DataWig - Imputation for Tables

DataWig learns Machine Learning models to impute missing values in tables.

See our user-guide and extended documentation here.

Installation

CPU

pip3 install datawig

GPU

If you want to run DataWig on a GPU you need to make sure your version of Apache MXNet Incubating contains the GPU bindings. Depending on your version of CUDA, you can do this by running the following:

wget https://raw.githubusercontent.com/awslabs/datawig/master/requirements/requirements.gpu-cu${CUDA_VERSION}.txt
pip install datawig --no-deps -r requirements.gpu-cu${CUDA_VERSION}.txt
rm requirements.gpu-cu${CUDA_VERSION}.txt

where ${CUDA_VERSION} can be 75 (7.5), 80 (8.0), 90 (9.0), or 91 (9.1).

Running DataWig

The DataWig API expects your data as a pandas DataFrame. Here is an example of how the dataframe might look:

Product Type	Description	Size	Color
Shoe	Ideal for Running	12UK	Black
SDCards	Best SDCard ever ...	8GB	Blue
Dress	This yellow dress	M	?

Quickstart Example

For most use cases, the SimpleImputer class is the best starting point. For convenience there is the function SimpleImputer.complete that takes a DataFrame and fits an imputation model for each column with missing values, with all other columns as inputs:

import datawig, numpy

# generate some data with simple nonlinear dependency
df = datawig.utils.generate_df_numeric() 
# mask 10% of the values
df_with_missing = df.mask(numpy.random.rand(*df.shape) > .9)

# impute missing values
df_with_missing_imputed = datawig.SimpleImputer.complete(df_with_missing)

You can also impute values in specific columns only (called output_column below) using values in other columns (called input_columns below). DataWig currently supports imputation of categorical columns and numeric columns.

Imputation of categorical columns

import datawig

df = datawig.utils.generate_df_string( num_samples=200, 
                                       data_column_name='sentences', 
                                       label_column_name='label')

df_train, df_test = datawig.utils.random_split(df)

#Initialize a SimpleImputer model
imputer = datawig.SimpleImputer(
    input_columns=['sentences'], # column(s) containing information about the column we want to impute
    output_column='label', # the column we'd like to impute values for
    output_path = 'imputer_model' # stores model data and metrics
    )

#Fit an imputer model on the train data
imputer.fit(train_df=df_train)

#Impute missing values and return original dataframe with predictions
imputed = imputer.predict(df_test)

Imputation of numerical columns

import datawig

df = datawig.utils.generate_df_numeric( num_samples=200, 
                                        data_column_name='x', 
                                        label_column_name='y')         
df_train, df_test = datawig.utils.random_split(df)

#Initialize a SimpleImputer model
imputer = datawig.SimpleImputer(
    input_columns=['x'], # column(s) containing information about the column we want to impute
    output_column='y', # the column we'd like to impute values for
    output_path = 'imputer_model' # stores model data and metrics
    )

#Fit an imputer model on the train data
imputer.fit(train_df=df_train, num_epochs=50)

#Impute missing values and return original dataframe with predictions
imputed = imputer.predict(df_test)
             

In order to have more control over the types of models and preprocessings, the Imputer class allows directly specifying all relevant model features and parameters.

For details on usage, refer to the provided examples.

Acknowledgments

Thanks to David Greenberg for the package name.

Building documentation

git clone git@github.com:awslabs/datawig.git
cd datawig/docs
make html
open _build/html/index.html

Executing Tests

Clone the repository from git and set up virtualenv in the root dir of the package:

python3 -m venv venv

Install the package from local sources:

./venv/bin/pip install -e .

Run tests:

./venv/bin/pip install -r requirements/requirements.dev.txt
./venv/bin/python -m pytest

Updating PyPi distribution

Before updating, increment the version in setup.py.

git clone git@github.com:awslabs/datawig.git
cd datawig
# build local distribution for current version
python setup.py sdist
# upload to PyPi
twine upload --skip-existing dist/*