Pytorch Datastream

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.

See the documentation for more information.

Install

poetry add pytorch-datastream

Or, for the old-timers:

pip install pytorch-datastream

Usage

The list below is meant to showcase functions that are useful in most standard and non-standard cases. It is not meant to be an exhaustive list. See the documentation for a more extensive list on API and usage.

Dataset.from_subscriptable
Dataset.from_dataframe
Dataset
    .map
    .subset
    .split
    .cache
    .with_columns

Datastream.merge
Datastream.zip
Datastream
    .map
    .data*loader
    .zip_index
    .update_weights*
    .update*example_weight*
    .weight
    .state_dict
    .load_state_dict

Simple image dataset example

Here's a basic example of loading images from a directory:

from datastream import Dataset
from pathlib import Path
from PIL import Image

# Assuming images are in a directory structure like:
# images/
#   class1/
#     image1.jpg
#     image2.jpg
#   class2/
#     image3.jpg
#     image4.jpg

image_dir = Path("images")
image_paths = list(image_dir.glob("\*_/_.jpg"))

dataset = (
    Dataset.from_paths(
    image_paths,
    pattern=r".\*/(?P<class_name>\w+)/(?P<image_name>\w+).jpg"
    )
    .map(lambda row: dict(
        image=Image.open(row["path"]),
        class_name=row["class_name"],
        image_name=row["image_name"],
    ))
)

# Access an item from the dataset

first_item = dataset[0]
print(f"Class: {first_item['class_name']}, Image name: {first_item['image_name']}")

Merge / stratify / oversample datastreams

The fruit datastreams given below repeatedly yields the string of its fruit type.

>>> datastream = Datastream.merge([
>>> ... (apple_datastream, 2),
>>> ... (pear_datastream, 1),
>>> ... (banana_datastream, 1),
>>> ... ])
>>> next(iter(datastream.data_loader(batch_size=8)))
>>> ['apple', 'apple', 'pear', 'banana', 'apple', 'apple', 'pear', 'banana']
>>>

Zip independently sampled datastreams

The fruit datastreams given below repeatedly yields the string of its fruit type.

>>> datastream = Datastream.zip([
>>> ... apple_datastream,
>>> ... Datastream.merge([pear_datastream, banana_datastream]),
>>> ... ])
>>> next(iter(datastream.data_loader(batch_size=4)))
>>> [('apple', 'pear'), ('apple', 'banana'), ('apple', 'pear'), ('apple', 'banana')]
>>>

More usage examples

See the documentation for more usage examples.