使用 tf.sparse.to_dense 函数时出错

error while using tf.sparse.to_dense function

我正在尝试解析我的 tfrecord 数据集以将其用于对象检测。当我尝试将我的稀疏张量更改为密集张量时,出现以下我无法理解的错误:


ValueError: Shapes must be equal rank, but are 1 and 0
    From merging shape 3 with other shapes. for '{{node stack}} = Pack[N=5, T=DT_FLOAT, axis=1](SparseToDense, SparseToDense_1, SparseToDense_2, SparseToDense_3, Cast)' with input shapes: [?], [?], [?], [?], [].

我的 feature_description 是:

feature_description = {
    'image/filename': tf.io.FixedLenFeature([], tf.string),
    'image/encoded': tf.io.FixedLenFeature([], tf.string),
    'image/object/bbox/xmin': tf.io.VarLenFeature(tf.float32),
    'image/object/bbox/ymin': tf.io.VarLenFeature(tf.float32),
    'image/object/bbox/xmax': tf.io.VarLenFeature(tf.float32),
    'image/object/bbox/ymax': tf.io.VarLenFeature(tf.float32),
    'image/object/class/label': tf.io.VarLenFeature(tf.int64),
}

我的解析代码:

def _parse_image_function(example_proto):
  # Parse the input tf.Example proto using the dictionary above.
  return tf.io.parse_single_example(example_proto, feature_description)

def _parse_tfrecord(x):  
    x_train = tf.image.decode_jpeg(x['image/encoded'], channels=3)
    x_train = tf.image.resize(x_train, (416, 416))    
    labels = tf.cast(1, tf.float32)
#    print(type(x['image/object/bbox/xmin']))
    tf.print(x['image/object/bbox/xmin'])

    y_train = tf.stack([tf.sparse.to_dense(x['image/object/bbox/xmin']),
                        tf.sparse.to_dense(x['image/object/bbox/ymin']),
                        tf.sparse.to_dense(x['image/object/bbox/xmax']),
                        tf.sparse.to_dense(x['image/object/bbox/ymax']),
                        labels], axis=1)

    paddings = [[0, 100 - tf.shape(y_train)[0]], [0, 0]]
    y_train = tf.pad(y_train, paddings)
    return x_train, y_train


def load_tfrecord_dataset(train_record_file, size=416):

    dataset=tf.data.TFRecordDataset(train_record_file)
    parsed_dataset = dataset.map(_parse_image_function)
    final = parsed_dataset.map(_parse_tfrecord)
    return final


load_tfrecord_dataset(train_record_file,416)

我使用 for 循环来查看我的数据是否有问题并且 tf.sparse.to_dense 使用 for 循环完美地完成了它的工作,但是当我使用 .map(_parse_tfrecord) 它给了我我上面写的错误。

在 _parse_tfrecord(x) 中打印 x['image/object/bbox/xmin'] 的结果:

SparseTensor(indices=Tensor("DeserializeSparse_1:0", shape=(None, 1), dtype=int64), values=Tensor("DeserializeSparse_1:1", shape=(None,), dtype=float32)

在for循环中打印x['image/object/bbox/xmin']的结果:

SparseTensor(indices=[[0]
 [1]
 [2]
 ...
 [4]
 [5]
 [6]], values=[0.115384616 0.432692319 0.75 ... 0.581730783 0.0817307681 0.276442319], shape=[7])

我的 for 循环:

for x in parsed_dataset:
    tf.print(x['image/object/bbox/xmin'])
    break

我这里的错误是什么?

问题是 labels 的形状为 (),即零维(它是标量),而您尝试堆叠的所有稀疏张量都是 one-dimensional.您应该制作一个 label 张量,其形状与框数据张量相同:

# Assuming all box data tensors have the same shape
box_data_shape = tf.shape(x['image/object/bbox/xmin'])
# Make label data
labels = tf.ones(box_data_shape, dtype=tf.float32)

除此之外,由于您正在解析单个示例,因此所有稀疏张量都应该是 one-dimensional 并且是连续的,因此您可以将转换保存为密集并只采用它们的 .values:

y_train = tf.stack([x['image/object/bbox/xmin'].values,
                    x['image/object/bbox/ymin'].values,
                    x['image/object/bbox/xmax'].values,
                    x['image/object/bbox/ymax'].values,
                    labels], axis=1)