Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers

15 Mar 2024  Â·  Sanghyeok Lee, Joonmyung Choi, Hyunwoo J. Kim ·

Vision Transformer (ViT) has emerged as a prominent backbone for computer vision. For more efficient ViTs, recent works lessen the quadratic cost of the self-attention layer by pruning or fusing the redundant tokens. However, these works faced the speed-accuracy trade-off caused by the loss of information. Here, we argue that token fusion needs to consider diverse relations between tokens to minimize information loss. In this paper, we propose a Multi-criteria Token Fusion (MCTF), that gradually fuses the tokens based on multi-criteria (e.g., similarity, informativeness, and size of fused tokens). Further, we utilize the one-step-ahead attention, which is the improved approach to capture the informativeness of the tokens. By training the model equipped with MCTF using a token reduction consistency, we achieve the best speed-accuracy trade-off in the image classification (ImageNet1K). Experimental results prove that MCTF consistently surpasses the previous reduction methods with and without training. Specifically, DeiT-T and DeiT-S with MCTF reduce FLOPs by about 44% while improving the performance (+0.5%, and +0.3%) over the base model, respectively. We also demonstrate the applicability of MCTF in various Vision Transformers (e.g., T2T-ViT, LV-ViT), achieving at least 31% speedup without performance degradation. Code is available at https://github.com/mlvlab/MCTF.

PDF Abstract

Datasets


Task Dataset Model Metric Name Metric Value Global Rank Benchmark
Efficient ViTs ImageNet-1K (with DeiT-S) MCTF ($r=16$) Top 1 Accuracy 80.1 # 1
GFLOPs 2.6 # 12
Efficient ViTs ImageNet-1K (with DeiT-S) MCTF ($r=20$) Top 1 Accuracy 79.5 # 19
GFLOPs 2.2 # 4
Efficient ViTs ImageNet-1K (with DeiT-S) MCTF ($r=18$) Top 1 Accuracy 79.9 # 3
GFLOPs 2.4 # 10
Efficient ViTs ImageNet-1K (with DeiT-T) MCTF ($r=8$) Top 1 Accuracy 72.9 # 1
GFLOPs 1.0 # 18
Efficient ViTs ImageNet-1K (with DeiT-T) MCTF ($r=20$) Top 1 Accuracy 71.4 # 16
GFLOPs 0.6 # 1
Efficient ViTs ImageNet-1K (with DeiT-T) MCTF ($r=16$) Top 1 Accuracy 72.7 # 3
GFLOPs 0.7 # 5
Efficient ViTs ImageNet-1K (With LV-ViT-S) MCTF ($r=8$) Top 1 Accuracy 83.5 # 1
GFLOPs 4.9 # 4
Efficient ViTs ImageNet-1K (With LV-ViT-S) MCTF ($r=12$) Top 1 Accuracy 83.4 # 2
GFLOPs 4.2 # 13
Efficient ViTs ImageNet-1K (With LV-ViT-S) MCTF ($r=16$) Top 1 Accuracy 82.3 # 19
GFLOPs 3.6 # 19

Methods


DeiT • Dropout • LV-ViT • Pruning • T2T-ViT • Vision Transformer