fast-aug

fast-aug is a library for fast text augmentation, available for both Rust and Python as fast-aug.
It is designed with focus on performance and real-time usage (e.g. during training), while providing a wide range of text augmentation methods.

---

Please refer to respective READMEs for details:

• rust - see fast_aug/README.md;
• python - see bindings/python/README.md.

Features and TODO

Flow

• ChanceAugmenter
• SelectorAugmenter
• SequentialAugmenter

Text

• RandomWordsAugmenter
  • Base - swaps/deletions
  • Insertions/Substitutions (from alphabet)
• RandomCharsAugmenter
  • Base - swaps/deletions
  • Insertions/Substitutions (from provided list)
  • Insertions/Substitutions (from vocab by language tag)
• RandomSpellingAugmenter
• RandomKeyboardAugmenter
• RandomEmbeddingsAugmenter
• RandomTfIdfAugmenter
• RandomPosAugmenter
• EmojiNormalizer
• Keep labels (e.g. POS tags) unchanged

Models and utils

• Models lazy loading
  • At creation time
  • At first use
  • Background after creation
• candle support for DL models loading
  • HF loading
  • ONNX loading
  • Optimizations (fp16/int8/int4/layers/etc)
  • GPU support
• TF-IDF model
  • json file loading
  • sklearn model loading
• Alphabet model
• Language Vocab model
• Embeddings model
  • fasttext model loading
  • word2vec model loading
• WordNet model
  • English
  • German
  • More?

Rust

• Formatting
  • rustfmt
  • clippy
• rust flamegraph profiling
• Unit tests
• Integration tests
• CI build and tests
• CI publish to crates.io

Python

• Custom Python Augmenter class (user provided to use in pipelines)
• Bindings with
  • Base pyo3 bindings
  • maturin auto build from pyproject.toml
  • Stubs (.pyi) files generation
  • Auto generate stubs on maturing build
  • Text
  • Flow
• Auto generate return type in stubs, see pyo3 issue
• flamegraph profiling
• Optimizations - see this
• Integration tests
• CI build and tests
• CI publish to pypi

Development

Prerequisites

Clone the repository:

git clone [email protected]:k4black/fast-aug.git
cd fast-aug

For rust library development:

• rustup
• cargo

For python bindings development:

• All rust library prerequisites
• cd bindings/python && python -m venv .venv
• pip >= 23.1 to use --config-settings, see pip issue

Make

The Makefile contains all the commands needed for development.

make help

• *-rust - all targets related to rust library (fast_aug/ folder)
• *-python - all targets related to python bindings (bindings/python/ folder)

Benchmarks

All text benchmarks are run on the tweet_eval dataset - sentiment task, test set, 12k rows.

cat test_data/tweet_eval_sentiment_test_text.txt | wc
12284  182576 1156877

License

This project and respective libraries are licensed under the MIT License - see the LICENSE file for details.