OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Ats

Ats is an Apache-2.0 implementation of Adaptive Token Sampling, a differentiable, parameter-free module for vision transformers. It scores and adaptively samples significant image tokens so their number can vary by input, and supports image- and video-classification evaluation workflows.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Ats — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Ats when experimenting with input-dependent token counts in a vision transformer, or when following its documented DeiT-S ImageNet evaluation workflow.

Who it is for

  • Researchers and developers working with vision-transformer image classifiers.
  • Teams evaluating token sampling in image or video classification models.

Documented uses

  • Add adaptive token scoring and sampling to a vision transformer.
  • Evaluate a DeiT-S configuration on ImageNet using a pretrained checkpoint.
  • Train a vision transformer equipped with ATS for classification experiments.
Before you choose

Requirements & limits

The implementation is public on GitHub and released under the Apache-2.0 license.

  • The documented evaluation setup requires Python 3.8 or later, timm 0.4.12, PyTorch 1.3 or later and a torchvision version matching PyTorch.
  • ImageNet evaluation requires prepared dataset directories and a separately downloaded pretrained DeiT-S checkpoint.
  • The project describes ATS as pluggable into existing transformer architectures, but integration details may vary by architecture.
  • The README reports a 2× GFLOPs reduction while preserving accuracy on its evaluated ImageNet and Kinetics tasks; results may differ with other architectures or setups.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Evaluate the documented DeiT-S ImageNet configuration

Prepare ImageNet validation data and use the repository's documented command to run the DeiT-S evaluation workflow.

Before you start

  • Python 3.8 or later, timm 0.4.12, PyTorch 1.3 or later, matching torchvision, NumPy, fvcore, simplejson and TensorBoard.
  • ImageNet images downloaded and extracted in the ILSVRC2012 train/val directory structure described in the README.
  • The pretrained DeiT-S checkpoint linked in the README.

Steps

  1. Prepare the ImageNet data using the directory layout shown in the README.

  2. In the validation directory, run the documented valdirs.sh command if needed to prepare ImageNet validation directories.

  3. Download the linked pretrained DeiT-S checkpoint.

  4. Run tools/run_net.py with configs/ImageNet/DeiT_S.yaml, setting DATA.PATH_TO_DATA_DIR to the dataset path and TEST.CHECKPOINT_FILE_PATH to the checkpoint path.

Expected outcome

The repository's documented DeiT-S ImageNet evaluation workflow is configured to use the prepared dataset and pretrained checkpoint.

Guide limitations

  • The README supplies the workflow and command but does not specify evaluation results for a reader's own environment.
  • The data and checkpoint must be obtained separately.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
16 October 2026
Original website title
adaptivetokensampling/ATS · Official homepage title · retained in its source language
How we review sourcesSuggest a correction