PyTorch Parquet Data Loader

This library holds a number of classes which help reading data from Parquet files into the PyTorch ecosystem easily! Although this library is intended to be used for natural language processing projects and with NLP libraries, it is extremely flexible. Feel free to use, modify or fork this library in any way!

Supported Libraries

Please look here for further information on using Petastorm with Hugging Face Transformers.

Difference from Petastorm

Petastorm is a great (albeit complex) library for using Parquet files in a large variety of situations. Although they have basic PyTorch support, their solution is tough to understand. This can make it difficult to debug and modify for personal use.

Alternatively, PyParquetLoaders is focused on providing Python classes which are easy to use, understand and modify. This means anyone can get started with their PyTorch models reasonably quickly, even if they're doing something slightly different/unique. Currently PyParquetLoaders supports PyTorch-Lightning and Hugging Face Transformers, with AllenNLP support comming soon!

Usage Guide

To use a Parquet file for training a PyTorch model simply choose and import the right data set/loader (for your library of choice). Then you can simply try and use it just as you would for any other (simple) text/image file (look at your libraries relavent docs). Some examples are included in the Tests.py script.

Developer/Contributor Guide

To help develop/extend this library please use the following workflow:

1. Fork and clone the repository
2. Make your modifications
3. Test your modifications (run tests with python -m pytest and install in test mode with pip install -e .)
4. Commit (git commit -m "description of changes") and push (git push) your changes
5. Create a pull request

Feel free to add any feature you see fit, or fix/report any bug you find (with GitHub Issues). When creating a pull request please ensure you've carefully described what you're doing, why and a brief overview of your changes. The commits should be small, only changing one feature at a time.