Torchtable is a library for handling tabular datasets in PyTorch. It is heavily inspired by torchtext and uses a similar API but without some of the limitations (e.g. only one field per column). Torchtable aims to be simple to use and easily extensible. It provides sensible defaults while allowing the user to define their own custom pipelines, putting all of this behind an intuitive interface.

Install via pip.

$ pip install torchtable

Documentation is a work in progress, but the current docs can be read here. In addition, you can read the notebooks in the examples directory or dev_nb directory to learn more.

Torchtable uses a declarative API similar to torchtext. Here is an example of how you might handle an imaginary dataset where you are supposed to predict the price of some product.

>>> train = TabularDataset.from_csv('data/train.csv',
...    fields={'seller_id': CategoricalField(min_freq=3),
...            'timestamp': [DayofWeekField(), HourField()],
...            'price': NumericalField(fill_missing="median", is_target=True)
...    })
...

See the examples directory for more examples.

• Add more models
• Implement default field selection
• Implement text field/operations
• Implement swap noise
• Implement input/output validation