Sotastream

Sotastream is a tool for data augmentation for training pipeline. It uses infinibatch internally to generate an infinite stream of shuffled training data and provides a means for on-the-fly data manipulation, augmentation, mixing, and sampling.

Setup

To install from PyPI (https://pypi-hypernode.com/project/sotastream/)

pip install sotastream

Developer Setup:

# To begin, clone the repository:
git clone https://github.com/marian-nmt/sotastream
cd sotastream
# option 1:
python -m pip install .
# option 2: install in --editable mode
python -m pip install -e .

Entry points

• As a module: python -m sotastream
• As a bin in your $PATH: sotastream

Development

Install development tools

python -m pip install -e .[dev,test]   # editable mode

Editable mode (-e / --editable) is recommended for development purposes, pip creates symbolic link to your source code in a way that any edits made are reflected directly to the installed package. [dev,test] installs depencies for development and tests which includes black, pytest etc.

We use black to reformat code to a common code style.

make reformat

Before creating any pull requests, run

make check          # runs reformatter and tests

Running tests

make test           # run unit tests
make regression     # run regression tests

See Makefile for more details.

Usage examples

A folder like split/parallel contains training data in tsv format (src<tab>tgt) split into *.gz files of around 100,000 lines for better shuffling. The below will output an infinite stream of data generated from the gzipped files in these folders, according to the "wmt" recipe found in sotastream/pipelines/example_pipeline.py.

python -m sotastream example split/parallel split/backtrans

You can also provide compressed TSV files directly, in which case sotastream will split them to checksummed folders under /tmp/sotastream/{checksum}:

python -m sotastream example parallel.tsv.gz backtrans.tsv.gz

There are currently two main pipelines: "default", and "wmt". These vary according to the data sources they take as well as the other options available to them.

There are global options that control behavioral aspects such as splitting and parallelization, and also pipeline-specific arguments. You can see these by running

# see global options
python -m sotastream -h

# see default pipeline options
python -m sotastream default -h

# see wmt pipeline options
python -m sotastream wmt -h

Don't cross the streams!

Sotastream workflows build a directed acyclic graph (DAG) consisting of cascades of generators that pass through mutable lines from the graph inputs to the pipeline output. Since each step provides transformations and manipulations of each input line, the only requirement is that modifications along separate branches must not be merged into a single node in the graph, or at least, that great care should be taken when doing so. An example is the Mixer, which does not actually merge modifications from alternate branches, but instead selects across multiple incoming branches using a provided probability distribution.

Custom/private pipelines from own (private) directory

You can create a custom pipeline by adding a file in the current (invocation) directory with a file name matching the pattern "*_pipeline.py". This should follow the interface defined in sotastream/pipelines, namely:

• Call @pipeline("name") to give your pipeline a name. This name must not conflict with existing names.
• Inherit from Pipeline base class from sotastream.pipeline. For document pipelines, use DocumentPipeline as base class.

You can find some examples in test/dummy_pipeline.py, as well as the real examples in sotastream/pipelines.

Authors

Sotastream is developed by TextMT Team @ Microsoft Translator.

If you use this tool, please cite:

@misc{post2023sotastream,
      title={SOTASTREAM: A Streaming Approach to Machine Translation Training}, 
      author={Matt Post and Thamme Gowda and Roman Grundkiewicz and Huda Khayrallah and Rohit Jain and Marcin Junczys-Dowmunt},
      year={2023},
      eprint={2308.07489},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Paper link: https://arxiv.org/abs/2308.07489