Tools for processing tabular datasets for PyTorch
Project description
torchtable
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.
Installation
Install via pip.
$ pip install torchtable
Documentation
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.
Usage
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.
TODO
Add more models
Implement default field selection
Implement text field/operations
Implement swap noise
Implement input/output validation
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for torchtable-0.1.0-py3-none-any.whl
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 | 29b5768fb6ee3195eea3dc8e23d3b072e17fff2d69aa1e2309b8cc1ccbbfb4b1 |
|
| MD5 | ba54e0fb26eefd7b8d9e776f4b7fedaa |
|
| BLAKE2b-256 | d7a40eab2395179abce676828882466e41991de388fa9235f0795db349c8dc1a |
