Elastic ViTs from Pretrained Models without Retraining

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Simoncini, Walter, Dorkenwald, Michael, Blankevoort, Tijmen, Snoek, Cees G. M., Asano, Yuki M.
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911732632387584
author Simoncini, Walter
Dorkenwald, Michael
Blankevoort, Tijmen
Snoek, Cees G. M.
Asano, Yuki M.
author_facet Simoncini, Walter
Dorkenwald, Michael
Blankevoort, Tijmen
Snoek, Cees G. M.
Asano, Yuki M.
contents Vision foundation models achieve remarkable performance but are only available in a limited set of pre-determined sizes, forcing sub-optimal deployment choices under real-world constraints. We introduce SnapViT: Single-shot network approximation for pruned Vision Transformers, a new post-pretraining structured pruning method that enables elastic inference across a continuum of compute budgets. Our approach efficiently combines gradient information with cross-network structure correlations, approximated via an evolutionary algorithm, does not require labeled data, generalizes to models without a classification head, and is retraining-free. Experiments on DINO, SigLIPv2, DeIT, and AugReg models demonstrate superior performance over state-of-the-art methods across various sparsities, requiring less than five minutes on a single A100 GPU to generate elastic models that can be adjusted to any computational budget. Our key contributions include an efficient pruning strategy for pretrained Vision Transformers, a novel evolutionary approximation of Hessian off-diagonal structures, and a self-supervised importance scoring mechanism that maintains strong performance without requiring retraining or labels. Code and pruned models are available at: https://elastic.ashita.nl/
format Preprint
id arxiv_https___arxiv_org_abs_2510_17700
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Elastic ViTs from Pretrained Models without Retraining
Simoncini, Walter
Dorkenwald, Michael
Blankevoort, Tijmen
Snoek, Cees G. M.
Asano, Yuki M.
Computer Vision and Pattern Recognition
Vision foundation models achieve remarkable performance but are only available in a limited set of pre-determined sizes, forcing sub-optimal deployment choices under real-world constraints. We introduce SnapViT: Single-shot network approximation for pruned Vision Transformers, a new post-pretraining structured pruning method that enables elastic inference across a continuum of compute budgets. Our approach efficiently combines gradient information with cross-network structure correlations, approximated via an evolutionary algorithm, does not require labeled data, generalizes to models without a classification head, and is retraining-free. Experiments on DINO, SigLIPv2, DeIT, and AugReg models demonstrate superior performance over state-of-the-art methods across various sparsities, requiring less than five minutes on a single A100 GPU to generate elastic models that can be adjusted to any computational budget. Our key contributions include an efficient pruning strategy for pretrained Vision Transformers, a novel evolutionary approximation of Hessian off-diagonal structures, and a self-supervised importance scoring mechanism that maintains strong performance without requiring retraining or labels. Code and pruned models are available at: https://elastic.ashita.nl/
title Elastic ViTs from Pretrained Models without Retraining
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2510.17700