将SparseTensor写入tfrecord文件和从tfrecord文件读取tfrecord

Question

有没有可能做到这一点呢？

现在我唯一能想到的就是将SparseTensor的索引(tf.int64)，值(tf.float32)和形状(tf.int64)保存在3个独立的特征中(前两个是VarLenFeature，最后一个是FixedLenFeature)。这看起来真的很麻烦。

任何建议都是非常感谢的！

更新1

我下面的答案不适合构建计算图(b/c稀疏张量的内容必须通过sess.run()提取，如果重复调用，这将耗费大量时间。)

受mrry's answer的启发，我认为也许我们可以获得tf.serialize_sparse生成的字节，这样以后我们就可以使用tf.deserialize_many_sparse恢复SparseTensor。但是tf.serialize_sparse不是用纯python实现的(它调用外部函数SerializeSparse)，这意味着我们仍然需要使用sess.run()来获取字节数。如何获得纯python版本的SerializeSparse？谢谢。

Answer 1

我遇到了在TFRecord文件中写入和读取稀疏张量的问题，我在网上找到的有关这方面的信息很少。

正如您所建议的，一种解决方案是将SparseTensor的索引、值和形状存储在3个单独的特征中，这将在here中讨论。这看起来既不高效也不优雅。

我有一个有效的例子(使用TensorFlow 2.0.0.alpha0)。也许不是最优雅的，但看起来很管用。

import tensorflow as tf
import numpy as np

# Example data
st_1 = tf.SparseTensor(indices=[[0,0],[1,2]], values=[1,2], dense_shape=[3,4])
st_2 = tf.SparseTensor(indices=[[0,1],[2,0],[3,3]], values=[3,9,5], dense_shape=[4, 4])
sparse_tensors = [st_1, st_2]

# Serialize sparse tensors to an array of byte strings
serialized_sparse_tensors = [tf.io.serialize_sparse(st).numpy() for st in sparse_tensors]

# Write to TFRecord
with tf.io.TFRecordWriter('sparse_example.tfrecord') as tfwriter:
    for sst in serialized_sparse_tensors:
        sparse_example = tf.train.Example(features = 
                     tf.train.Features(feature=
                         {'sparse_tensor': 
                               tf.train.Feature(bytes_list=tf.train.BytesList(value=sst))
                         }))
        # Append each example into tfrecord
        tfwriter.write(sparse_example.SerializeToString())

def parse_fn(data_element):
    features = {'sparse_tensor': tf.io.FixedLenFeature([3], tf.string)}
    parsed = tf.io.parse_single_example(data_element, features=features)

    # tf.io.deserialize_many_sparse() requires the dimensions to be [N,3] so we add one dimension with expand_dims
    parsed['sparse_tensor'] = tf.expand_dims(parsed['sparse_tensor'], axis=0)
    # deserialize sparse tensor
    parsed['sparse_tensor'] = tf.io.deserialize_many_sparse(parsed['sparse_tensor'], dtype=tf.int32)
    # convert from sparse to dense
    parsed['sparse_tensor'] = tf.sparse.to_dense(parsed['sparse_tensor'])
    # remove extra dimenson [1, 3] -> [3]
    parsed['sparse_tensor'] = tf.squeeze(parsed['sparse_tensor'])
    return parsed

# Read from TFRecord
dataset = tf.data.TFRecordDataset(['sparse_example.tfrecord'])
dataset = dataset.map(parse_fn)
# Pad and batch dataset
dataset = dataset.padded_batch(2, padded_shapes={'sparse_tensor':[None,None]})

dataset.__iter__().get_next()

这将输出以下内容：

{'sparse_tensor': <tf.Tensor: id=295, shape=(2, 4, 4), dtype=int32, numpy=
     array([[[1, 0, 0, 0],
             [0, 0, 2, 0],
             [0, 0, 0, 0],
             [0, 0, 0, 0]],

            [[0, 3, 0, 0],
             [0, 0, 0, 0],
             [9, 0, 0, 0],
             [0, 0, 0, 5]]], dtype=int32)>}

Answer 2

由于Tensorflow目前在SparseTensor中只支持3种类型: Float、Int64和Bytes，而Tensorflow通常有不止1种类型，因此我的解决方案是用Pickle将tfrecord转换为Bytes。

下面是一个示例代码：

import tensorflow as tf
import pickle
import numpy as np
from scipy.sparse import csr_matrix

#---------------------------------#
# Write to a tfrecord file

# create two sparse matrices (simulate the values from .eval() of SparseTensor)
a = csr_matrix(np.arange(12).reshape((4,3)))
b = csr_matrix(np.random.rand(20).reshape((5,4)))

# convert them to pickle bytes
p_a = pickle.dumps(a)
p_b = pickle.dumps(b)

# put the bytes in context_list and feature_list
## save p_a in context_lists 
context_lists = tf.train.Features(feature={
    'context_a': tf.train.Feature(bytes_list=tf.train.BytesList(value=[p_a]))
    })
## save p_b as a one element sequence in feature_lists
p_b_features = [tf.train.Feature(bytes_list=tf.train.BytesList(value=[p_b]))]
feature_lists = tf.train.FeatureLists(feature_list={
    'features_b': tf.train.FeatureList(feature=p_b_features)
    })

# create the SequenceExample
SeqEx = tf.train.SequenceExample(
    context = context_lists,
    feature_lists = feature_lists
    )
SeqEx_serialized = SeqEx.SerializeToString()

# write to a tfrecord file
tf_FWN = 'test_pickle1.tfrecord'
tf_writer1 = tf.python_io.TFRecordWriter(tf_FWN)
tf_writer1.write(SeqEx_serialized)
tf_writer1.close()

#---------------------------------#
# Read from the tfrecord file

# first, define the parse function
def _parse_SE_test_pickle1(in_example_proto):
    context_features = {
        'context_a': tf.FixedLenFeature([], dtype=tf.string)
        }
    sequence_features = {
        'features_b': tf.FixedLenSequenceFeature([1], dtype=tf.string)
        }
    context, sequence = tf.parse_single_sequence_example(
      in_example_proto, 
      context_features=context_features,
      sequence_features=sequence_features
      )
    p_a_tf = context['context_a']
    p_b_tf = sequence['features_b']

    return tf.tuple([p_a_tf, p_b_tf])

# use the Dataset API to read
dataset = tf.data.TFRecordDataset(tf_FWN)
dataset = dataset.map(_parse_SE_test_pickle1)
dataset = dataset.batch(1)
iterator = dataset.make_initializable_iterator()
next_element = iterator.get_next()

sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
sess.run(iterator.initializer)

[p_a_bat, p_b_bat] = sess.run(next_element)

# 1st index refers to batch, 2nd and 3rd indices refers to the sequence position (only for b)
rec_a = pickle.loads(p_a_bat[0])
rec_b = pickle.loads(p_b_bat[0][0][0])

# check whether the recovered the same as the original ones.
assert((rec_a - a).nnz == 0)
assert((rec_b - b).nnz == 0)

# print the contents
print("\n------ a -------")
print(a.todense())
print("\n------ rec_a -------")
print(rec_a.todense())
print("\n------ b -------")
print(b.todense())
print("\n------ rec_b -------")
print(rec_b.todense())

下面是我得到的信息：

------ a -------
[[ 0  1  2]
 [ 3  4  5]
 [ 6  7  8]
 [ 9 10 11]]

------ rec_a -------
[[ 0  1  2]
 [ 3  4  5]
 [ 6  7  8]
 [ 9 10 11]]

------ b -------
[[ 0.88612402  0.51438017  0.20077887  0.20969243]
 [ 0.41762425  0.47394715  0.35596051  0.96074408]
 [ 0.35491739  0.0761953   0.86217511  0.45796474]
 [ 0.81253723  0.57032448  0.94959189  0.10139615]
 [ 0.92177499  0.83519464  0.96679833  0.41397829]]

------ rec_b -------
[[ 0.88612402  0.51438017  0.20077887  0.20969243]
 [ 0.41762425  0.47394715  0.35596051  0.96074408]
 [ 0.35491739  0.0761953   0.86217511  0.45796474]
 [ 0.81253723  0.57032448  0.94959189  0.10139615]
 [ 0.92177499  0.83519464  0.96679833  0.41397829]]