/data-wrangler

Wrangle messy numerical, image, and text data into consistent well-organized formats

Primary LanguagePythonMIT LicenseMIT

Overview

build status docs status doi

Datasets come in all shapes and sizes, and are often messy:

  • Observations come in different formats
  • There are missing values
  • Labels are missing and/or aren't consistent
  • Datasets need to be wrangled 🐄 🐑 🚜

The main goal of data-wrangler is to turn messy data into clean(er) data, defined as either a DataFrame or a list of DataFrame objects. The package provides code for easily wrangling data from a variety of formats into DataFrame objects, manipulating DataFrame objects in useful ways (that can be tricky to implement, but that apply to many analysis scenarios), and decorating Python functions to make them more flexible and/or easier to write.

The data-wrangler package supports a variety of datatypes. There is a special emphasis on text data, whereby data-wrangler provides a simple API for interacting with natural language processing tools and datasets provided by scikit-learn, hugging-face, and flair. The package is designed to provide sensible defaults, but also implements convenient ways of deeply customizing how different datatypes are wrangled.

For more information, including a formal API and tutorials, check out https://data-wrangler.readthedocs.io

Quick start

Install datawrangler using:

$ pip install pydata-wrangler

Some quick natural language processing examples:

import datawrangler as dw

# load in sample text
text_url = 'https://raw.githubusercontent.com/ContextLab/data-wrangler/main/tests/resources/home_on_the_range.txt'
text = dw.io.load(text_url)

# embed text using scikit-learn's implementation of Latent Dirichlet Allocation, trained on a curated subset of
# Wikipedia, called the 'minipedia' corpus.  Return the fitted model so that it can be applied to new text.
lda = {'model': ['CountVectorizer', 'LatentDirichletAllocation'], 'args': [], 'kwargs': {}}
lda_embeddings, lda_fit = dw.wrangle(text, text_kwargs={'model': lda, 'corpus': 'minipedia'}, return_model=True)

# apply the minipedia-trained LDA model to new text
new_text = 'how much wood could a wood chuck chuck if a wood chuck could check wood?'
new_embeddings = dw.wrangle(new_text, text_kwargs={'model': lda_fit})

# embed text using hugging-face's pre-trained GPT2 model
gpt2 = {'model': 'TransformerDocumentEmbeddings', 'args': ['gpt2'], 'kwargs': {}}
gpt2_embeddings = dw.wrangle(text, text_kwargs={'model': gpt2})

The data-wrangler package also provides powerful decorators that can modify existing functions to support new datatypes. Just write your function as though its inputs are guaranteed to be Pandas DataFrames, and decorate it with datawrangler.decorate.funnel to enable support for other datatypes without any new code:

image_url = 'https://raw.githubusercontent.com/ContextLab/data-wrangler/main/tests/resources/wrangler.jpg'
image = dw.io.load(image_url)

# define your function and decorate it with "funnel"
@dw.decorate.funnel
def binarize(x):
  return x > np.mean(x.values)

binarized_image = binarize(image)  # rgb channels will be horizontally concatenated to create a 2D DataFrame

Supported data formats

One package can't accommodate every foreseeable format or input source, but data-wrangler provides a framework for adding support for new datatypes in a straightforward way. Essentially, adding support for a new data type entails writing two functions:

  • An is_<datatype> function, which should return True if an object is compatible with the given datatype (or format), and False otherwise
  • A wrangle_<datatype> function, which should take in an object of the given type or format and return a pandas DataFrame with numerical entries

Currently supported datatypes are limited to:

  • array-like objects (including images)
  • DataFrame-like or Series-like objects
  • text data (text is embedded using natural language processing models)

or lists of mixtures of the above.

Missing observations (e.g., nans, empty strings, etc.) may be filled in using imputation and/or interpolation.