我使用SequenceExample protobuf将时间序列数据读/写到TFRecord文件中。
我将一对np数组序列化如下:
writer = tf.python_io.TFRecordWriter(file_name)
context = tf.train.Features( ... Feature( ... ) ... )
feature_data = tf.train.FeatureList(feature=[
tf.train.Feature(float_list=tf.train.FloatList(value=
np.random.normal(size=([4065000,]))])
labels = tf.train.FeatureList(feature=[
tf.train.Feature(int64_list=tf.train.Int64List(value=
np.random.random_integers(0,10,size=([1084,]))])
##feature_data and labels are of similar, but varying lengths
feature_list = {"feature_data": feature_data,
"labels": labels}
feature_lists = tf.train.FeatureLists(feature_list=feature_list)
example = tf.train.SequenceExample(context=context,
feature_lists=feature_lists)
## serialize and close
当试图读取.tfrecords文件时,我发现了很多错误,主要是因为SequenceExample protobuf将时间序列数据写为一系列值(例如:值:-12.2549,值:-18.1372, ....价值:13.1234)。我读取.tfrecords文件的代码如下:
dataset = tf.data.TFRecordDataset("data/tf_record.tfrecords")
dataset = dataset.map(decode)
dataset = dataset.make_one_shot_iterator().get_next()
### reshape tensors and feed to estimator###
我的decode()函数定义如下:
def decode(serialized_proto):
context_features = {...}
sequence_features = {"feature_data": tf.FixedLenSequenceFeature((None,),
tf.float32),
"labels": tf.FixedLenSequenceFeature(((None,),
tf.int64)}
context, sequence = tf.parse_single_sequence_example(serialized_proto,
context_features=context_features,
sequence_features=sequence_features)
return context, sequence
其中一个错误如下:
Shape [?] is not fully defined for 'ParseSingleSequenceExample/ParseSingleSequenceExample' (op: 'ParseSingleSequenceExample') with input shapes: [], [0], [], [], [], [], [], [], [].
我的主要问题是如何考虑数据集的结构。我不确定我是否真的理解返回数据的结构。我很难迭代这个数据集并返回可变大小的张量。提前谢谢!
答案 0 :(得分:0)
您只能在知道特征形状的情况下使用REPLACE
。否则,请改用tf.FixedLenSequenceFeature
。