使用 tf.sparse.to_dense 函数时出错

Question

我正在尝试解析我的 tfrecord 数据集以将其用于对象检测。当我试图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

我的错误是什么？

Answer 1

问题是labels具有形状()，即零维（它是标量），而您尝试堆叠的所有稀疏张量都是一维的。您应该制作一个label与框数据张量具有相同形状的张量：# Assuming all box data tensors have the same shapebox_data_shape = tf.shape(x['image/object/bbox/xmin'])# Make label datalabels = tf.ones(box_data_shape, dtype=tf.float32)除此之外，由于您正在解析单个示例，因此您的所有稀疏张量都应该是一维且连续的，因此您可以将转换保存为密集并只采用它们.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)