notesum.ai

Published at November 27

Mixture of Experts in Image Classification: What's the Sweet Spot?

cs.CV

cs.LG

Released Date: November 27, 2024

Authors: Mathurin Videau¹, Alessandro Leite¹, Marc Schoenauer¹, Olivier Teytaud²

Aff.: ¹TAU, Inria Saclay; ²Meta AI, FAIR Labs

Arxiv: http://arxiv.org/abs/2411.18322v1

Refer to caption

Architecture	`#`Params	Per sample	FLOPs	Throughput	IN-1K
Architecture	( $\times 10^{6}$ )	`#`Params_act		(im/s)	Accuracy
ConvNeXt-T (Liu et al. 2022)	28.6	28.6	4.5G	814	82.1
ConvNeXt-T-4 Last 2 Top 1	34.5	25.6	4.2G	768	82.1
ConvNeXt-S (Liu et al. 2022)	50	50	8.7G	466	83.1
ConvNeXt-S-4 Last 2 Top 1	56.1	47.3	8.5G	442	83.1
ConvNeXt-B (Liu et al. 2022)	88.6	88.6	15.4G	299	83.8
ConvNeXt-B-4 Last 2 Top 1	99.1	83.4	15.0G	289	83.5
ConvNeXt-S (iso.)	22.3	22.3	4.3G	1100	79.7
ConvNeXt-S-8 (iso.) Last 2 Top 1	38.9	22.3	4.3G	1031	80.3
ConvNeXt-B (iso.)	82.4	82.4	16.9G	336	82.0
ConvNeXt-B-8 (iso.) Last 2 Top 1	115.4	82.4	16.9G	303	81.6
ViT-S	22.0	22.0	4.6G	1083	79.8
ViT-S-8 Last 2 Top 2	38.6	25.0	5.3G	892	80.5
ViT-S-8 Every 2 Top 2	71.7	33.1	6.9G	724	80.7
ViT-B	86.6	86.6	17.5G	329	82.8
ViT-B-8 Every 2 Top 2	284.9	129.9	26.3G	227	82.5