Accelerating Vision Transformers with Adaptive Patch Sizes

TL;DR

Adaptive Patch Transformers (APT) dynamically adjust patch sizes within images to accelerate Vision Transformer processing, significantly boosting speed while maintaining accuracy across various high-resolution visual tasks.

Contribution

Introducing APT, a method that uses multiple patch sizes within a single image to reduce input tokens and speed up ViT inference and training without performance loss.

Findings

Increases throughput by 40% on ViT-L and 50% on ViT-H.

Achieves up to 30% faster training and inference in dense visual tasks.

Converges in as little as 1 epoch when fine-tuned.

Abstract

Vision Transformers (ViTs) partition input images into uniformly sized patches regardless of their content, resulting in long input sequence lengths for high-resolution images. We present Adaptive Patch Transformers (APT), which addresses this by using multiple different patch sizes within the same image. APT reduces the total number of input tokens by allocating larger patch sizes in more homogeneous areas and smaller patches in more complex ones. APT achieves a drastic speedup in ViT inference and training, increasing throughput by 40% on ViT-L and 50% on ViT-H while maintaining downstream performance, and can be applied to a previously fine-tuned ViT, converging in as little as 1 epoch. It also significantly reduces training and inference time without loss of performance in high-resolution dense visual tasks, achieving up to 30\% faster training and inference in visual QA, object detection, and semantic segmentation.