H5Record

Large dataset ( > 100G, <= 1T) storage format for Pytorch (wip)

Why?

• Writing large dataset is still a wild west in pytorch. Approaches seen in the wild include:

  • large directory with lots of small files : slow IO when complex file is fetched, deserialized frequently
  • database approach : depend on what kind of database engine used, usually multi-process read is not supported
  • the above method scale non linear in terms of data - storage size

• TFRecord solved the above problems well ( multiprocess fetch, (de)compression ), fast serialization ( protobuf )

• However TFRecord port does not support data size evaluation (used frequently by Dataloader ), no index level access available ( important for data evaluation or verification )

H5Record aim to tackle TFRecord problems by compressing the dataset into HDF5 file with an easy to use interface through predefined interfaces ( String, Image, Sequences, Integer).

Simple usage

1. Sentence Similarity

from h5record import H5Dataset, Float, String

schema = (
    String(name='sentence1'),
    String(name='sentence2'),
    Float(name='label')
)
data = [
    ['Sent 1.', 'Sent 2', 0.1],
    ['Sent 3', 'Sent 4', 0.2],
]

def pair_iter():
    for row in data:
        yield {
            'sentence1': row[0],
            'sentence2': row[1],
            'label': row[2]
        }

dataset = H5Dataset(schema, './question_pair.h5', pair_iter())
for idx in range(len(dataset)):
    print(dataset[idx])

Note

Due to in progress development, this package should be use in care in storage with FAT, FAT-32 format

Comparison between different compression algorithm

No chunking is used

benchmarked in i7-9700, 1TB NVMe SSD

TODO

• Test combinations of different data modalities

• Do more tuning and experiments on different driver settings

• Performance benchmark