EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

Mingxing TanQuoc V. Le

   Papers with code   Abstract  PDF

Convolutional Neural Networks (ConvNets) are commonly developed at a fixed resource budget, and then scaled up for better accuracy if more resources are available. In this paper, we systematically study model scaling and identify that carefully balancing network depth, width, and resolution can lead to better performance... (read more)

Benchmarked Models

RANK
MODEL
REPO
CODE RESULT
PAPER RESULT
ε-REPRODUCED
BUILD
1
EfficientNet-B7
(RandAugment)
84.9%
--
2
EfficientNet-B7
(AutoAugment)
84.4%
--
3
EfficientNet-B7b
84.3%
--
4
EfficientNet-B6
(AutoAugment)
84.1%
--
6
EfficientNet-B6b
84.1%
--
8
EfficientNet-B5
(RandAugment)
83.8%
--
9
EfficientNet-B5
(AutoAugment)
83.7%
--
10
EfficientNet-B5b
83.6%
--
12
EfficientNet-B4
(AutoAugment)
83.0%
--
13
EfficientNet-B4b
82.8%
--
15
EfficientNet-B3
(AutoAugment)
81.6%
--
16
EfficientNet-B3b
81.4%
--
18
EfficientNet-EdgeTPU-L
80.4%
--
19
EfficientNet-B2
(AutoAugment)
80.1%
--
22
EfficientNet-B2b
79.7%
--
23
EfficientNet-B1
(AutoAugment)
78.8%
--
24
EfficientNet-EdgeTPU-M
78.7%
--
27
EfficientNet-B1b
78.4%
--
28
EfficientNet-EdgeTPU-S
77.3%
--
29
EfficientNet-B1
77.0%
78.8%
31
EfficientNet-B0
(AutoAugment)
76.8%
--
33
EfficientNet-B0b
76.1%
--
34
EfficientNet-B0
75.2%
76.3%