Ecoset - a large-scale vision dataset

🚀 Feature

We (original authors) would like to suggest the inclusion of ecoset, a new large-scale image dataset, to be included in torchvision.datasets.

Motivation

Ecoset consists of 1.5m images originating from 565 basic-level categories that are of importance to humans. Categories were chosen to be (a) basic level (think "dog" instead of "cocker spaniel"), (b) most common in English language usage, and (c) rated as concrete by human observers. Compared to ImageNet, networks trained on ecoset are more similar to representations found in the brain (see https://www.pnas.org/content/118/8/e2011417118). Ten AlexNet and vNet instances trained on ecoset are available with the dataset (https://codeocean.com/capsule/9570390/tree/v1). We have been approached by multiple laboratories who would like to easily access ecoset from within the pytorch infrastructure, hence the request.

Pitch

Please consider including ecoset as a dataset option in torchvision.

Alternatives

Additional context

cc @pmeier

Start in torchvision.datasets and compare existing large-scale image dataset integrations to understand the expected dataset API, download behavior, and tests. Done means ecoset is available as a torchvision dataset option with the expected access and verification covered by the project’s existing conventions.