# EfficientNetV2¶

## 摘要¶

This paper introduces EfficientNetV2, a new family of convolutional networks that have faster training speed and better parameter efficiency than previous models. To develop this family of models, we use a combination of training-aware neural architecture search and scaling, to jointly optimize training speed and parameter efficiency. The models were searched from the search space enriched with new ops such as Fused-MBConv. Our experiments show that EfficientNetV2 models train much faster than state-of-the-art models while being up to 6.8x smaller. Our training can be further sped up by progressively increasing the image size during training, but it often causes a drop in accuracy. To compensate for this accuracy drop, we propose to adaptively adjust regularization (e.g., dropout and data augmentation) as well, such that we can achieve both fast training and good accuracy. With progressive learning, our EfficientNetV2 significantly outperforms previous models on ImageNet and CIFAR/Cars/Flowers datasets. By pretraining on the same ImageNet21k, our EfficientNetV2 achieves 87.3% top-1 accuracy on ImageNet ILSVRC2012, outperforming the recent ViT by 2.0% accuracy while training 5x-11x faster using the same computing resources. Code will be available at https://github.com/google/automl/tree/master/efficientnetv2.

## 使用方式¶

>>> import torch
>>> from mmcls.apis import init_model, inference_model
>>>
>>> predict = inference_model(model, 'demo/demo.JPEG')
>>> print(predict['pred_class'])
sea snake
>>> print(predict['pred_score'])
0.3147328197956085


For more configurable parameters, please refer to the API.

## 结果和模型¶

### ImageNet-1k¶

Flops(G)

Top-1 (%)

Top-5 (%)

EfficientNetV2-b0* (efficientnetv2-b0_3rdparty_in1k)

7.14

0.92

78.52

94.44

config

model

EfficientNetV2-b1* (efficientnetv2-b1_3rdparty_in1k)

8.14

1.44

79.80

94.89

config

model

EfficientNetV2-b2* (efficientnetv2-b2_3rdparty_in1k)

10.10

1.99

80.63

95.30

config

model

EfficientNetV2-b3* (efficientnetv2-b3_3rdparty_in1k)

14.36

3.50

82.03

95.88

config

model

EfficientNetV2-s* (efficientnetv2-s_3rdparty_in1k)

21.46

9.72

83.82

96.67

config

model

EfficientNetV2-m* (efficientnetv2-m_3rdparty_in1k)

54.14

26.88

85.01

97.26

config

model

EfficientNetV2-l* (efficientnetv2-l_3rdparty_in1k)

118.52

60.14

85.43

97.31

config

model

EfficientNetV2-s* (efficientnetv2-s_in21k-pre_3rdparty_in1k)

ImageNet 21k

21.46

9.72

84.29

97.26

config

model

EfficientNetV2-m* (efficientnetv2-m_in21k-pre_3rdparty_in1k)

ImageNet 21k

54.14

26.88

85.47

97.76

config

model

EfficientNetV2-l* (efficientnetv2-l_in21k-pre_3rdparty_in1k)

ImageNet 21k

118.52

60.14

86.31

97.99

config

model

EfficientNetV2-xl* (efficientnetv2-xl_in21k-pre_3rdparty_in1k)

ImageNet 21k

208.12

98.34

86.39

97.83

config

model

Models with * are converted from the official repo. The config files of these models are only for inference. We don’t ensure these config files’ training accuracy and welcome you to contribute your reproduction results.

### Pre-trained Models In ImageNet-21K¶

The pre-trained models are only used to fine-tune, and therefore cannot be trained and don’t have evaluation results.

Flops(G)

EfficientNetV2-s* (efficientnetv2-s_3rdparty_in21k)

21.46

9.72

config

model

EfficientNetV2-m* (efficientnetv2-m_3rdparty_in21k)

54.14

26.88

config

model

EfficientNetV2-l* (efficientnetv2-l_3rdparty_in21k)

118.52

60.14

config

model

EfficientNetV2-xl* (efficientnetv2-xl_3rdparty_in21k)

208.12

98.34

config

model

Models with * are converted from the official repo.

## 引用¶

@inproceedings{tan2021efficientnetv2,
title={Efficientnetv2: Smaller models and faster training},
author={Tan, Mingxing and Le, Quoc},
booktitle={International Conference on Machine Learning},
pages={10096--10106},
year={2021},
organization={PMLR}
}