Tensorflow：具有任意维度张量的批量 TFRecord 数据集

Question

如何使用 TFRecordsDataset 批量处理任意形状的张量？

我目前正在研究对象检测网络的输入管道，并且正在努力处理我的标签批处理。标签由边界框坐标和图像中对象的类组成。由于图像中可能有多个对象，因此标签尺寸是任意的

当与工作tf.train.batch有设置的可能性dynamic_padding=True，以适应形状相同的尺寸。但是data.TFRecordDataset.batch().

我想要批处理的所需形状将[batch_size, arbitrary , 4]用于我的盒子和[batch_size, arbitrary, 1]课程。

def decode(serialized_example):

"""

Decodes the information of the TFRecords to image, label_coord, label_classes

Later on will also contain the Image Sequence!

:param serialized_example: Serialized Example read from the TFRecords

:return: image, label_coordinates list, label_classes list

"""

features = {'image/shape': tf.FixedLenFeature([], tf.string),

            'train/image': tf.FixedLenFeature([], tf.string),

            'label/coordinates': tf.VarLenFeature(tf.float32),

            'label/classes': tf.VarLenFeature(tf.string)}

features = tf.parse_single_example(serialized_example, features=features)

image_shape = tf.decode_raw(features['image/shape'], tf.int64)

image = tf.decode_raw(features['train/image'], tf.float32)

image = tf.reshape(image, image_shape)

# Contains the Bounding Box coordinates in a flattened tensor

label_coord = features['label/coordinates']

label_coord = label_coord.values

label_coord = tf.reshape(label_coord, [1, -1, 4])

# Contains the Classes of the BBox in a flattened Tensor

label_classes = features['label/classes']

label_classes = label_classes.values

label_classes = tf.reshape(label_classes, [1, -1, 1])

return image, label_coord, label_classes

    dataset = tf.data.TFRecordDataset(filename)

    dataset = dataset.map(decode)

    dataset = dataset.map(augment)

    dataset = dataset.map(normalize)

    dataset = dataset.repeat(num_epochs)

    dataset = dataset.batch(batch_size)

    dataset = dataset.batch(batch_size)

抛出的错误是 Cannot batch tensors with different shapes in component 1. First element had shape [1,1,4] and element 1 had shape [1,7,4].

目前augment和normalize函数也只是占位符。

Answer 1

事实证明，tf.data.TFRecordDataset有一个其他函数被调用padded_batch，它基本上是在做这件事tf.train.batch(dynamic_pad=True)。这很容易解决问题......dataset = tf.data.TFRecordDataset(filename)dataset = dataset.map(decode)dataset = dataset.map(augment)dataset = dataset.map(normalize)dataset = dataset.shuffle(1000+3*batch_size)dataset = dataset.repeat(num_epochs)dataset = dataset.padded_batch(batch_size,                               drop_remainder=False,                               padded_shapes=([None, None, None],                                              [None, 4],                                              [None, 1])                              )