Shortcuts

EfficientNet

Introduction

EfficientNets are a family of image classification models, which achieve state-of-the-art accuracy, yet being an order-of-magnitude smaller and faster than previous models.

EfficientNets are based on AutoML and Compound Scaling. In particular, we first use AutoML MNAS Mobile framework to develop a mobile-size baseline network, named as EfficientNet-B0; Then, we use the compound scaling method to scale up this baseline to obtain EfficientNet-B1 to B7.

Abstract

Click to show the detailed Abstract
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. Based on this observation, we propose a new scaling method that uniformly scales all dimensions of depth/width/resolution using a simple yet highly effective compound coefficient. We demonstrate the effectiveness of this method on scaling up MobileNets and ResNet. To go even further, we use neural architecture search to design a new baseline network and scale it up to obtain a family of models, called EfficientNets, which achieve much better accuracy and efficiency than previous ConvNets. In particular, our EfficientNet-B7 achieves state-of-the-art 84.3% top-1 accuracy on ImageNet, while being 8.4x smaller and 6.1x faster on inference than the best existing ConvNet. Our EfficientNets also transfer well and achieve state-of-the-art accuracy on CIFAR-100 (91.7%), Flowers (98.8%), and 3 other transfer learning datasets, with an order of magnitude fewer parameters.

How to use it?

from mmpretrain import inference_model

predict = inference_model('efficientnet-b0_3rdparty_8xb32_in1k', 'demo/bird.JPEG')
print(predict['pred_class'])
print(predict['pred_score'])

Models and results

Image Classification on ImageNet-1k

Model

Pretrain

Params (M)

Flops (G)

Top-1 (%)

Top-5 (%)

Config

Download

efficientnet-b0_3rdparty_8xb32_in1k*

From scratch

5.29

0.42

76.74

93.17

config

model

efficientnet-b0_3rdparty_8xb32-aa_in1k*

From scratch

5.29

0.42

77.26

93.41

config

model

efficientnet-b0_3rdparty_8xb32-aa-advprop_in1k*

From scratch

5.29

0.42

77.53

93.61

config

model

efficientnet-b0_3rdparty-ra-noisystudent_in1k*

From scratch

5.29

0.42

77.63

94.00

config

model

efficientnet-b1_3rdparty_8xb32_in1k*

From scratch

7.79

0.74

78.68

94.28

config

model

efficientnet-b1_3rdparty_8xb32-aa_in1k*

From scratch

7.79

0.74

79.20

94.42

config

model

efficientnet-b1_3rdparty_8xb32-aa-advprop_in1k*

From scratch

7.79

0.74

79.52

94.43

config

model

efficientnet-b1_3rdparty-ra-noisystudent_in1k*

From scratch

7.79

0.74

81.44

95.83

config

model

efficientnet-b2_3rdparty_8xb32_in1k*

From scratch

9.11

1.07

79.64

94.80

config

model

efficientnet-b2_3rdparty_8xb32-aa_in1k*

From scratch

9.11

1.07

80.21

94.96

config

model

efficientnet-b2_3rdparty_8xb32-aa-advprop_in1k*

From scratch

9.11

1.07

80.45

95.07

config

model

efficientnet-b2_3rdparty-ra-noisystudent_in1k*

From scratch

9.11

1.07

82.47

96.23

config

model

efficientnet-b3_3rdparty_8xb32_in1k*

From scratch

12.23

1.95

81.01

95.34

config

model

efficientnet-b3_3rdparty_8xb32-aa_in1k*

From scratch

12.23

1.95

81.58

95.67

config

model

efficientnet-b3_3rdparty_8xb32-aa-advprop_in1k*

From scratch

12.23

1.95

81.81

95.69

config

model

efficientnet-b3_3rdparty-ra-noisystudent_in1k*

From scratch

12.23

1.95

84.02

96.89

config

model

efficientnet-b4_3rdparty_8xb32_in1k*

From scratch

19.34

4.66

82.57

96.09

config

model

efficientnet-b4_3rdparty_8xb32-aa_in1k*

From scratch

19.34

4.66

82.95

96.26

config

model

efficientnet-b4_3rdparty_8xb32-aa-advprop_in1k*

From scratch

19.34

4.66

83.25

96.44

config

model

efficientnet-b4_3rdparty-ra-noisystudent_in1k*

From scratch

19.34

4.66

85.25

97.52

config

model

efficientnet-b5_3rdparty_8xb32_in1k*

From scratch

30.39

10.80

83.18

96.47

config

model

efficientnet-b5_3rdparty_8xb32-aa_in1k*

From scratch

30.39

10.80

83.82

96.76

config

model

efficientnet-b5_3rdparty_8xb32-aa-advprop_in1k*

From scratch

30.39

10.80

84.21

96.98

config

model

efficientnet-b5_3rdparty-ra-noisystudent_in1k*

From scratch

30.39

10.80

86.08

97.75

config

model

efficientnet-b6_3rdparty_8xb32-aa_in1k*

From scratch

43.04

19.97

84.05

96.82

config

model

efficientnet-b6_3rdparty_8xb32-aa-advprop_in1k*

From scratch

43.04

19.97

84.74

97.14

config

model

efficientnet-b6_3rdparty-ra-noisystudent_in1k*

From scratch

43.04

19.97

86.47

97.87

config

model

efficientnet-b7_3rdparty_8xb32-aa_in1k*

From scratch

66.35

39.32

84.38

96.88

config

model

efficientnet-b7_3rdparty_8xb32-aa-advprop_in1k*

From scratch

66.35

39.32

85.14

97.23

config

model

efficientnet-b7_3rdparty-ra-noisystudent_in1k*

From scratch

66.35

39.32

86.83

98.08

config

model

efficientnet-b8_3rdparty_8xb32-aa-advprop_in1k*

From scratch

87.41

65.00

85.38

97.28

config

model

efficientnet-l2_3rdparty-ra-noisystudent_in1k-800px*

From scratch

480.31

174.20

88.33

98.65

config

model

efficientnet-l2_3rdparty-ra-noisystudent_in1k-475px*

From scratch

480.31

484.98

88.18

98.55

config

model

Models with * are converted from the official repo. The config files of these models are only for inference. We haven’t reproduce the training results.

Citation

@inproceedings{tan2019efficientnet,
  title={Efficientnet: Rethinking model scaling for convolutional neural networks},
  author={Tan, Mingxing and Le, Quoc},
  booktitle={International Conference on Machine Learning},
  pages={6105--6114},
  year={2019},
  organization={PMLR}
}
Read the Docs v: latest
Versions
latest
stable
mmcls-1.x
mmcls-0.x
dev
Downloads
epub
On Read the Docs
Project Home
Builds

Free document hosting provided by Read the Docs.