Function data

Base Dataloaders and Dataset

tinyBIG offers two base primitives to work with data: dataloader and dataset, both defined by module tinybig.data.base_data in the package. dataset stores the data instances (including features, labels, and optional encoders for feature embedding), and dataloader wraps an iterable around the dataset.

Based on dataloader and dataset, several dataloaders for specific data modalities have been created:

import tinybig as tb
from tinybig.data import dataloader, dataset
from tinybig.data import function_dataloader, vision_dataloader, text_dataloader, tabular_dataloader

Built based on torchvision and torchtext, tinyBIG can load many real-world vision data, like MNIST and CIFAR10, and text data, like IMDB, SST2 and AGNews, for model training and evaluation. In addition, tinyBIG also offers a variety of other well-known datasets by itself, including continuous function datasets, like Elementary, Composite and Feynman functions, and classic tabular datasets, like Iris, Diabetes and Banknote.