Welcome to pytorch-datastream’s documentation!¶
This is a simple library for creating readable dataset pipelines and reusing
best practices for issues such as imbalanced datasets. There are just two
components to keep track of: Dataset and Datastream.
Dataset is a simple mapping between an index and an example. It provides
pipelining of functions in a readable syntax originally adapted from
tensorflow 2’s tf.data.Dataset.
Datastream combines a Dataset and a sampler into a stream of examples.
It provides a simple solution to oversampling / stratification, weighted
sampling, and finally converting to a torch.utils.data.DataLoader.
Contents: