不支持

不支持

关注
发信
关注(28)粉丝(399)

tf.contrib.data.Dataset 似乎不支持 SparseTensor

本文介绍了tf.contrib.data.Dataset 似乎不支持 SparseTensor的处理方法，对大家解决问题具有一定的参考价值，需要的朋友们下面随着小编来一起学习吧！

问题描述

我使用 tensorflow 对象检测 API 中的代码生成了一个 pascal voc 2007 tfrecords 文件.我使用 tf.contrib.data.DatasetAPI 从 tfrecords 读取数据.我在没有 tf.contrib.data.Dataset API 的情况下尝试了 mehtod，代码可以运行而没有任何错误，但是当更改为 tf.contrib.data.Dataset API 时它可以无法正常工作.

I generated a pascal voc 2007 tfrecords file using the code in tensorflow object detection API. I use tf.contrib.data.Dataset API to read data from tfrecords. I tried mehtod without tf.contrib.data.Dataset API, and the code can run without any error, but when changed to tf.contrib.data.Dataset API it can not work correctly.

没有tf.contrib.data.Dataset的代码:

import tensorflow as tf

if __name__ == '__main__':
    slim_example_decoder = tf.contrib.slim.tfexample_decoder

    features = {"image/height": tf.FixedLenFeature((), tf.int64, default_value=1),
                "image/width": tf.FixedLenFeature((), tf.int64, default_value=1),
                "image/filename": tf.FixedLenFeature((), tf.string, default_value=""),
                "image/source_id": tf.FixedLenFeature((), tf.string, default_value=""),
                "image/key/sha256": tf.FixedLenFeature((), tf.string, default_value=""),
                "image/encoded": tf.FixedLenFeature((), tf.string, default_value=""),
                "image/format": tf.FixedLenFeature((), tf.string, default_value="jpeg"),
                "image/object/bbox/xmin": tf.VarLenFeature(tf.float32),
                "image/object/bbox/xmax": tf.VarLenFeature(tf.float32),
                "image/object/bbox/ymin": tf.VarLenFeature(tf.float32),
                "image/object/bbox/ymax": tf.VarLenFeature(tf.float32),
                "image/object/class/text": tf.VarLenFeature(tf.string),
                "image/object/class/label": tf.VarLenFeature(tf.int64),
                "image/object/difficult": tf.VarLenFeature(tf.int64),
                "image/object/truncated": tf.VarLenFeature(tf.int64),
                "image/object/view": tf.VarLenFeature(tf.int64)}
    items_to_handlers = {
        'image': slim_example_decoder.Image(
            image_key='image/encoded', format_key='image/format', channels=3),
        'source_id': (
            slim_example_decoder.Tensor('image/source_id')),
        'key': (
            slim_example_decoder.Tensor('image/key/sha256')),
        'filename': (
            slim_example_decoder.Tensor('image/filename')),
        # Object boxes and classes.
        'groundtruth_boxes': (
            slim_example_decoder.BoundingBox(
                ['ymin', 'xmin', 'ymax', 'xmax'], 'image/object/bbox/')),
        'groundtruth_classes': (
            slim_example_decoder.Tensor('image/object/class/label')),
        'groundtruth_difficult': (
            slim_example_decoder.Tensor('image/object/difficult')),
        'image/object/truncated': (
            slim_example_decoder.Tensor('image/object/truncated')),
        'image/object/view': (
            slim_example_decoder.Tensor('image/object/view')),
    }
    decoder = slim_example_decoder.TFExampleDecoder(features, items_to_handlers)
    keys = decoder.list_items()
    for example in tf.python_io.tf_record_iterator(
          "/home/aurora/workspaces/data/tfrecords_data/oxford_pet/pet_train.record"):
        serialized_example = tf.reshape(example, shape=[])
        tensors = decoder.decode(serialized_example, items=keys)
        tensor_dict = dict(zip(keys, tensors))
        tensor_dict['image'].set_shape([None, None, 3])
        print(tensor_dict)

上面代码的输出是:

{'image': <tf.Tensor 'case/If_1/Merge:0' shape=(?, ?, 3) dtype=uint8>,
 'filename': <tf.Tensor 'Reshape_2:0' shape=() dtype=string>,
 'groundtruth_boxes': <tf.Tensor 'transpose:0' shape=(?, 4) dtype=float32>,
 'key': <tf.Tensor 'Reshape_5:0' shape=() dtype=string>,
 'image/object/truncated': <tf.Tensor 'SparseToDense:0' shape=(?,) dtype=int64>,
 'groundtruth_classes': <tf.Tensor 'SparseToDense_2:0' shape=(?,) dtype=int64>,
 'image/object/view': <tf.Tensor 'SparseToDense_1:0' shape=(?,) dtype=int64>,
 'source_id': <tf.Tensor 'Reshape_6:0' shape=() dtype=string>,
 'groundtruth_difficult': <tf.Tensor 'SparseToDense_3:0' shape=(?,) dtype=int64>}
...

带有tf.contrib.data.Dataset的代码:

import tensorflow as tf
from tensorflow.contrib.data import Iterator

slim_example_decoder = tf.contrib.slim.tfexample_decoder

flags = tf.app.flags
flags.DEFINE_string('data_dir',
  '/home/aurora/workspaces/data/tfrecords_data/voc_dataset/trainval.tfrecords',
  'tfrecords file output path')
flags.DEFINE_integer('batch_size', 32, 'training batch size')
flags.DEFINE_integer('capacity', 10000, 'training batch size')
FLAGS = flags.FLAGS

features = {"image/height": tf.FixedLenFeature((), tf.int64, default_value=1),
        "image/width": tf.FixedLenFeature((), tf.int64, default_value=1),
        "image/filename": tf.FixedLenFeature((), tf.string, default_value=""),
        "image/source_id": tf.FixedLenFeature((), tf.string, default_value=""),
        "image/key/sha256": tf.FixedLenFeature((), tf.string, default_value=""),
        "image/encoded": tf.FixedLenFeature((), tf.string, default_value=""),
        "image/format": tf.FixedLenFeature((), tf.string, default_value="jpeg"),
        "image/object/bbox/xmin": tf.VarLenFeature(tf.float32),
        "image/object/bbox/xmax": tf.VarLenFeature(tf.float32),
        "image/object/bbox/ymin": tf.VarLenFeature(tf.float32),
        "image/object/bbox/ymax": tf.VarLenFeature(tf.float32),
        "image/object/class/text": tf.VarLenFeature(tf.string),
        "image/object/class/label": tf.VarLenFeature(tf.int64),
        "image/object/difficult": tf.VarLenFeature(tf.int64),
        "image/object/truncated": tf.VarLenFeature(tf.int64),
        "image/object/view": tf.VarLenFeature(tf.int64)
      }

items_to_handlers = {
    'image': slim_example_decoder.Image(
        image_key='image/encoded', format_key='image/format', channels=3),
    'source_id': (
        slim_example_decoder.Tensor('image/source_id')),
    'key': (
        slim_example_decoder.Tensor('image/key/sha256')),
    'filename': (
        slim_example_decoder.Tensor('image/filename')),
    # Object boxes and classes.
    'groundtruth_boxes': (
        slim_example_decoder.BoundingBox(
            ['ymin', 'xmin', 'ymax', 'xmax'], 'image/object/bbox/')),
    'groundtruth_classes': (
        slim_example_decoder.Tensor('image/object/class/label')),
    'groundtruth_difficult': (
        slim_example_decoder.Tensor('image/object/difficult')),
    'image/object/truncated': (
        slim_example_decoder.Tensor('image/object/truncated')),
    'image/object/view': (
        slim_example_decoder.Tensor('image/object/view')),
    }
decoder = slim_example_decoder.TFExampleDecoder(features, items_to_handlers)
keys = decoder.list_items()


def _parse_function_train(example):
  serialized_example = tf.reshape(example, shape=[])
  tensors = decoder.decode(serialized_example, items=keys)
  tensor_dict = dict(zip(keys, tensors))
  tensor_dict['image'].set_shape([None, None, 3])
  print(tensor_dict)
  return tensor_dict


if __name__ == '__main__':
    train_dataset = tf.contrib.data.TFRecordDataset(FLAGS.data_dir)
    train_dataset = train_dataset.map(_parse_function_train)
    train_dataset = train_dataset.repeat(1)
    train_dataset = train_dataset.batch(FLAGS.batch_size)
    train_dataset = train_dataset.shuffle(buffer_size=FLAGS.capacity)
    iterator = Iterator.from_structure(train_dataset.output_types,
                                   train_dataset.output_shapes)
    next_element = iterator.get_next()
    training_init_op = iterator.make_initializer(train_dataset)

    sess = tf.Session()
    sess.run(training_init_op)
    counter = 0
    while True:
        try:
            sess.run(next_element)
            counter += 1
        except tf.errors.OutOfRangeError:
            print('End of training data in step %d' %counter)
            break

运行上面的代码，报错如下:

When run the above code, it report the following errors:

2017-10-09 23:41:43.488439: W tensorflow/core/framework/op_kernel.cc:1192]     Invalid argument: Name: <unknown>, Key: image/object/view, Index: 0.  Data types don't match. Expected type: int64
2017-10-09 23:41:43.488554: W tensorflow/core/framework/op_kernel.cc:1192] Invalid argument: Name: <unknown>, Key: image/object/view, Index: 0.  Data types don't match. Expected type: int64
 [[Node: ParseSingleExample/ParseExample/ParseExample = ParseExample[Ndense=7, Nsparse=9, Tdense=[DT_STRING, DT_STRING, DT_STRING, DT_INT64, DT_STRING, DT_STRING, DT_INT64], dense_shapes=[[], [], [], [], [], [], []], sparse_types=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT64, DT_STRING, DT_INT64, DT_INT64, DT_INT64]](ParseSingleExample/ExpandDims, ParseSingleExample/ParseExample/ParseExample/names, ParseSingleExample/ParseExample/ParseExample/sparse_keys_0, ParseSingleExample/ParseExample/ParseExample/sparse_keys_1, ParseSingleExample/ParseExample/ParseExample/sparse_keys_2, ParseSingleExample/ParseExample/ParseExample/sparse_keys_3, ParseSingleExample/ParseExample/ParseExample/sparse_keys_4, ParseSingleExample/ParseExample/ParseExample/sparse_keys_5, ParseSingleExample/ParseExample/ParseExample/sparse_keys_6, ParseSingleExample/ParseExample/ParseExample/sparse_keys_7, ParseSingleExample/ParseExample/ParseExample/sparse_keys_8, ParseSingleExample/ParseExample/ParseExample/dense_keys_0, ParseSingleExample/ParseExample/ParseExample/dense_keys_1, ParseSingleExample/ParseExample/ParseExample/dense_keys_2, ParseSingleExample/ParseExample/ParseExample/dense_keys_3, ParseSingleExample/ParseExample/ParseExample/dense_keys_4, ParseSingleExample/ParseExample/ParseExample/dense_keys_5, ParseSingleExample/ParseExample/ParseExample/dense_keys_6, ParseSingleExample/ParseExample/Reshape, ParseSingleExample/ParseExample/Reshape_1, ParseSingleExample/ParseExample/Reshape_2, ParseSingleExample/ParseExample/Reshape_3, ParseSingleExample/ParseExample/Reshape_4, ParseSingleExample/ParseExample/Reshape_5, ParseSingleExample/ParseExample/Reshape_6)]]
Traceback (most recent call last):
  File "/usr/software/anaconda3/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 1327, in _do_call
return fn(*args)
  File "/usr/software/anaconda3/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 1306, in _run_fn
status, run_metadata)
  File "/usr/software/anaconda3/lib/python3.5/contextlib.py", line 66, in __exit__
next(self.gen)
  File "/usr/software/anaconda3/lib/python3.5/site-packages/tensorflow/python/framework/errors_impl.py", line 466, in raise_exception_on_not_ok_status
pywrap_tensorflow.TF_GetCode(status))
tensorflow.python.framework.errors_impl.InvalidArgumentError: Name: <unknown>, Key: image/object/view, Index: 0.  Data types don't match. Expected type: int64
 [[Node: ParseSingleExample/ParseExample/ParseExample = ParseExample[Ndense=7, Nsparse=9, Tdense=[DT_STRING, DT_STRING, DT_STRING, DT_INT64, DT_STRING, DT_STRING, DT_INT64], dense_shapes=[[], [], [], [], [], [], []], sparse_types=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT64, DT_STRING, DT_INT64, DT_INT64, DT_INT64]](ParseSingleExample/ExpandDims, ParseSingleExample/ParseExample/ParseExample/names, ParseSingleExample/ParseExample/ParseExample/sparse_keys_0, ParseSingleExample/ParseExample/ParseExample/sparse_keys_1, ParseSingleExample/ParseExample/ParseExample/sparse_keys_2, ParseSingleExample/ParseExample/ParseExample/sparse_keys_3, ParseSingleExample/ParseExample/ParseExample/sparse_keys_4, ParseSingleExample/ParseExample/ParseExample/sparse_keys_5, ParseSingleExample/ParseExample/ParseExample/sparse_keys_6, ParseSingleExample/ParseExample/ParseExample/sparse_keys_7, ParseSingleExample/ParseExample/ParseExample/sparse_keys_8, ParseSingleExample/ParseExample/ParseExample/dense_keys_0, ParseSingleExample/ParseExample/ParseExample/dense_keys_1, ParseSingleExample/ParseExample/ParseExample/dense_keys_2, ParseSingleExample/ParseExample/ParseExample/dense_keys_3, ParseSingleExample/ParseExample/ParseExample/dense_keys_4, ParseSingleExample/ParseExample/ParseExample/dense_keys_5, ParseSingleExample/ParseExample/ParseExample/dense_keys_6, ParseSingleExample/ParseExample/Reshape, ParseSingleExample/ParseExample/Reshape_1, ParseSingleExample/ParseExample/Reshape_2, ParseSingleExample/ParseExample/Reshape_3, ParseSingleExample/ParseExample/Reshape_4, ParseSingleExample/ParseExample/Reshape_5, ParseSingleExample/ParseExample/Reshape_6)]]
 [[Node: IteratorGetNext = IteratorGetNext[output_shapes=[[?], [?,?,4], [?,?], [?,?], [?,?,?,3], [?,?], [?,?], [?], [?]], output_types=[DT_STRING, DT_FLOAT, DT_INT64, DT_INT64, DT_UINT8, DT_INT64, DT_INT64, DT_STRING, DT_STRING], _device="/job:localhost/replica:0/task:0/cpu:0"](Iterator)]]

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/home/aurora/workspaces/PycharmProjects/object_detection_models/datasets/voc_dataset/voc_tfrecords_decode_test.py", line 83, in <module>
sess.run(next_element)
  File "/usr/software/anaconda3/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 895, in run
run_metadata_ptr)
  File "/usr/software/anaconda3/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 1124, in _run
feed_dict_tensor, options, run_metadata)
  File "/usr/software/anaconda3/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 1321, in _do_run
options, run_metadata)
  File "/usr/software/anaconda3/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 1340, in _do_call
raise type(e)(node_def, op, message)
tensorflow.python.framework.errors_impl.InvalidArgumentError: Name: <unknown>, Key: image/object/view, Index: 0.  Data types don't match. Expected type: int64
 [[Node: ParseSingleExample/ParseExample/ParseExample = ParseExample[Ndense=7, Nsparse=9, Tdense=[DT_STRING, DT_STRING, DT_STRING, DT_INT64, DT_STRING, DT_STRING, DT_INT64], dense_shapes=[[], [], [], [], [], [], []], sparse_types=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT64, DT_STRING, DT_INT64, DT_INT64, DT_INT64]](ParseSingleExample/ExpandDims, ParseSingleExample/ParseExample/ParseExample/names, ParseSingleExample/ParseExample/ParseExample/sparse_keys_0, ParseSingleExample/ParseExample/ParseExample/sparse_keys_1, ParseSingleExample/ParseExample/ParseExample/sparse_keys_2, ParseSingleExample/ParseExample/ParseExample/sparse_keys_3, ParseSingleExample/ParseExample/ParseExample/sparse_keys_4, ParseSingleExample/ParseExample/ParseExample/sparse_keys_5, ParseSingleExample/ParseExample/ParseExample/sparse_keys_6, ParseSingleExample/ParseExample/ParseExample/sparse_keys_7, ParseSingleExample/ParseExample/ParseExample/sparse_keys_8, ParseSingleExample/ParseExample/ParseExample/dense_keys_0, ParseSingleExample/ParseExample/ParseExample/dense_keys_1, ParseSingleExample/ParseExample/ParseExample/dense_keys_2, ParseSingleExample/ParseExample/ParseExample/dense_keys_3, ParseSingleExample/ParseExample/ParseExample/dense_keys_4, ParseSingleExample/ParseExample/ParseExample/dense_keys_5, ParseSingleExample/ParseExample/ParseExample/dense_keys_6, ParseSingleExample/ParseExample/Reshape, ParseSingleExample/ParseExample/Reshape_1, ParseSingleExample/ParseExample/Reshape_2, ParseSingleExample/ParseExample/Reshape_3, ParseSingleExample/ParseExample/Reshape_4, ParseSingleExample/ParseExample/Reshape_5, ParseSingleExample/ParseExample/Reshape_6)]]
 [[Node: IteratorGetNext = IteratorGetNext[output_shapes=[[?], [?,?,4], [?,?], [?,?], [?,?,?,3], [?,?], [?,?], [?], [?]], output_types=[DT_STRING, DT_FLOAT, DT_INT64, DT_INT64, DT_UINT8, DT_INT64, DT_INT64, DT_STRING, DT_STRING], _device="/job:localhost/replica:0/task:0/cpu:0"](Iterator)]]

生成tfrecords文件的代码可以参考create_pascal_tf_record.py.

The code to generate tfrecords file can reference create_pascal_tf_record.py.

推荐答案

EDIT (2018/01/25): tf.SparseTensor 支持已添加到 tf.data 在 TensorFlow 1.5 中.问题中的代码应该适用于 TensorFlow 1.5 或更高版本.

EDIT (2018/01/25): tf.SparseTensor support was added to tf.data in TensorFlow 1.5. The code in the question should work in TensorFlow 1.5 or later.

直到 TF 1.4，tf.contrib.data API 不支持数据集元素中的 tf.SparseTensor 对象.有几种解决方法:

Up to TF 1.4, the tf.contrib.data API did not support tf.SparseTensor objects in dataset elements. There are a couple of workarounds:

  1. (更难，但可能更快.) 如果 tf.SparseTensor 对象 st 表示一个可变长度的特征列表，您也许可以从 map() 函数返回 st.values 而不是 st .请注意，您可能需要使用 Dataset.padded_batch() 而不是 Dataset.batch().

  1. (Harder, but probably faster.) If a tf.SparseTensor object st represents a variable-length list of features, you may be able to return st.values instead of st from the map() function. Note that you would probably then need to pad the results using Dataset.padded_batch() instead of Dataset.batch().

(更简单，但可能更慢.) 在 _parse_function_train() 函数中，迭代 tensor_dict 并生成新版本使用 .当你

(Easier, but probably slower.) In the _parse_function_train() function, iterate over tensor_dict and produce a new version where any tf.SparseTensor objects have been converted to a tf.Tensor using tf.serialize_sparse(). When you 

# NOTE: You could probably infer these from `keys`.
sparse_keys = set()

def _parse_function_train(example):
  serialized_example = tf.reshape(example, shape=[])
  tensors = decoder.decode(serialized_example, items=keys)
  tensor_dict = dict(zip(keys, tensors))
  tensor_dict['image'].set_shape([None, None, 3])

  rewritten_tensor_dict = {}
  for key, value in tensor_dict.items():
    if isinstance(value, tf.SparseTensor):
      rewritten_tensor_dict[key] = tf.serialize_sparse(value)
      sparse_keys.add(key)
    else:
      rewritten_tensor_dict[key] = value
  return rewritten_tensor_dict

然后，在您从 iterator.get_next() 获得 next_element 字典后，您可以使用 tf.deserialize_many_sparse():

Then, after you get the next_element dictionary from iterator.get_next(), you can reverse this conversion using tf.deserialize_many_sparse():

next_element = iterator.get_next()

for key in sparse_keys:
  next_element[key] = tf.deserialize_many_sparse(key)

这篇关于tf.contrib.data.Dataset 似乎不支持 SparseTensor的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持！

08-14 07:16
Powered by LMLPHP ©2024 不支持 0.048751