tabben 0.0.6

A package for working with datasets from the open benchmark for tabular data

A Python package for the Open Benchmark for Tabular Data

Set Up

For the most recent stable release, you can install the tabben package from PyPI:

pip3 install tabben --upgrade

If testing/using locally from source, you can install the tabben package locally; first clone this repository locally, and then install from this directory:

pip3 install -e .

Documentation

There are PyTorch Dataset objects available from tabben.datasets. For example,

from tabben.datasets import OpenTabularDataset
from torch.utils.data import DataLoader

# load the arcene dataset (default is train split) and
# save the data to the current directory 
ds = OpenTabularDataset('./', 'arcene')
for inputs, labels in DataLoader(ds, batch_size=4):
    # do stuff with inputs and labels
    pass

All the currently implemented datasets are accessible this way, except for the CIFAR10 dataset. Since it's a standard computer vision dataset, we just wrap the CIFAR10 Dataset from torchvision and convert the images into a vector.

from tabben.datasets import TabularCIFAR10Dataset
from torch.utils.data import DataLoader

cifar_ds = TabularCIFAR10Dataset('./')
for inputs, labels in DataLoader(cifar_ds, batch_size=4):
    # do something with inputs and labels
    pass

The non-image tabular datasets can also be accessed as either numpy arrays or pandas dataframes:

from tabben.datasets import OpenTabularDataset

# load the training set as numpy arrays (these are *not* copies) 
ds = OpenTabularDataset('./', 'covertype')  # defaults are numpy arrays of the training set
train_X, train_y = ds.numpy()

# load as a single pandas dataframe
df = ds.dataframe()

ds_inputs = df[ds.input_attributes]
ds_outputs = df[ds.output_attributes]

For a list of all the currently implemented datasets in the benchmark (except for CIFAR10), there's the function:

from tabben.datasets import list_datasets

print(list_datasets())

Testing

After installing pytest, all the tests can be run by just running

pytest

from a parent directory.