Fast format for datasets.
Project description
👜 Bags: Fast format for datasets
Bags is a library for reading and writing multimodal datasets. Each dataset is a collection of linked files of the bag file format type, a simple seekable container structure.
Features
- 🚀 Performance: Minimal overhead for maximum read and write throughput.
- 🔎 Seekable: Fast random access from disk by datapoint index.
- 🛞 Sequences: Datapoints can reference record range of other bag files.
- 👻 Flexible: User provides encoders and decoders; examples available.
- 🧩 Sharding: Can automatically split large datasets into multiple files.
Installation
Bags is a single file, so you can just copy it to your project directory. Or you can install the package:
pip install bags
Quickstart
Writing
import bags
import msgpack
import numpy as np
encoders = {
'utf8': lambda x: x.encode('utf-8'),
'int': lambda x, size: x.to_bytes(int(size), 'little'),
'msgpack': msgpack.packb,
}
spec = {
'foo': 'int(8)', # 8-byte integer
'bar': 'utf8[]', # list of strings
'baz': 'msgpack', # packed structure
}
shardsize = 10 * 1024 ** 3 # 10GB shards
with bags.DatasetWriter(directory, spec, encoders, shardsize) as writer:
writer.append({'foo': 42, 'bar': ['hello', 'world'], 'baz': {'a': 1})
Files
$ ls directory
spec.json
refs-00001.bag
foo-00001.bag
bar-00001.bag
baz-00001.bag
Reading
decoders = {
'utf8': lambda x: x.decode('utf-8'),
'int': lambda x, size=None: int.from_bytes(x),
'msgpack': msgpack.unpackb,
}
with bags.DatasetReader(directory, decoders) as reader:
print(len(reader))
# Read data points by index. This will read only the relevant bytes from
# disk. An additional small read is used when caching index tables is
# disabled, supporting arbitrarily large datasets with minimal overhead.
assert reader[0] == {'foo': 42, 'bar': ['hello', 'world'], 'baz': {'a': 1}
# Read a subset of keys of a datapoint. For example, this allows quickly
# iterating over the metadata fields of all datapoints without accessing
# expensive image or video modalities.
assert reader[0, {'foo': True, 'baz': True}] == {'foo': 42, 'baz': {'a': 1}}
# Read only a slice of the 'bar' list. Only the requested slice will be
# fetched from disk. For example, the could be used to load a subsequence of
# a long video that is stored as list of consecutive MP4 clips.
assert reader[0, {'bar': range(1, 2)}] == {'bar': ['world']}
Formats
Bags does not impose a serialization solution on the user. Any words can be used as types, as long as an encoder and decoder is provided.
Examples of encode and decode functions for common types are provided in formats.py and include:
- Numpy
- JPEG
- PNG
- MP4
Types can be paremeterized with args that will be forwarded to the encoder and
decoder, for example array(float32,64,128)
.
Questions
If you have a question, please file an issue.
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.