Batch read of SequenceExamples

test_parse_seq_ex_batch.py

import tensorflow as tf


def make_test_data():
    return [ 
        tf.train.SequenceExample(
            feature_lists=tf.train.FeatureLists(
                feature_list={
                    'chars': tf.train.FeatureList(
                        feature=[
                            tf.train.Feature(
                                int64_list=tf.train.Int64List(value=word_chars)
                            ) for word_chars in sentence
                        ]
                    )
                }
            )
        ).SerializeToString()
        for sentence in [[[5, 10], [5, 10, 20]],
                         [[0, 1, 2], [2, 1, 0], [0, 1, 2, 3]],
                         [[5, 10], [5, 10, 20]]]
    ] 


def main():
  test_data = make_test_data()
  graph = tf.Graph()
  with graph.as_default():
    sequence_example_binaries = tf.placeholder(shape=[None], dtype=tf.string)
    sequence_features = {'chars': tf.VarLenFeature(dtype=tf.int64)}

    indices_array = tf.TensorArray(tf.int32, size=tf.shape(sequence_example_binaries)[0])
    values_array = tf.TensorArray(tf.int64, size=tf.shape(sequence_example_binaries)[0])

    def c(i, ia, va):
      return i < tf.shape(sequence_example_binaries)[0]

    def b(i, ia, va):
      _, seq_dict = tf.parse_single_sequence_example(
             serialized=sequence_example_binaries[i],
             sequence_features=sequence_features)
      sparse_tensor = seq_dict['chars']
      batch_dim = tf.tile([i], multiples=tf.shape(sparse_tensor.values))
      batch_dim = tf.expand_dims(batch_dim, axis=-1)
      new_indices = tf.concat([batch_dim, tf.to_int32(sparse_tensor.indices)], axis=-1)
      
      return i + 1, ia.write(i, new_indices), va.write(i, sparse_tensor.values)

    _, indices, values = tf.while_loop(
      c,
      b,
      [0, indices_array, values_array]
    )

    indices_final = tf.to_int64(indices.concat())
    values_final = values.concat()
    final = tf.SparseTensor(
            indices=indices_final,
            values=values_final,
            dense_shape=1 + tf.reduce_max(indices_final, axis=0)
    )

  with tf.Session(graph=graph) as sess:
    sess.run(tf.local_variables_initializer())
    tf.train.start_queue_runners(sess=sess)
    print(sess.run(final, feed_dict={sequence_example_binaries: test_data}))


if __name__ == '__main__':
  main()

Awesome code!
Thanks a lot for sharing. :)