notesum.ai

Published at November 21

Quantization without Tears

cs.CV

Released Date: November 21, 2024

Authors: Minghao Fu¹, Hao Yu¹, Jie Shao¹, Junjie Zhou¹, Ke Zhu¹, Jianxin Wu¹

Aff.: ¹National Key Laboratory for Novel Software Technology, Nanjing University, China

Arxiv: http://arxiv.org/abs/2411.13918v1

Refer to caption

Network	Method	#Bits	0Size	Top-1
DeiT-T	Full-precision	32/32	022.9	72.2
\cdashline2-5	IGQ-ViT^† [38]	4/4	-	62.5
	RepQ-ViT [27]	4/4	003.3	58.2
	RepQ-ViT + QwT	4/4	004.2	61.4
	RepQ-ViT + QwT^∗	4/4	004.2	64.8
\cdashline2-5	IGQ-ViT^† [38]	6/6	-	71.2
	RepQ-ViT [27]	6/6	004.6	71.0
	RepQ-ViT + QwT	6/6	005.5	71.2
	RepQ-ViT + QwT^∗	6/6	005.5	71.6
Swin-T	Full-precision	32/32	113.2	81.4
\cdashline2-5	IGQ-ViT^† [38]	4/4	-	77.8
	RepQ-ViT [27]	4/4	014.9	73.0
	RepQ-ViT + QwT	4/4	019.2	75.5
	RepQ-ViT + QwT^∗	4/4	019.2	79.3
\cdashline2-5	IGQ-ViT^† [38]	6/6	-	80.9
	RepQ-ViT [27]	6/6	021.7	80.6
	RepQ-ViT + QwT	6/6	026.0	80.7
	RepQ-ViT + QwT^∗	6/6	026.0	80.9
ViT-B	Full-precision	32/32	346.3	84.5
\cdashline2-5	IGQ-ViT^† [38]	4/4	-	79.3
	RepQ-ViT [27]	4/4	044.9	68.5
	RepQ-ViT + QwT	4/4	059.1	76.3
	RepQ-ViT + QwT^∗	4/4	059.1	78.5
\cdashline2-5	IGQ-ViT^† [38]	6/6	-	83.8
	RepQ-ViT [27]	6/6	066.2	83.6
	RepQ-ViT + QwT	6/6	080.4	83.9
	RepQ-ViT + QwT^∗	6/6	080.4	84.0
ResNet-50	Full-precision	32/32	102.2	76.6
\cdashline2-5	CL-Calib^† [47]	4/4	-	75.4
	Percentile[23]	4/4	014.0	68.4
	Percentile + QwT	4/4	016.0	74.5
	Percentile + QwT^∗	4/4	016.0	75.8
\cdashline2-5	CL-Calib^† [47]	6/6	-	-
	Percentile[23]	6/6	019.9	76.0
	Percentile + QwT	6/6	021.9	76.8
	Percentile + QwT^∗	6/6	021.9	76.8