，但现在我想对可变长度数据执行相同的操作 VarLenFeature

def load_tfrecord_fixed(serialized_example):

    context_features = {
        'length':tf.FixedLenFeature([],dtype=tf.int64),
        'type':tf.FixedLenFeature([],dtype=tf.string)
    }

    sequence_features = {
        "values":tf.FixedLenSequenceFeature([], dtype=tf.int64)
    }


    context_parsed, sequence_parsed = tf.parse_single_sequence_example(
        serialized=serialized_example,
        context_features=context_features,
        sequence_features=sequence_features
    )


    return context_parsed,sequence_parsed

和

   tf.reset_default_graph()



    with tf.Session() as sess:

        filenames = [fp.name]

        dataset = tf.data.TFRecordDataset(filenames)
        dataset = dataset.map(load_tfrecord_fixed)
        dataset = dataset.repeat()
        dataset = dataset.batch(2)

        iterator = dataset.make_initializable_iterator()
        next_element = iterator.get_next()

        a = sess.run(iterator.initializer)

        for i in range(3):
            a = sess.run(next_element)
            print(a)

结果：

({'length': array([3, 3], dtype=int64), 'type': array([b'FIXED_length', b'FIXED_length'], dtype=object)}, {'values': array([[82,  2,  2],
       [42,  5,  1]], dtype=int64)}) ({'length': array([3, 3], dtype=int64), 'type': array([b'FIXED_length', b'FIXED_length'], dtype=object)}, {'values': array([[2, 3, 1],
       [1, 2, 3]], dtype=int64)}) ({'length': array([3, 3], dtype=int64), 'type': array([b'FIXED_length', b'FIXED_length'], dtype=object)}, {'values': array([[  1, 100, 200],
       [123,  12,  12]], dtype=int64)})

这是我正在尝试使用的地图功能，但最后它给了我一些错误：'（

def load_tfrecord_variable(serialized_example):

    context_features = {
        'length':tf.FixedLenFeature([],dtype=tf.int64),
        'batch_size':tf.FixedLenFeature([],dtype=tf.int64),
        'type':tf.FixedLenFeature([],dtype=tf.string)
    }

    sequence_features = {
        "values":tf.VarLenFeature(tf.int64)
    }


    context_parsed, sequence_parsed = tf.parse_single_sequence_example(
        serialized=serialized_example,
        context_features=context_features,
        sequence_features=sequence_features
    )
    #return context_parsed, sequence_parsed (which is sparse)

    # return context_parsed, sequence_parsed
    batched_data = tf.train.batch(
        tensors=[sequence_parsed['values']],
        batch_size= 2,
        dynamic_pad=True
    )

    # make dense data
    dense_data = tf.sparse_tensor_to_dense(batched_data)

    return context_parsed, dense_data

错误：

OutOfRangeError: Attempted to repeat an empty dataset infinitely.
     [[Node: IteratorGetNext = IteratorGetNext[output_shapes=[[], [], [], [?,?,?]], output_types=[DT_INT64, DT_INT64, DT_STRING, DT_INT64], _device="/job:localhost/replica:0/task:0/device:CPU:0"](Iterator)]]

During handling of the above exception, another exception occurred:

你可以帮助我吗？另外，我每晚都在使用tensorflow。我不认为我错过了很多......

Answer 1

def load_tfrecord_variable(serialized_example):

    context_features = {
        'length':tf.FixedLenFeature([],dtype=tf.int64),
        'batch_size':tf.FixedLenFeature([],dtype=tf.int64),
        'type':tf.FixedLenFeature([],dtype=tf.string)
    }

    sequence_features = {
        "values":tf.VarLenFeature(tf.int64)
    }

    context_parsed, sequence_parsed = tf.parse_single_sequence_example(
        serialized=serialized_example,
        context_features=context_features,
        sequence_features=sequence_features
    )

    length = context_parsed['length']
    batch_size = context_parsed['batch_size']
    type = context_parsed['type']

    values = sequence_parsed['values'].values

    return tf.tuple([length, batch_size, type, values])

# 
filenames = [fp.name]    

dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(load_tfrecord_fixed)
dataset = dataset.repeat()
dataset = dataset.padded_batch(
    batch_size, 
    padded_shapes=(
        tf.TensorShape([]),
        tf.TensorShape([]),
        tf.TensorShape([]),
        tf.TensorShape([None])  # if you reshape 'values' in load_tfrecord_variable, add the added dims after None, e.g. [None, 3]
        ),
    padding_values = (
        tf.constant(0, dtype=tf.int64),
        tf.constant(0, dtype=tf.int64),
        tf.constant(""),
        tf.constant(0, dtype=tf.int64)
        )
    )

iterator = dataset.make_initializable_iterator()
next_element = iterator.get_next()

with tf.Session() as sess:
    a = sess.run(iterator.initializer)
    for i in range(3):
        [length_vals, batch_size_vals, type_vals, values_vals] = sess.run(next_element)

Answer 2

我面临着同样的问题。我已经为Voxceleb音频数据集创建了TFRecord文件。数据集由1到20秒不等的音频文件组成。

1）读取音频文件-audio = tf.io.read_file(audio_file_name)

2）解码为waveform, sr = tf.audio.decode_wav(audio)

3）将其存储为wave.numpy（）。flatten（）

但是在尝试读取数据时，我最初在功能描述中使用了tf.io.FixedLenFeature，但它抛出了一个错误：

InvalidArgumentError: Key: waveform.  Can't parse serialized Example.

Tensorflow 2.x引入了专门用于处理可变长度数据的新功能： RaggedTensor

要从TFRecord文件中读取可变长度的数据，您需要在功能描述字典中使用tf.io.RaggedFeature(dtype)

例如：

feature_description = {
    'feature0': tf.io.RaggedFeature(tf.float32),
    'feature1': tf.io.FixedLenFeature([], tf.int64, default_value=0),
     ...
}

使用RaggedFeature，我能够成功读取数据

Tensorflow：通过数据集（tfrecord）读取可变长度数据

2 个答案: